# Accelerator Security Features ## Device attestation is documented for 12 of the 35 accelerators in Scrutica’s chip catalogue, including 9 of the 17 restricted for China. In attestation, a chip signs a report of its firmware for a remote verifier. Below are [the register](https://scrutica.com/verification/hardware#hw-register) of attestation and ten other features, [what has shipped](https://scrutica.com/verification/hardware#hw-shipped), [when each feature arrived](https://scrutica.com/verification/hardware#hw-timeline), [the attestation path](https://scrutica.com/verification/hardware#hw-path), and [what vendors and lawmakers have said about kill switches, backdoors, tracking and location](https://scrutica.com/verification/hardware#hw-statements). - Documented on the device - Partial, or host or cloud only - Attribution to the chip unverified - Not documented Restricted for China · 17 9 documented on the device Restricted for China (17): 9 documented on the device; 3 partial, or host or cloud only; 1 with attribution to the chip unverified; 4 not documented. Other catalogued · 18 3 documented on the device Other catalogued (18): 3 documented on the device; 1 partial, or host or cloud only; 5 with attribution to the chip unverified; 9 not documented. “Restricted for China” is the export status in Scrutica’s chip catalogue. Its 3 chips under case-by-case review (H200 SXM, H20 SXM and MI325X) are not among the 17. VendorAll vendorsNVIDIAAMDHuaweiIntelGoogleAmazonCerebrasTenstorrentMicrosoftMetaMoore ThreadsRestricted for China only - Documented on the device - Partial, or host or cloud only - Attribution to the chip unverified - Vendor states absent - Not documented A documented feature is a capability: these sources do not show whether it is enabled in deployment. Select any square to read what the vendor’s documents say about it.Select any accelerator for its eleven features, and any feature for what the vendor’s documents say about it. China export status, as the [chip catalogue](https://scrutica.com/export-controls/chips) names and defines it Restricted Case-by-Case Review Licence required, reviewed case by case, with conditions attached. Designed Compliant Designed to meet export-control thresholds. Domestic (China) Unassessed No China export classification is recorded for these chips.

		Root and boot			Identity and attestation		Protection in use				Location and remote disable
Accelerator	China export status
Documented on the device, of the 35		19	19	4	9	12	9	2	0	3	0	0
NVIDIA · sold · 51 documents read
A100 SXM	Restricted	Root of trust: documented	Secure boot: documented	Rollback protection: documented	Identity key: documented, unverified	Device attestation: partial	Confidential computing: partial	Multi-device protection: not documented	TEE-I/O: vendor states absent	Link encryption: partial, unverified	Location feature: partial	Position on remote disable: vendor states absent
A800	Restricted	Root of trust: not documented	Secure boot: not documented	Rollback protection: not documented	Identity key: not documented	Device attestation: not documented	Confidential computing: not documented	Multi-device protection: not documented	TEE-I/O: vendor states absent	Link encryption: not documented	Location feature: not documented	Position on remote disable: vendor states absent
H100 SXM5	Restricted	Root of trust: documented	Secure boot: documented	Rollback protection: documented, unverified	Identity key: documented	Device attestation: documented	Confidential computing: documented	Multi-device protection: partial	TEE-I/O: vendor states absent	Link encryption: partial	Location feature: partial	Position on remote disable: vendor states absent
H100 PCIe	Restricted	Root of trust: documented	Secure boot: documented	Rollback protection: documented, unverified	Identity key: documented	Device attestation: documented	Confidential computing: documented	Multi-device protection: vendor states absent	TEE-I/O: vendor states absent	Link encryption: partial	Location feature: partial	Position on remote disable: vendor states absent
H100 NVL	Restricted	Root of trust: documented	Secure boot: documented	Rollback protection: documented, unverified	Identity key: documented	Device attestation: documented	Confidential computing: documented	Multi-device protection: vendor states absent	TEE-I/O: vendor states absent	Link encryption: partial	Location feature: partial	Position on remote disable: vendor states absent
H800 SXM5	Restricted	Root of trust: documented	Secure boot: documented	Rollback protection: documented, unverified	Identity key: documented	Device attestation: documented	Confidential computing: documented	Multi-device protection: partial	TEE-I/O: vendor states absent	Link encryption: partial	Location feature: partial	Position on remote disable: vendor states absent
H800 PCIe	Restricted	Root of trust: documented	Secure boot: documented	Rollback protection: documented, unverified	Identity key: documented	Device attestation: documented	Confidential computing: documented	Multi-device protection: vendor states absent	TEE-I/O: vendor states absent	Link encryption: partial	Location feature: partial	Position on remote disable: vendor states absent
H200 SXM	Case-by-Case Review	Root of trust: documented	Secure boot: documented	Rollback protection: documented, unverified	Identity key: documented	Device attestation: documented	Confidential computing: documented	Multi-device protection: partial	TEE-I/O: vendor states absent	Link encryption: partial	Location feature: partial	Position on remote disable: vendor states absent
H20 SXM	Case-by-Case Review	Root of trust: documented	Secure boot: documented	Rollback protection: documented, unverified	Identity key: documented	Device attestation: documented	Confidential computing: documented	Multi-device protection: partial	TEE-I/O: vendor states absent	Link encryption: partial	Location feature: partial	Position on remote disable: vendor states absent
H20E (HBM3E refresh)	Unassessed	Root of trust: documented, unverified	Secure boot: documented, unverified	Rollback protection: documented, unverified	Identity key: documented, unverified	Device attestation: documented, unverified	Confidential computing: documented, unverified	Multi-device protection: partial, unverified	TEE-I/O: vendor states absent, unverified	Link encryption: partial, unverified	Location feature: partial, unverified	Position on remote disable: vendor states absent, unverified
B100	Restricted	Root of trust: documented	Secure boot: documented	Rollback protection: not documented	Identity key: not documented	Device attestation: not documented	Confidential computing: partial	Multi-device protection: not documented	TEE-I/O: partial	Link encryption: not documented	Location feature: partial	Position on remote disable: vendor states absent
B200	Restricted	Root of trust: documented	Secure boot: documented	Rollback protection: documented, unverified	Identity key: documented	Device attestation: documented	Confidential computing: documented	Multi-device protection: documented	TEE-I/O: partial	Link encryption: documented	Location feature: partial	Position on remote disable: vendor states absent
B300	Restricted	Root of trust: documented	Secure boot: documented	Rollback protection: documented, unverified	Identity key: documented	Device attestation: documented	Confidential computing: documented	Multi-device protection: documented	TEE-I/O: partial	Link encryption: documented	Location feature: partial	Position on remote disable: vendor states absent
B30A	Designed Compliant	Root of trust: not documented	Secure boot: not documented	Rollback protection: not documented	Identity key: not documented	Device attestation: not documented	Confidential computing: not documented	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: not documented	Location feature: not documented	Position on remote disable: vendor states absent
Rubin GPU	Restricted	Root of trust: documented	Secure boot: not documented	Rollback protection: not documented	Identity key: not documented	Device attestation: partial	Confidential computing: partial	Multi-device protection: partial	TEE-I/O: not documented	Link encryption: partial	Location feature: partial	Position on remote disable: vendor states absent
AMD · sold · 59 documents read
MI300X	Restricted	Root of trust: documented	Secure boot: documented	Rollback protection: documented	Identity key: partial, unverified	Device attestation: partial	Confidential computing: not documented	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: not documented	Location feature: not documented	Position on remote disable: not documented
