Confidential GPU

A hardware-enforced boundary within the GPU that cryptographically isolates a workload's data and code from all other processes, including the host OS and hypervisor. The TEE encrypts data in use—the third state of data protection alongside data at rest and data in transit. On NVIDIA H100 GPUs, this is implemented via a confidential computing mode that creates isolated virtual machine instances with dedicated memory regions inaccessible to the host.

• Memory pages within the TEE are encrypted with AES-GCM at the hardware level
• Attestation reports cryptographically verify the GPU's firmware and security state
• Even infrastructure administrators with physical access cannot inspect enclave contents

A cryptographic verification mechanism that allows a remote party to confirm that a workload is running on genuine, untampered confidential GPU hardware with the correct firmware and security configuration. The GPU generates a signed attestation report containing measurements of the TEE's initial state, which can be validated against a trusted reference.

• Uses a hardware root of trust embedded in the GPU silicon
• Reports include firmware measurements, security version numbers, and platform identity
• Enables zero-trust deployment where workloads only execute on verified hardware
• Critical for regulated industries requiring proof of execution environment integrity

A dedicated inline hardware encryption engine on the PCIe bus that transparently encrypts and decrypts all data moving between the CPU and the confidential GPU. This prevents bus snooping attacks where an attacker with physical access to the PCIe interconnect could capture model weights or inference data in transit.

• Operates at line rate with negligible latency overhead
• Secures both command buffers and data payloads
• Prevents DMA attacks from compromised PCIe devices
• Complements TEE memory encryption for end-to-end data-in-use protection

Confidential GPUs enable secure workload co-location where multiple tenants or departments can share a single physical GPU without risk of data leakage. Each tenant's execution environment is cryptographically isolated, making this ideal for cloud providers offering GPU-as-a-Service and enterprises consolidating sensitive workloads.

• Hardware-enforced isolation prevents cross-tenant memory access
• Each tenant receives a unique encryption key managed by the GPU's key management unit
• Enables confidential multi-party computation where competing organizations can run models on shared infrastructure
• Supports MIG partitioning with per-instance confidentiality guarantees

Confidential GPUs protect proprietary model weights during inference by ensuring the model is only decrypted inside the TEE. This prevents infrastructure providers, cloud operators, or malicious insiders from extracting valuable intellectual property. The model owner can set release policies that require successful attestation before decryption keys are released.

• Model weights remain encrypted until verified inside the TEE
• Prevents model theft even with physical access to the GPU node
• Enables model marketplace scenarios where inference can be monetized without exposing IP
• Supports encrypted container images with runtime decryption inside the enclave

Confidential GPUs provide the technical controls required to meet data sovereignty and privacy regulations such as GDPR, HIPAA, and the EU AI Act. By cryptographically proving that data is processed only within approved geographies and on verified hardware, organizations can demonstrate technical enforcement of compliance policies rather than relying solely on contractual agreements.

• Attestation logs provide auditable proof of processing environment
• Enables sovereign AI deployments where data never leaves jurisdictional boundaries in plaintext
• Supports data processing agreements with cryptographic enforcement
• Reduces compliance scope by removing infrastructure providers from the trust boundary

A secure area within a main processor that guarantees code and data loaded inside is protected with respect to confidentiality and integrity. In confidential GPUs, the TEE creates a hardware-enforced boundary that isolates sensitive AI workloads from the host operating system, hypervisor, and even the infrastructure provider. Data in use—including model weights, training data, and inference inputs—remains encrypted within the TEE and is only decrypted inside the processor package.

A hardware component integrated into the GPU's memory controller that transparently encrypts all data written to off-chip memory and decrypts data read back. This prevents physical attacks such as cold boot attacks, bus snooping, and DRAM interposers from extracting plaintext data. Modern confidential GPUs use AES-XTS encryption with per-session keys, ensuring that even if an attacker physically accesses the memory modules, the data remains cryptographically protected.

A cryptographic mechanism that allows a remote party to verify the identity and integrity of the confidential GPU environment before releasing secrets or data. The GPU generates a signed attestation report containing measurements of the firmware, TEE configuration, and platform state. This report is verified against a trusted authority, ensuring the GPU is running genuine, untampered code. Critical for establishing trust in multi-tenant cloud and sovereign deployments.

A hardware mechanism that maintains a protected mapping table between guest physical addresses and host physical addresses within the GPU's memory management unit. This prevents the hypervisor or host OS from remapping pages to snoop on confidential workload data. The mapping table itself is encrypted and integrity-protected, ensuring that even a compromised hypervisor cannot redirect memory accesses to extract sensitive AI model parameters or training data.

Techniques implemented in confidential GPU hardware and firmware to prevent indirect information leakage through observable physical phenomena:

• Cache timing attacks: Partitioned cache lines prevent cross-workload interference
• Power analysis: Constant-time cryptographic operations mask key material
• Page fault attacks: Secure nested paging prevents the hypervisor from inducing observable faults These mitigations are essential for defending against sophisticated attackers with physical access to sovereign infrastructure.