MI325X	Case-by-Case Review	Root of trust: documented	Secure boot: documented	Rollback protection: not documented	Identity key: not documented	Device attestation: not documented	Confidential computing: not documented	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: not documented	Location feature: not documented	Position on remote disable: not documented
MI350X	Restricted	Root of trust: documented	Secure boot: documented	Rollback protection: not documented	Identity key: partial	Device attestation: documented	Confidential computing: not documented	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: partial	Location feature: not documented	Position on remote disable: not documented
MI355X	Restricted	Root of trust: documented	Secure boot: documented	Rollback protection: not documented	Identity key: partial	Device attestation: documented	Confidential computing: not documented	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: partial	Location feature: not documented	Position on remote disable: not documented
MI400	Restricted	Root of trust: documented, unverified	Secure boot: documented	Rollback protection: not documented	Identity key: documented, unverified	Device attestation: documented, unverified	Confidential computing: host or cloud only, unverified	Multi-device protection: host or cloud only, unverified	TEE-I/O: documented, unverified	Link encryption: documented	Location feature: not documented	Position on remote disable: not documented
Huawei · sold in China · 10 documents read
Ascend 910B	Domestic (China)	Root of trust: not documented	Secure boot: not documented	Rollback protection: not documented	Identity key: documented, unverified	Device attestation: not documented	Confidential computing: not documented	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: partial, unverified	Location feature: not documented	Position on remote disable: not documented
Ascend 910C	Domestic (China)	Root of trust: not documented	Secure boot: not documented	Rollback protection: not documented	Identity key: partial, unverified	Device attestation: not documented	Confidential computing: not documented	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: partial, unverified	Location feature: not documented	Position on remote disable: not documented
Ascend 910D	Domestic (China)	Root of trust: not documented	Secure boot: not documented	Rollback protection: not documented	Identity key: not documented	Device attestation: not documented	Confidential computing: not documented	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: not documented	Location feature: not documented	Position on remote disable: not documented
Ascend 920	Domestic (China)	Root of trust: not documented	Secure boot: not documented	Rollback protection: not documented	Identity key: not documented	Device attestation: not documented	Confidential computing: not documented	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: not documented	Location feature: not documented	Position on remote disable: not documented
Intel · sold · 14 documents read
Gaudi 2	Restricted	Root of trust: documented	Secure boot: documented	Rollback protection: documented	Identity key: not documented	Device attestation: not documented	Confidential computing: not documented	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: not documented	Location feature: not documented	Position on remote disable: not documented
Gaudi 3	Restricted	Root of trust: documented	Secure boot: documented	Rollback protection: documented	Identity key: not documented	Device attestation: not documented	Confidential computing: not documented	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: not documented	Location feature: not documented	Position on remote disable: not documented
Google · built for its cloud · 22 documents read
TPU v4	Unassessed	Root of trust: host or cloud only, unverified	Secure boot: host or cloud only, unverified	Rollback protection: host or cloud only, unverified	Identity key: host or cloud only, unverified	Device attestation: host or cloud only, unverified	Confidential computing: vendor states absent from its cloud service	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: not documented	Location feature: not documented	Position on remote disable: not documented
TPU v5e	Unassessed	Root of trust: host or cloud only, unverified	Secure boot: host or cloud only, unverified	Rollback protection: host or cloud only, unverified	Identity key: host or cloud only, unverified	Device attestation: host or cloud only, unverified	Confidential computing: vendor states absent from its cloud service	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: not documented	Location feature: not documented	Position on remote disable: not documented
TPU v5p	Unassessed	Root of trust: host or cloud only, unverified	Secure boot: host or cloud only, unverified	Rollback protection: host or cloud only, unverified	Identity key: host or cloud only, unverified	Device attestation: host or cloud only, unverified	Confidential computing: vendor states absent from its cloud service	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: not documented	Location feature: not documented	Position on remote disable: not documented
TPU v6e (Trillium)	Unassessed	Root of trust: host or cloud only, unverified	Secure boot: host or cloud only, unverified	Rollback protection: host or cloud only, unverified	Identity key: host or cloud only, unverified	Device attestation: host or cloud only, unverified	Confidential computing: vendor states absent from its cloud service	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: not documented	Location feature: not documented	Position on remote disable: not documented
TPU v7 (Ironwood)	Unassessed	Root of trust: documented	Secure boot: documented	Rollback protection: host or cloud only, unverified	Identity key: partial, unverified	Device attestation: documented	Confidential computing: vendor states absent from its cloud service	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: not documented	Location feature: not documented	Position on remote disable: not documented
Amazon · built for its cloud · 16 documents read
Trainium3	Unassessed	Root of trust: host or cloud only	Secure boot: host or cloud only	Rollback protection: not documented	Identity key: not documented	Device attestation: host or cloud only	Confidential computing: host or cloud only	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: host or cloud only	Location feature: not documented	Position on remote disable: not documented
Cerebras · sold · 15 documents read
WSE-3	Unassessed	Root of trust: not documented	Secure boot: not documented	Rollback protection: not documented	Identity key: not documented	Device attestation: not documented	Confidential computing: not documented	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: not documented	Location feature: not documented	Position on remote disable: not documented
Tenstorrent · sold · 39 documents read
Wormhole n300	Unassessed	Root of trust: not documented	Secure boot: not documented	Rollback protection: not documented	Identity key: not documented	Device attestation: not documented	Confidential computing: not documented	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: not documented	Location feature: not documented	Position on remote disable: not documented
Blackhole p150	Unassessed	Root of trust: not documented	Secure boot: partial	Rollback protection: partial	Identity key: not documented	Device attestation: not documented	Confidential computing: not documented	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: not documented	Location feature: not documented	Position on remote disable: not documented

27 accelerators outside Scrutica’s chip catalogue

		Root and boot			Identity and attestation		Protection in use				Location and remote disable
Accelerator	China export status
NVIDIA · sold · 51 documents read
A100 80GB PCIe	Not in the catalogue	Root of trust: documented	Secure boot: documented	Rollback protection: documented	Identity key: documented, unverified	Device attestation: partial	Confidential computing: partial	Multi-device protection: not documented	TEE-I/O: vendor states absent	Link encryption: partial, unverified	Location feature: partial	Position on remote disable: vendor states absent
H800 NVL	Not in the catalogue	Root of trust: documented	Secure boot: documented	Rollback protection: documented, unverified	Identity key: documented	Device attestation: documented	Confidential computing: documented	Multi-device protection: vendor states absent	TEE-I/O: vendor states absent	Link encryption: partial	Location feature: partial	Position on remote disable: vendor states absent
H200 NVL	Not in the catalogue	Root of trust: documented	Secure boot: documented	Rollback protection: documented, unverified	Identity key: documented	Device attestation: documented	Confidential computing: documented	Multi-device protection: vendor states absent	TEE-I/O: vendor states absent	Link encryption: partial	Location feature: partial	Position on remote disable: vendor states absent
GH200 superchip	Not in the catalogue	Root of trust: documented	Secure boot: documented	Rollback protection: not documented	Identity key: not documented	Device attestation: not documented	Confidential computing: not documented	Multi-device protection: not documented	TEE-I/O: vendor states absent	Link encryption: not documented	Location feature: partial	Position on remote disable: vendor states absent
L40S	Not in the catalogue	Root of trust: documented	Secure boot: documented	Rollback protection: not documented	Identity key: not documented	Device attestation: not documented	Confidential computing: vendor states absent	Multi-device protection: vendor states absent	TEE-I/O: vendor states absent	Link encryption: not documented	Location feature: partial	Position on remote disable: vendor states absent
GB200 superchip	Not in the catalogue	Root of trust: documented	Secure boot: documented	Rollback protection: not documented	Identity key: not documented	Device attestation: not documented	Confidential computing: not documented	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: not documented	Location feature: partial	Position on remote disable: vendor states absent
GB300 superchip	Not in the catalogue	Root of trust: documented	Secure boot: documented	Rollback protection: not documented	Identity key: not documented	Device attestation: not documented	Confidential computing: not documented	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: not documented	Location feature: partial	Position on remote disable: vendor states absent
RTX PRO 6000 Server	Not in the catalogue	Root of trust: documented	Secure boot: documented	Rollback protection: documented, unverified	Identity key: documented	Device attestation: documented	Confidential computing: documented	Multi-device protection: not documented	TEE-I/O: partial	Link encryption: documented	Location feature: partial	Position on remote disable: vendor states absent
AMD · sold · 59 documents read
MI250X	Not in the catalogue	Root of trust: documented	Secure boot: documented	Rollback protection: not documented	Identity key: not documented	Device attestation: not documented	Confidential computing: not documented	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: not documented	Location feature: not documented	Position on remote disable: not documented
MI300A	Not in the catalogue	Root of trust: documented	Secure boot: documented	Rollback protection: not documented	Identity key: not documented	Device attestation: not documented	Confidential computing: not documented	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: not documented	Location feature: not documented	Position on remote disable: not documented
MI308X	Not in the catalogue	Root of trust: not documented	Secure boot: not documented	Rollback protection: not documented	Identity key: not documented	Device attestation: not documented	Confidential computing: not documented	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: not documented	Location feature: not documented	Position on remote disable: not documented
Huawei · sold in China · 10 documents read
Ascend 910A	Not in the catalogue	Root of trust: documented	Secure boot: documented	Rollback protection: partial	Identity key: documented	Device attestation: not documented	Confidential computing: not documented	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: not documented	Location feature: not documented	Position on remote disable: not documented
Google · built for its cloud · 22 documents read
TPU 8t	Not in the catalogue	Root of trust: not documented	Secure boot: not documented	Rollback protection: not documented	Identity key: not documented	Device attestation: not documented	Confidential computing: vendor states absent from its cloud service	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: not documented	Location feature: not documented	Position on remote disable: not documented
TPU 8i	Not in the catalogue	Root of trust: not documented	Secure boot: not documented	Rollback protection: not documented	Identity key: not documented	Device attestation: not documented	Confidential computing: vendor states absent from its cloud service	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: not documented	Location feature: not documented	Position on remote disable: not documented
Amazon · built for its cloud · 16 documents read
Trainium	Not in the catalogue	Root of trust: host or cloud only	Secure boot: host or cloud only	Rollback protection: not documented	Identity key: not documented	Device attestation: host or cloud only	Confidential computing: host or cloud only	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: host or cloud only	Location feature: not documented	Position on remote disable: not documented
Inferentia2	Not in the catalogue	Root of trust: host or cloud only	Secure boot: host or cloud only	Rollback protection: not documented	Identity key: not documented	Device attestation: host or cloud only	Confidential computing: host or cloud only	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: not documented	Location feature: not documented	Position on remote disable: not documented
Trainium2	Not in the catalogue	Root of trust: host or cloud only	Secure boot: host or cloud only	Rollback protection: not documented	Identity key: not documented	Device attestation: host or cloud only	Confidential computing: host or cloud only	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: host or cloud only	Location feature: not documented	Position on remote disable: not documented
Cerebras · sold · 15 documents read
WSE-3 Turbo	Not in the catalogue	Root of trust: not documented	Secure boot: not documented	Rollback protection: not documented	Identity key: not documented	Device attestation: not documented	Confidential computing: not documented	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: not documented	Location feature: not documented	Position on remote disable: not documented
Microsoft · built for its own services · 15 documents read
Maia 100	Not in the catalogue	Root of trust: host or cloud only, unverified	Secure boot: host or cloud only, unverified	Rollback protection: not documented	Identity key: not documented	Device attestation: not documented	Confidential computing: not documented	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: documented	Location feature: not documented	Position on remote disable: not documented
Maia 200	Not in the catalogue	Root of trust: documented, unverified	Secure boot: host or cloud only, unverified	Rollback protection: not documented	Identity key: not documented	Device attestation: not documented	Confidential computing: not documented	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: documented, unverified	Location feature: not documented	Position on remote disable: not documented
Meta · built for its own services · 8 documents read
MTIA v1	Not in the catalogue	Root of trust: not documented	Secure boot: partial	Rollback protection: not documented	Identity key: not documented	Device attestation: not documented	Confidential computing: not documented	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: not documented	Location feature: not documented	Position on remote disable: not documented
MTIA v2	Not in the catalogue	Root of trust: not documented	Secure boot: partial	Rollback protection: not documented	Identity key: not documented	Device attestation: not documented	Confidential computing: not documented	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: not documented	Location feature: not documented	Position on remote disable: not documented
MTIA 300	Not in the catalogue	Root of trust: not documented	Secure boot: partial	Rollback protection: not documented	Identity key: not documented	Device attestation: not documented	Confidential computing: not documented	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: not documented	Location feature: not documented	Position on remote disable: not documented
MTIA 400	Not in the catalogue	Root of trust: not documented	Secure boot: not documented	Rollback protection: not documented	Identity key: not documented	Device attestation: not documented	Confidential computing: not documented	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: not documented	Location feature: not documented	Position on remote disable: not documented
MTIA 450	Not in the catalogue	Root of trust: not documented	Secure boot: not documented	Rollback protection: not documented	Identity key: not documented	Device attestation: not documented	Confidential computing: not documented	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: not documented	Location feature: not documented	Position on remote disable: not documented
MTIA 500	Not in the catalogue	Root of trust: not documented	Secure boot: not documented	Rollback protection: not documented	Identity key: not documented	Device attestation: not documented	Confidential computing: not documented	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: not documented	Location feature: not documented	Position on remote disable: not documented
Moore Threads · sold in China · 4 documents read
MTT S4000	Not in the catalogue	Root of trust: not documented	Secure boot: not documented	Rollback protection: not documented	Identity key: not documented	Device attestation: not documented	Confidential computing: partial	Multi-device protection: not documented	TEE-I/O: not documented	Link encryption: not documented	Location feature: not documented	Position on remote disable: not documented

How to read the register The 11 features run in stack order, from the root of trust to the position on remote disable. The register’s 35 accelerators are the chips in Scrutica’s chip catalogue, with their export status for China from it; the 27 researched outside the catalogue have no export status and are listed apart. Vendors follow the catalogue’s order, with those that have no chip in it last; each heading gives how the vendor’s parts are offered and how many of its documents were read. A feature documented only for the host server or the cloud service counts as partial. A cell is dotted, and not counted as documented, where its attribution or its application to the chip is unverified, or where only press or analyst sources describe the feature. A feature documented for an architecture counts on that architecture’s chips. The location and remote-disable columns record what vendors state. Where a vendor states that a feature is absent, public documents cannot confirm it. On Hopper, NVIDIA documents attestation only in confidential-computing mode. Device attestation, across the 35 catalogued accelerators - 12 documented on the device: 9 of the 17 restricted for China, 3 of the 18 others - 4 partial, or host or cloud only - 6 with attribution to the chip unverified - 13 not documented Select any square to read what the vendor’s documents say about it. Each bar is all the compute shipped, split by what the documents say: - Documented on the device - Attribution to the chip unverified - Partial, or host or cloud only - One estimate covering chips documented differently - Vendor states absent - Vendor states absent from its cloud service - Not documented Share documented on the device The feature the chip-type table lists - Root of trust75% - Secure boot75% - Rollback protection5% - Identity key22% - Device attestation27%43%: B200, B300 and H20 - Confidential computing22% - Multi-device protection0% - TEE-I/O0% - Link encryption0% - Location feature0% - Position on remote disable0% Chip types forRoot of trustSecure bootRollback protectionIdentity keyDevice attestationConfidential computingMulti-device protectionTEE-I/OLink encryptionLocation featurePosition on remote disable Italic: outside Scrutica’s chip catalogue. Device attestation: 24 chip types, largest first, in thousands of H100-equivalents Italic: outside Scrutica’s chip catalogue. | Chip type | Documentation | H100-equivalents (thousands) | | --- | --- | --- | | NVIDIA B200 | One estimate covering chips documented differently | 4,847 | | NVIDIA H100 | Documented on the device | 4,317 | | NVIDIA B300 | One estimate covering chips documented differently | 3,792 | | Google TPU v6e | Attribution to the chip unverified | 2,291 | | [Amazon Trainium2](https://scrutica.com/verification/hardware#hw-aws-trainium2--device_attestation) | Partial, or host or cloud only | 919 | 19 more chip types Device attestation: chip types 6 to 24 | Chip type | Documentation | H100-equivalents (thousands) | | --- | --- | --- | | AMD Instinct MI300X | Partial, or host or cloud only | 695 | | Google TPU v7 | Documented on the device | 687 | | Huawei Ascend 910C | Not documented | 499 | | Google TPU v5e | Attribution to the chip unverified | 475 | | Google TPU v5p | Attribution to the chip unverified | 312 | | NVIDIA A100 | Partial, or host or cloud only | 303 | | NVIDIA H20 | One estimate covering chips documented differently | 223 | | Huawei Ascend 910B | Not documented | 209 | | AMD Instinct MI350X | Documented on the device | 192 | | AMD Instinct MI355X | Documented on the device | 124 | | NVIDIA H800 | Documented on the device | 115 | | [AMD Instinct MI300A](https://scrutica.com/verification/hardware#hw-amd-mi300a--device_attestation) | Not documented | 114 | | AMD Instinct MI325X | Not documented | 104 | | [Amazon Trainium1](https://scrutica.com/verification/hardware#hw-aws-trainium1--device_attestation) | Partial, or host or cloud only | 94 | | NVIDIA A800 | Not documented | 38 | | Google TPU v4 | Attribution to the chip unverified | 36 | | [AMD Instinct MI308X](https://scrutica.com/verification/hardware#hw-amd-mi308x--device_attestation) | Not documented | 10 | | [AMD Instinct MI250X](https://scrutica.com/verification/hardware#hw-amd-mi250x--device_attestation) | Not documented | 9 | | Google TPU v4i | Not in the register | 2 | The accelerators each chip type counts (24) - NVIDIA B200 counts two rows that read differently: B200, documented on the device; _GB200 superchip_, not documented. Epoch links this series to its GB200 hardware entry; NVIDIA documents HGX B200 boards and does not name GB200 systems, so both rows count. - NVIDIA H100 counts H100 SXM5, H100 PCIe and H100 NVL. - NVIDIA B300 counts two rows that read differently: B300, documented on the device; _GB300 superchip_, not documented. Epoch links this series to its GB300 hardware entry; NVIDIA documents HGX B300 boards and does not name GB300 systems, so both rows count. - Google TPU v6e counts TPU v6e (Trillium). - Amazon Trainium2 counts _Trainium2_. - AMD Instinct MI300X counts MI300X. - Google TPU v7 counts TPU v7 (Ironwood). - Huawei Ascend 910C counts Ascend 910C. - Google TPU v5e counts TPU v5e. - Google TPU v5p counts TPU v5p. - NVIDIA A100 counts A100 SXM and _A100 80GB PCIe_. - NVIDIA H20 counts two rows that read differently: H20 SXM, documented on the device; H20E (HBM3E refresh), attribution to the chip unverified. - Huawei Ascend 910B counts Ascend 910B. - AMD Instinct MI350X counts MI350X. - AMD Instinct MI355X counts MI355X. - NVIDIA H800 counts H800 SXM5, H800 PCIe and _H800 NVL_. - AMD Instinct MI300A counts _MI300A_. - AMD Instinct MI325X counts MI325X. - Amazon Trainium1 counts _Trainium_. - NVIDIA A800 counts A800. - Google TPU v4 counts TPU v4. - AMD Instinct MI308X counts _MI308X_. - AMD Instinct MI250X counts _MI250X_. - Google TPU v4i has no row in the register. How shipments are weighted Epoch AI’s median estimates of compute shipped from each designer’s series start (2022 to 2024) through 31 Dec 2025: 20.4 million H100-equivalents over 24 chip types. Series start by designer: NVIDIA’s from 2022, Google’s from 2023 and AMD’s, Huawei’s and Amazon’s from 2024. The medians are added; Epoch’s 90% intervals (5th to 95th percentile) are left out. Accelerator models are recorded for 9 facilities (enrichment records of 19 Apr 2026), against 4,266 published in the facility register (snapshot 3 Oct 2026). That coverage is too sparse to weight by installed hardware; shipment estimates supply the weights. Cambricon Siyuan 590 (38 thousand H100-equivalents) is left out: Epoch’s series for Cambricon restarts each year, and only series with a single starting point enter this total. Source: [Epoch AI, AI Chip Sales](https://epoch.ai/data/ai-chip-sales) (CC BY 4.0, snapshot 19 Apr 2026). - Documented - Partial, or host or cloud only - Early access to release - Announced - Removed - Export control - Earlier - Attribution to the chip unverified Italic: outside Scrutica’s chip catalogue Not drawn: Huawei and Moore Threads, which have no dated chip in the register; Cerebras, whose register rows have no feature documented on the device. Label every event Previous8 of 31Next 29 Jul 2023 · NVIDIA Hopper · Released Attestation service live for H100 Device attestation Quotation and source > General Availability (GA) release of the NVIDIA Remote Attestation Service (NRAS). Provides GPU attestation support for H100 GPUs. Generates signed EAT (Entity Attestation Token) as the attestation response. Source (1) 1. [Remote Attestation Service: Releases](https://docs.nvidia.com/attestation/cloud-services/latest/nras/nras_releases.html) · primary · 29 Jul 2023 to 5 Aug 2025 (release entries) 29 Jul 2023, NVIDIA Hopper: Attestation service live for H100 How the marks are dated The first documented security features by architecture and vendor, beside the export-control actions in Scrutica’s chip catalogue. NVIDIA’s marks are its dated releases. For other vendors, each mark is the first chip to document a feature, at the date of the chip’s launch, availability or deployment; the document can be later. A ring marks a chip that had only been announced. Marks of one lane that would touch are set side by side, a little off their dates, or on a phone on a second line in the lane first; the card gives each date as its source does. All 31 events as a list Italic: outside Scrutica’s chip catalogue 1. NVIDIA AmpereVolta V100 generationBefore the axis: firmware authenticated since Volta V100 [Source](https://developer.nvidia.com/blog/confidential-computing-on-h100-gpus-for-secure-and-trustworthy-ai/) 2. AMD Instinct8 Nov 2021MI250X: root of trust, secure boot [Source](https://www.amd.com/en/products/accelerators/instinct/mi200/mi250x.html) 3. NVIDIA Ampere25 Feb 2022A100 PCIe variants without firmware attestation [Source](https://www.colfax-intl.com/downloads/NVIDIA-End-Customer-Communication-Data-Center-CEC-Defeaturing.pdf) 4. NVIDIA Ampere22 Mar 2022Confidential VM preview with A100 (Azure) [Source](https://azure.microsoft.com/en-us/blog/azure-confidential-computing-with-nvidia-gpus-for-trustworthy-ai/) 5. Intel Gaudi10 May 2022Gaudi 2: root of trust, secure boot, rollback protection [Source](https://habana.ai/press/habana-labs-launches-second-generation-ai-deep-learning-processors/) 6. Export controls21 Oct 2022October 2022 Advanced Computing Rule [87 FR 62186](https://www.govinfo.gov/content/pkg/FR-2022-10-13/pdf/2022-21658.pdf) 7. Meta MTIA18 May 2023MTIA v1: partial secure boot [Source](https://ai.meta.com/blog/meta-training-inference-accelerator-AI-MTIA/) 8. NVIDIA Hopper29 Jul 2023Attestation service live for H100 [Source](https://docs.nvidia.com/attestation/cloud-services/latest/nras/nras_releases.html) 9. NVIDIA Hopper3 Aug 2023 – Feb 2024Confidential computing, early access to GA [Source](https://developer.nvidia.com/blog/confidential-computing-on-h100-gpus-for-secure-and-trustworthy-ai/) 10. Microsoft Maia15 Nov 2023Maia 100: link encryption [Source](https://news.microsoft.com/source/features/ai/in-house-chips-silicon-to-service-to-meet-ai-demand/) 11. Export controls17 Nov 2023October 2023 Advanced Computing Update [88 FR 73458](https://www.federalregister.gov/documents/2023/10/25/2023-23055/implementation-of-additional-export-controls-certain-advanced-computing-items-supercomputer-and) 12. AMD Instinct6 Dec 2023MI300X: rollback protection, partial identity key (unverified), partial device attestation [Source](https://www.amd.com/en/products/accelerators/instinct/mi300/mi300x.html) 13. NVIDIA HopperJun 2024 – 23 Apr 2025Multi-GPU, NVLink unencrypted [Source](https://docs.nvidia.com/550trd3-nvidia-trusted-computing-solutions-release-notes.pdf) 14. Export controls2 Dec 2024December 2024 Expansion [89 FR 96790](https://www.govinfo.gov/content/pkg/FR-2024-12-05/pdf/2024-28270.pdf) 15. NVIDIA Blackwell29 Jan 2025Attestation service adds Blackwell [Source](https://docs.nvidia.com/attestation/cloud-services/latest/nras/nras_releases.html) 16. Tenstorrent3 Apr 2025Blackhole p150: partial secure boot, partial rollback protection [Source](https://tenstorrent.com/vision/tenstorrent-launches-blackhole-developer-products-at-tenstorrent-dev-day) 17. Export controls9 Apr 2025Licence required for H20 [Source](https://www.sec.gov/Archives/edgar/data/1045810/000104581025000082/nvda-20250409.htm) 18. NVIDIA BlackwellMay–Aug 2025Confidential computing on B200 [Source](https://docs.nvidia.com/575trd1-trusted-computing-solutions-release-notes.pdf) 19. AMD Instinct12 Jun 2025MI350X, MI355X: device attestation, partial link encryption [Source](https://www.amd.com/en/products/accelerators/instinct/mi350/mi350x.html) 20. Export controls14 Jul 2025H20 licence applications [Source](https://blogs.nvidia.com/blog/nvidia-ceo-promotes-ai-in-dc-and-china/) 21. Amazon23 Sep 2025EC2 instance attestation (cloud level) [Source](https://aws.amazon.com/about-aws/whats-new/2025/09/aws-announces-ec2-instance-attestation/) 22. NVIDIA BlackwellDec 2025Multi-GPU with encrypted NVLink [Source](https://docs.nvidia.com/590trd1-trusted-computing-solutions-release-notes.pdf) 23. NVIDIA Blackwell10 Dec 2025 – 11 May 2026Fleet service, operator-declared location [Source](https://blogs.nvidia.com/blog/optional-data-center-fleet-management-software/) 24. NVIDIA Rubin5 Jan 2026Rack-scale confidential computing announced [Source](https://nvidianews.nvidia.com/news/rubin-platform-ai-supercomputer) 25. Export controls15 Jan 2026January 2026 License Review Policy Revision [91 FR 1684](https://www.federalregister.gov/documents/2026/01/15/2026-00789/revision-to-license-review-policy-for-advanced-computing-commodities) 26. Microsoft Maia26 Jan 2026Maia 200: root of trust (unverified) [Source](https://blogs.microsoft.com/blog/2026/01/26/maia-200-the-ai-accelerator-built-for-inference/) 27. NVIDIA Blackwell4 Feb 2026Attestation outside confidential-computing mode [Source](https://github.com/NVIDIA/attestation-sdk/releases) 28. Google TPU31 Mar 2026TPU v7 (Ironwood): root of trust, secure boot, partial identity key (unverified), device attestation [Source](https://docs.cloud.google.com/tpu/docs/release-notes) 29. NVIDIA Hopper11 May 2026Fleet service, operator-declared location [Source](https://developer.nvidia.com/blog/introducing-nvidia-fleet-intelligence-for-real-time-gpu-fleet-visibility-and-optimization/) 30. NVIDIA Rubin11 May 2026Fleet service, operator-declared location [Source](https://developer.nvidia.com/blog/introducing-nvidia-fleet-intelligence-for-real-time-gpu-fleet-visibility-and-optimization/) 31. AMD Instinct23 Jul 2026MI400: identity key (unverified), TEE-I/O (unverified), link encryption [Source](https://newsroom.amd.com/news/aai-2026-mi400-instinct-update/) From NVIDIA’s white papers, for a GPU in a confidential virtual machine (CVM), and AMD’s and Intel’s documents for the CPU. The CA is NVIDIA’s certificate authority. 1. Factory and CA 2. GPU 3. Host 4. CVM 5. Verifier 6. Relying party 1. #### 0 At the factory: the GPU gets an identity NVIDIA factory and CA → GPU: Fused identity key, certified by NVIDIA’s CA A device-unique private key is fused into each GPU and NVIDIA's certificate authority (CA) certifies the public key. All later steps trust this. Scrutica’s inference IEEE 802.1AR is listed as the standard for a device identity its supplier installs (an IDevID); neither NVIDIA white paper cited on this path mentions it. Sources (3) Specified by [NVIDIA GPU confidential computing](https://images.nvidia.com/aem-dam/en-zz/Solutions/data-center/HCC-Whitepaper-v1.0.pdf), [IEEE 802.1AR device identity](https://standards.ieee.org/ieee/802.1AR/6995/). Source: [Confidential Compute on NVIDIA Hopper H100 (WP-11459-001 v1.0)](https://images.nvidia.com/aem-dam/en-zz/Solutions/data-center/HCC-Whitepaper-v1.0.pdf), p. 22. 2. #### 1 At power-on: secure boot GPU, on its own: Root of trust checks NVIDIA’s firmware signature The GPU's on-die root of trust only runs NVIDIA-signed firmware and locks it until reset. Scrutica’s inference DICE (the Device Identifier Composition Engine of the Trusted Computing Group, TCG) and Caliptra (an open root of trust for data-centre chips, specified by the Open Compute Project) are listed as the standards-based pattern for this step; NVIDIA's documents do not say Hopper/Blackwell use Caliptra or TCG DICE. Sources (4) Specified by [NVIDIA GPU confidential computing](https://images.nvidia.com/aem-dam/en-zz/Solutions/data-center/HCC-Whitepaper-v1.0.pdf), [TCG DICE hardware requirements](https://trustedcomputinggroup.org/resource/hardware-requirements-for-a-device-identifier-composition-engine/), [OCP Caliptra](https://www.opencompute.org/documents/caliptra-2-0-pdf). Source: [NVIDIA Secure AI with Blackwell and Hopper GPUs (White Paper WP-12554-001\_v1.3)](https://docs.nvidia.com/nvidia-secure-ai-with-blackwell-and-hopper-gpus-whitepaper.pdf), p. 7. 3. #### 2 The host switches on confidential computing Host and hypervisor → GPU: Mode switch; the GPU resets and scrubs memory Switching confidential-computing mode requires a GPU reset; memory is locked and scrubbed before hand-over. Sources (2) Specified by [NVIDIA GPU confidential computing](https://images.nvidia.com/aem-dam/en-zz/Solutions/data-center/HCC-Whitepaper-v1.0.pdf). Source: [Confidential Compute on NVIDIA Hopper H100 (WP-11459-001 v1.0)](https://images.nvidia.com/aem-dam/en-zz/Solutions/data-center/HCC-Whitepaper-v1.0.pdf), p. 21. 4. #### 3 The host launches a confidential virtual machine Host and hypervisor → Confidential VM: TDX trust domain or SEV-SNP guest The CVM runs as an Intel TDX (Trust Domain Extensions) trust domain or an AMD SEV-SNP (Secure Encrypted Virtualization with Secure Nested Paging) guest, memory encrypted and isolated from the hypervisor. Sources (4) Specified by [Intel TDX](https://cc-enabling.trustedservices.intel.com/intel-tdx-enabling-guide/03/hardware_selection/), [AMD SEV-SNP](https://www.amd.com/en/developer/sev.html), [Confidential Computing Consortium definition](https://confidentialcomputing.io/). Source: [NVIDIA Secure AI with Blackwell and Hopper GPUs (White Paper WP-12554-001\_v1.3)](https://docs.nvidia.com/nvidia-secure-ai-with-blackwell-and-hopper-gpus-whitepaper.pdf), p. 10. 5. #### 4 The GPU is assigned to the virtual machine Host and hypervisor → Confidential VM: GPU assigned; bounce buffers, with TDISP and IDE a future option on TEE-I/O platforms Hopper: data moves through encrypted bounce buffers in shared memory. Blackwell, on TEE-I/O platforms (which extend the trusted execution environment, or TEE, to include the GPU) NVIDIA describes as a future option: TDISP (the TEE Device Interface Security Protocol) locks the device interface into the CVM and IDE (Integrity and Data Encryption) encrypts the PCIe (PCI Express) link. > In the future, as an alternative to the bounce buffer method, inline encryption can be used with systems that have Blackwell B100/B200 GPUs and TDISP/IDE-compatible CPUs. [NVIDIA Secure AI with Blackwell and Hopper GPUs (White Paper WP-12554-001\_v1.3)](https://docs.nvidia.com/nvidia-secure-ai-with-blackwell-and-hopper-gpus-whitepaper.pdf), p. 21 (printed); PDF p. 25 · primary · 14 Aug 2025 Later or contrary evidence On 22 Sep 2026 NVIDIA wrote that, under B200 confidential computing, host-to-device transfers pass through a software-encrypted bounce buffer. [Enabling Private High-Performance Production AI Inference with NVIDIA Confidential Computing](https://developer.nvidia.com/blog/enabling-private-high-performance-production-ai-inference-with-nvidia-confidential-computing/) · primary Sources (6) Specified by [NVIDIA GPU confidential computing](https://images.nvidia.com/aem-dam/en-zz/Solutions/data-center/HCC-Whitepaper-v1.0.pdf), [PCI-SIG TDISP](https://pcisig.com/PCI%20Express/ECN/Base/TEEDeviceInterfaceSecurityProtocol), [PCI-SIG IDE](https://pcisig.com/PCI%20Express/ECN/Base/IntegrityandDataEncryption), [Intel TDX](https://cc-enabling.trustedservices.intel.com/intel-tdx-enabling-guide/03/hardware_selection/), [AMD SEV-SNP](https://www.amd.com/en/developer/sev.html). Source: [NVIDIA Secure AI with Blackwell and Hopper GPUs (White Paper WP-12554-001\_v1.3)](https://docs.nvidia.com/nvidia-secure-ai-with-blackwell-and-hopper-gpus-whitepaper.pdf), p. 8. 6. #### 5 Driver and GPU open an encrypted session Confidential VM ↔ GPU: Versions, algorithms, certificates, key exchange Inside the CVM the NVIDIA driver negotiates with GPU firmware (version, capabilities, algorithms and certificates under SPDM, the Security Protocol and Data Model) and runs a Diffie-Hellman key exchange to establish an encrypted channel. Sources (4) Specified by [DMTF SPDM (DSP0274)](https://www.dmtf.org/dsp/DSP0274), [DMTF Secured Messages (DSP0277)](https://www.dmtf.org/dsp/DSP0277), [NVIDIA GPU confidential computing](https://images.nvidia.com/aem-dam/en-zz/Solutions/data-center/HCC-Whitepaper-v1.0.pdf). Source: [NVIDIA Secure AI with Blackwell and Hopper GPUs (White Paper WP-12554-001\_v1.3)](https://docs.nvidia.com/nvidia-secure-ai-with-blackwell-and-hopper-gpus-whitepaper.pdf), p. 11. 7. #### 6 The GPU produces evidence GPU → Verifier: Measurements signed over the verifier’s nonce A fresh nonce (a number used once) from the verifier goes to the GPU; the GPU returns SPDM MEASUREMENTS signed with an attestation key that its device identity key has certified. Sources (4) Specified by [DMTF SPDM (DSP0274)](https://www.dmtf.org/dsp/DSP0274), [NVIDIA GPU confidential computing](https://images.nvidia.com/aem-dam/en-zz/Solutions/data-center/HCC-Whitepaper-v1.0.pdf), [IETF RATS architecture (RFC 9334)](https://www.rfc-editor.org/rfc/rfc9334). Source: [Confidential Compute on NVIDIA Hopper H100 (WP-11459-001 v1.0)](https://images.nvidia.com/aem-dam/en-zz/Solutions/data-center/HCC-Whitepaper-v1.0.pdf), p. 24, step 4. 8. #### 7 A verifier appraises the GPU Verifier, on its own: Chain to NVIDIA’s root, revocation, nonce, reference manifest NVIDIA’s Remote Attestation Service (NRAS), or the local verifier, checks the certificate chain to NVIDIA's root and its revocation status, the signature and nonce, then compares the measurements with NVIDIA's signed reference integrity manifest (RIM) for that driver and GPU model. Sources (4) Specified by [IETF RATS architecture (RFC 9334)](https://www.rfc-editor.org/rfc/rfc9334), [TCG Reference Integrity Manifest](https://trustedcomputinggroup.org/resource/tcg-reference-integrity-manifest-rim-information-model/), [NVIDIA GPU confidential computing](https://images.nvidia.com/aem-dam/en-zz/Solutions/data-center/HCC-Whitepaper-v1.0.pdf). Source: [Confidential Compute on NVIDIA Hopper H100 (WP-11459-001 v1.0)](https://images.nvidia.com/aem-dam/en-zz/Solutions/data-center/HCC-Whitepaper-v1.0.pdf), p. 25, step 18. 9. #### 8 The verifier issues an attestation result Verifier → Relying party: Signed Entity Attestation Token, directly or by way of the VM The result is a signed Entity Attestation Token (EAT) in JSON Web Token (JWT) form with claims such as hardware model, UEID (universal entity ID), debug status and measurement result. Sources (4) Specified by [IETF Entity Attestation Token (RFC 9711)](https://www.rfc-editor.org/rfc/rfc9711), [NVIDIA GPU confidential computing](https://images.nvidia.com/aem-dam/en-zz/Solutions/data-center/HCC-Whitepaper-v1.0.pdf). Sources: [Confidential Compute on NVIDIA Hopper H100 (WP-11459-001 v1.0)](https://images.nvidia.com/aem-dam/en-zz/Solutions/data-center/HCC-Whitepaper-v1.0.pdf), p. 25, step 21; [NVIDIA Attestation, GPU Claims](https://docs.nvidia.com/attestation/advanced-documentation/latest/claims-guide/gpu_claims.html), claims version 3.0. 10. #### 9 The CPU is attested too Confidential VM → Verifier: TDX quote or SEV-SNP report The CVM obtains its own attestation from the CPU. With Intel TDX it is a quote from the TD-quoting enclave, which generates a trust domain’s remote attestation, certified through a chain of PCK (Provisioning Certification Key) certificates to Intel’s root. With AMD SEV-SNP it is a report signed by the chip’s VCEK (Versioned Chip Endorsement Key), whose certificate comes from AMD’s key distribution service (KDS) and chains to AMD’s root key (ARK) through its signing key (ASK). Sources (5) Specified by [Intel TDX](https://cc-enabling.trustedservices.intel.com/intel-tdx-enabling-guide/03/hardware_selection/), [AMD SEV-SNP](https://www.amd.com/en/developer/sev.html). Sources: [AMD, SEV-SNP: Strengthening VM Isolation with Integrity Protection and More (white paper)](https://docs.amd.com/v/u/en-US/SEV-SNP-strengthening-vm-isolation-with-integrity-protection-and-more), p. 18; [Intel Trust Domain Extensions white paper (TDX-Whitepaper-February2022)](https://cdrdv2-public.intel.com/690419/TDX-Whitepaper-February2022.pdf), §E; [API Documentation - Intel SGX and Intel TDX services (Provisioning Certification Service)](https://api.portal.trustedservices.intel.com/content/documentation.html). 11. #### 10 CPU and GPU results are combined Verifier, on its own: CPU and GPU evidence appraised together A verifier can appraise CPU and GPU evidence together; a key the CVM generated can be bound into the CPU report so secrets go only to that VM. Scrutica’s inference No standard found specifies how a deployment binds the GPU's evidence to one particular CVM on the host; the sources examined describe composite appraisal and the binding of keys into report data. Later or contrary evidence Researchers publishing as TEE.fail (Oct 2025) reported that, once keys are extracted from the CPU’s trusted execution environment, NVIDIA attestations can be borrowed from other devices, since they are bound to no particular VM. [TEE.fail: Breaking Trusted Execution Environments via DDR5 Memory Bus Interposition (project site)](https://tee.fail/) · research Sources (4) Specified by [Intel TDX](https://cc-enabling.trustedservices.intel.com/intel-tdx-enabling-guide/03/hardware_selection/), [AMD SEV-SNP](https://www.amd.com/en/developer/sev.html), [IETF Conceptual Message Wrapper (RFC 9999)](https://www.rfc-editor.org/rfc/rfc9999). Source: [Intel Trust Authority documentation, Introduction](https://docs.trustauthority.intel.com/main/articles/articles/ita/introduction.html). 12. #### 11 The relying party decides Relying party → Confidential VM: Allows the device to communicate, or blocks it The Relying Party checks the token's signature against the verifier's published keys, applies its own appraisal policy and, on its decision, “will allow the device to communicate or block it”; CUDA programs (software written for NVIDIA GPUs) run only after attestation succeeds. Scrutica’s inference Scrutica interprets the relying party as a key service that releases data or model keys only after this decision; NVIDIA’s white papers describe no key release. Sources (6) Specified by [IETF RATS architecture (RFC 9334)](https://www.rfc-editor.org/rfc/rfc9334), [IETF Entity Attestation Token (RFC 9711)](https://www.rfc-editor.org/rfc/rfc9711), [NVIDIA GPU confidential computing](https://images.nvidia.com/aem-dam/en-zz/Solutions/data-center/HCC-Whitepaper-v1.0.pdf), [AMD SEV-SNP](https://www.amd.com/en/developer/sev.html). Sources: [Confidential Compute on NVIDIA Hopper H100 (WP-11459-001 v1.0)](https://images.nvidia.com/aem-dam/en-zz/Solutions/data-center/HCC-Whitepaper-v1.0.pdf), p. 26, step 30; [NVIDIA Secure AI with Blackwell and Hopper GPUs (White Paper WP-12554-001\_v1.3)](https://docs.nvidia.com/nvidia-secure-ai-with-blackwell-and-hopper-gpus-whitepaper.pdf), p. 11. 13. #### Not shown by any step: location, workload history, firmware defects Nothing above measures where the hardware physically is, what it computed before or after, or that the firmware is free of bugs; the EAT location claim is self-reported, and proposals to bind location to attestation (V-PEA) are drafts that admit sensor spoofing. Sources (5) Specified by [IETF Entity Attestation Token (RFC 9711)](https://www.rfc-editor.org/rfc/rfc9711), [V-PEA geofence (individual Internet-Draft)](https://www.ietf.org/archive/id/draft-lkspa-rats-verifiable-geo-fence-03.txt), [NIST IR 8320](https://csrc.nist.gov/pubs/ir/8320/final), [OCP S.A.F.E.](https://www.opencompute.org/projects/security/ocp-safe-program/) Source: [IETF Internet-Draft draft-lkspa-rats-verifiable-geo-fence-03](https://www.ietf.org/archive/id/draft-lkspa-rats-verifiable-geo-fence-03.txt), §6.2.1. 1. 15 May 2025 Bill summary Congressional Research Service “Specifically, the bill directs the Department of Commerce to require any covered integrated circuit product to have chip security mechanisms that implement location verification before the product is exported, reexported, or in-country transferred to or in a foreign country.” Full statement Summary of the Chip Security Act (H.R. 3447) as introduced in the House - > Specifically, the bill directs the Department of Commerce to require any covered integrated circuit product to have chip security mechanisms that implement location verification before the product is exported, reexported, or in-country transferred to or in a foreign country. [CRS bill summary, H.R. 3447 Chip Security Act (govinfo BILLSUM)](https://www.govinfo.gov/bulkdata/BILLSUM/119/hr/BILLSUM-119hr3447.xml) · research · 15 May 2025 2. Jul 2025 Lobbying disclosure NVIDIA, AMD, Intel and Google Each lists the Chip Security Act among the issues it lobbied on; the reports name issues and record no position. Full statement Lobbying reports (LD-2), second quarter 2025 to second quarter 2026 Each lists the Chip Security Act among the issues it lobbied on; the reports name issues and record no position. Of the 164 lobbying filings for those quarters that name the bill, 35 are for these four companies, counting amendments and reports by firms they hired: NVIDIA 11, AMD 9, Intel 5 and Google 10. A filing by each company for the second quarters of 2025 and 2026 (8) - [LD-2, NVIDIA Corporation, 2025 Q2 (lda.gov API)](https://lda.gov/api/v1/filings/65879724-313e-4683-a5e5-95c3dbc310e0/) · primary · filed 18 Jul 2025 - [LD-2, NVIDIA Corporation, 2026 Q2 (lda.gov API)](https://lda.gov/api/v1/filings/4caec807-4163-44a7-9fbb-2001daedab24/) · primary · filed 18 Jul 2026 - [LD-2, Advanced Micro Devices, Inc., 2025 Q2 (lda.gov API)](https://lda.gov/api/v1/filings/434071e0-5dae-4e40-8e7f-cb2cf022f185/) · primary · filed 21 Jul 2025 - [LD-2, Advanced Micro Devices, Inc., 2026 Q2 (lda.gov API)](https://lda.gov/api/v1/filings/a4917840-6c3b-4bdc-abf8-63a88f704f96/) · primary · filed 20 Jul 2026 - [LD-2, Intel Corporation, 2025 Q2 (lda.gov API)](https://lda.gov/api/v1/filings/787b8ebc-8a6d-47e0-96d7-cf25aab6d180/) · primary · filed 18 Jul 2025 - [LD-2, Intel Corporation, 2026 Q2 (lda.gov API)](https://lda.gov/api/v1/filings/d3c5f124-cbb7-4a36-879f-44c5b0d64634/) · primary · filed 17 Jul 2026 - [LD-2, Google Client Services LLC, 2025 Q2 (lda.gov API)](https://lda.gov/api/v1/filings/6e9426dd-f07f-4504-9c0c-369f90ab9f52/) · primary · filed 21 Jul 2025 - [LD-2, Google Client Services LLC, 2026 Q2 (lda.gov API)](https://lda.gov/api/v1/filings/0676fd59-1d30-4323-aad2-83e03d2ff407/) · primary · filed 20 Jul 2026 3. 31 Jul 2025 Government notice Cyberspace Administration of China Translation by Scrutica, in EnglishOn 31 July 2025 the Cyberspace Administration of China summoned NVIDIA and required it to explain the vulnerability and backdoor security risks of the H20 computing chips it sells to China, and to submit supporting evidence. Full statement Notice of the summons on its website - Original, in Chinese > 美人工智能领域专家透露，英伟达算力芯片“追踪定位”“远程关闭”技术已成熟。 Translation by Scrutica, in EnglishExperts in the US artificial-intelligence field have disclosed that the “tracking and positioning” and “remote shutdown” technology for NVIDIA’s computing chips is mature. - Original, in Chinese > 国家互联网信息办公室于2025年7月31日约谈了英伟达公司，要求英伟达公司就对华销售的H20算力芯片漏洞后门安全风险问题进行说明并提交相关证明材料。 Translation by Scrutica, in EnglishOn 31 July 2025 the Cyberspace Administration of China summoned NVIDIA and required it to explain the vulnerability and backdoor security risks of the H20 computing chips it sells to China, and to submit supporting evidence. [国家互联网信息办公室就H20算力芯片漏洞后门安全风险约谈英伟达公司](https://www.cac.gov.cn/2025-07/31/c_1755675743897163.htm) · primary · 31 Jul 2025 4. Five NVIDIA statements, July to December 2025. Each says its chips have none of what is marked; the blog post of 5 August adds that they should not have them. - 31 Jul 2025 Statement reported by AP and The Register, marked: backdoors. “NVIDIA does not have ‘backdoors’ in our chips that would give anyone a remote way to access or control them,” an Nvidia spokesperson said in a statement to AP. Full statement Spokesperson statement after the summons - > Cybersecurity is critically important to us. NVIDIA does not have ‘backdoors’ in our chips that would give anyone a remote way to access or control them, an Nvidia spokesperson said in a statement to AP. [China summons Nvidia over ‘backdoor safety risks’ in H20 chips (AP, syndicated by WSLS)](https://www.wsls.com/news/politics/2025/07/31/china-summons-nvidia-over-backdoor-safety-risks-in-h20-chips/) · press or analyst · 31 Jul 2025 · A statement to the press, with no copy found on an NVIDIA site · also [Beijing asks Nvidia about backdoors in China-bound AI chips (The Register)](https://www.theregister.com/on-prem/2025/07/31/beijing-asks-nvidia-about-backdoors-in-china-bound-ai-chips/1289382), 31 Jul 2025 - 5 Aug 2025 Blog post, marked: backdoors; kill switches or remote disable. “NVIDIA GPUs do not and should not have kill switches and backdoors.” Full statement NVIDIA blog, by its chief security officer - > NVIDIA GPUs do not and should not have kill switches and backdoors. - > Some point to smartphone features like ‘find my phone’ or ‘remote wipe’ as models for a GPU kill switch. That comparison doesn’t hold water — optional software features, controlled by the user, are not hardware backdoors. [No Backdoors. No Kill Switches. No Spyware](https://blogs.nvidia.com/blog/no-backdoors-no-kill-switches-no-spyware/) · primary · 5 Aug 2025 - 10–11 Aug 2025 Statement reported by The Paper, marked: backdoors. Translation by Scrutica, in EnglishNVIDIA chips have no ‘backdoors’ and will not let anyone access or control the chips remotely. Full statement Reply after an article of 10 August by the account Yuyuan Tantian, relayed in Chinese - Original, in Chinese > 英伟达回应称：“网络安全对我们至关重要。英伟达的芯片不存在‘后门’，不会让任何人通过远程方式访问或控制芯片。” Translation by Scrutica, in EnglishNVIDIA responded: “Cybersecurity is critically important to us. NVIDIA chips have no ‘backdoors’ and will not let anyone access or control the chips remotely.” [英伟达再回应H20芯片“后门”风险](https://www.thepaper.cn/newsDetail_forward_31370247) · press or analyst · 11 Aug 2025 · Relayed by the press; the English original was not located on an NVIDIA host. - 10 Dec 2025 Blog post, marked: backdoors; kill switches or remote disable; hardware tracking. “NVIDIA GPUs do not have hardware tracking technology, kill switches and backdoors.” Full statement NVIDIA blog on the fleet management service - > This optional service provides real-time monitoring by each GPU system communicating and sharing GPU metrics with the external cloud service. NVIDIA GPUs do not have hardware tracking technology, kill switches and backdoors. [Opt-In NVIDIA Software Enables Data Center Fleet Management](https://blogs.nvidia.com/blog/optional-data-center-fleet-management-software/) · primary · 10 Dec 2025 - 10 Dec 2025 Statement reported by Reuters, marked: kill switches or remote disable. “There is no feature within Nvidia GPUs that allow Nvidia or a remote actor to disable the Nvidia GPU,” Nvidia said. Full statement - > There is no feature within Nvidia GPUs that allow Nvidia or a remote actor to disable the Nvidia GPU, Nvidia said. [Nvidia builds location verification tech that could help fight chip smuggling (Reuters, syndicated by Yahoo Finance)](https://finance.yahoo.com/news/exclusive-nvidia-builds-location-verification-044501961.html) · press or analyst · 10 Dec 2025 · Relayed by the press. - Reuters’ report of what an unnamed NVIDIA official said > The software was built to allow customers to track a chip’s overall computing performance - a common practice among companies that buy fleets of processors for large data centers - and would use the time delay in communicating with servers run by Nvidia to give a sense of the chip’s location on par with what other internet-based services can provide, according to an Nvidia official. [Nvidia builds location verification tech that could help fight chip smuggling (Reuters, syndicated by Yahoo Finance)](https://finance.yahoo.com/news/exclusive-nvidia-builds-location-verification-044501961.html) · press or analyst · 10 Dec 2025 · Relayed by the press from an unnamed NVIDIA official; none of the NVIDIA documents read describes it. 5. 25 Feb 2026 Annual report (Form 10-K), risk factors NVIDIA “\[the USG\] may impose additional conditions such as requiring chip tracking and throttling mechanisms that could disable or impair GPUs if certain events, including unauthorized system configuration, use, or location, are detected. Such government mandates in chip designs could introduce system vulnerabilities and expose us to significant risk and potential liability \[…\]” Full statement - > \[the USG\] may impose additional conditions such as requiring chip tracking and throttling mechanisms that could disable or impair GPUs if certain events, including unauthorized system configuration, use, or location, are detected. Such government mandates in chip designs could introduce system vulnerabilities and expose us to significant risk and potential liability, negatively impact demand for our products, and could have a material impact on our business, operating results, and financial condition. - > China’s government publicly questioned whether our H20 products have built-in vulnerabilities, discouraging customers from purchasing our products. We provided a public response explaining that our GPUs, including H20, do not include such built-in vulnerabilities, and will respond to any follow-up questions we receive. [NVIDIA Form 10-K for fiscal year ended 2026-01-25](https://www.sec.gov/Archives/edgar/data/1045810/000104581026000021/nvda-20260125.htm) · primary · 25 Feb 2026 (filed) 6. 27 Feb 2026 Blog post Information Technology Industry Council “While there are valid commercial reasons for companies to pursue location verification, a government chip tracking mandate would create the impression of deepening U.S. government control over the American AI stack \[…\]” Full statement ITI blog, on the Chip Security Act - > While there are valid commercial reasons for companies to pursue location verification, a government chip tracking mandate would create the impression of deepening U.S. government control over the American AI stack, fueling questions about the security, reliability, and privacy of U.S. technology and pushing the very countries that should be core customers of U.S. providers toward alternatives. - > Instead of requiring U.S. companies to track AI chips, policymakers should address export control avoidance by increasing funding and staffing for the Department of Commerce’s Bureau of Industry and Security (BIS), which administers and enforces U.S. export controls. [The Unintended Consequence of the Chip Security Act (ITI TechWonk blog, Mike Flynn)](https://www.itic.org/news-events/techwonk-blog/the-unintended-consequence-of-the-chip-security-act) · primary · 27 Feb 2026 7. 2 Mar 2026 Statement Semiconductor Industry Association “While we understand policymakers’ interest in addressing this issue, we cannot support blanket mandates for new, untested, and potentially infeasible on-chip mechanisms, such as what is being proposed in the Chip Security Act.” Full statement Statement by John Neuffer, its president and chief executive, on the Chip Security Act - > SIA members are fully committed to complying with export controls, and we strongly oppose the illicit diversion and misuse of our chip technologies. While we understand policymakers’ interest in addressing this issue, we cannot support blanket mandates for new, untested, and potentially infeasible on-chip mechanisms, such as what is being proposed in the Chip Security Act. Rushing to legislate complex, costly, and unproven security features risks undermining global trust in American semiconductor technologies and efforts to promote the export of the American AI technology stack, eroding our global leadership and competitiveness. - > Our companies work closely with government agencies, law enforcement, and other relevant stakeholders to prevent and detect illegal diversion and misuse of our products. [SIA Statement on Chip Security Act (Semiconductor Industry Association)](https://www.semiconductors.org/sia-statement-on-chip-security-act/) · primary · 2 Mar 2026 8. 26 Mar 2026 Press release Rep. Bill Huizenga “The Chip Security Act enhances protections on AI chip exports by scaling up best security practices and encouraging innovative technologies,” said South and Central Asia Subcommittee Chairman Bill Huizenga. Full statement House Foreign Affairs Committee release on advancing the Chip Security Act, which he introduced - > The Chip Security Act enhances protections on AI chip exports by scaling up best security practices and encouraging innovative technologies, said South and Central Asia Subcommittee Chairman Bill Huizenga. - > This bill directly supports the Trump administration’s AI Action Plan recommendations and benefits from both the administration’s technical assistance and industry feedback. [Chairman Mast, HFAC, Advances Chip Security Act (House Foreign Affairs Committee)](https://foreignaffairs.house.gov/news/press-releases/chairman-mast-hfac-advances-chip-security-act) · primary · 26 Mar 2026 Statements on other topics (6) None of these addresses kill switches, backdoors, tracking or location verification. - [AMD describes its export-compliance screening, in written answers to the Senate Commerce Committee](https://www.govinfo.gov/content/pkg/CHRG-119shrg61426/html/CHRG-119shrg61426.htm) - [Huawei guarantees that its commitment to cyber security “will never be outweighed by the consideration of commercial interests”](https://www.huawei.com/en/declarations/cyber-security) - [Google recommends research on chip security and more resources for BIS](https://files.nitrd.gov/90-fr-9088/Google-RFI-2025.pdf) - [Cerebras discloses security conditions on its export licences for the UAE](https://www.sec.gov/Archives/edgar/data/2021728/000162828026056357/cbrs-20260630.htm) - [Jim Keller, Tenstorrent’s chief executive, says Blackhole will need to be de-featured to ship to China, as EE Times reports](https://www.eetimes.com/jim-keller-whatever-nvidia-does-well-do-the-opposite/) - [Microsoft asks that the AI diffusion rule’s quantitative caps go and its qualitative security standards stay](https://blogs.microsoft.com/on-the-issues/2025/02/27/trump-administration-ai-global-race/) The instruments that ask for verification, among them [the Chip Security Act](https://scrutica.com/verification/policy#pol-us-chip-security-act-hr3447), are quoted and dated in [the policy register](https://scrutica.com/verification/policy).