Full Report
NVIDIA is warning users to activate System Level Error-Correcting Code mitigation to protect against Rowhammer attacks on graphical processors with GDDR6 memory. [...]
Analysis Summary
# Vulnerability: Rowhammer Attacks Against GDDR6 GPUs
## CVE Details
- CVE ID: Not explicitly listed in the provided text. This appears to be a class of hardware vulnerability (Rowhammer) being addressed by vendor guidance rather than a specific, published CVE at the time of the article.
- CVSS Score: Not specified.
- CWE: Potentially CWE-798 (Use of Hard-coded Credentials) or a hardware-specific weakness relating to memory cell interference, if a specific CVE were assigned (e.g., related to unauthorized access/manipulation via physical side-channel).
## Affected Systems
- Products: NVIDIA GPUs utilizing GDDR6 memory, including:
- GeForce RTX 30 and 40 Series
- Quadro RTX 6000, RTX 8000, and Volta GV100
- Jetson AGX Orin Industrial
- IGX Orin
- Versions: Any configuration of the listed GPUs where appropriate memory protection mechanisms might be absent or misconfigured (specifically when System Level ECC is not active).
- Configurations: Systems where System Level ECC is not enabled or where the memory controller is susceptible to bit flips induced by aggressive access patterns.
## Vulnerability Description
The vulnerability relates to the **Rowhammer** hardware side-channel attack, which can induce bit-flips in adjacent memory rows in DRAM chips (specifically GDDR6 in this context) by repeatedly accessing a target row. In a multi-tenant environment like cloud servers, successful exploitation could lead to data corruption or unauthorized privilege escalation by manipulating sensitive data stored in GPU memory. Newer NVIDIA architectures (Blackwell and Hopper series) mitigate this vulnerability through integrated on-die Error Correcting Code (ECC) protection.
## Exploitation
- Status: Academic/Theoretical concern, difficulty highlighted.
- Complexity: High (Requires specific conditions, high access rates, and precise control to exploit reliably).
- Attack Vector: Likely Local or Adjacent (requires proximity or access to the host system/GPU resources).
## Impact
- Confidentiality: Potential (Manipulation of sensitive data in memory).
- Integrity: High (Data corruption or unauthorized state changes possible).
- Availability: Low to Medium (Potential for denial of service through memory corruption).
## Remediation
### Patches
Actual software patches are not detailed, as the issue is mitigated by hardware features or configuration changes:
- **Future Architectures:** Blackwell RTX 50 Series, GB200, B200, B100, and Hopper H100, H200, H20, GH200 feature built-in on-die ECC protection, mitigating this specific class of Rowhammer vulnerability.
### Workarounds
1. **Enable System Level ECC:** Configure the system/hardware such that ECC memory protection is active for the affected GDDR6 memory.
2. **Driver/Firmware Updates:** Apply vendor-specific guidance to ensure memory controllers are configured defensively (implied by the guidance release).
## Detection
- **Out-of-Band:** Use the system's **BMC (Baseboard Management Controller)** and hardware interface software (e.g., **Redfish API**) to check the `"ECCModeEnabled"` status.
- **In-Band:** Use the **`nvidia-smi` command-line utility** from the system CPU to check and enable ECC where supported. (Note: Access to tools like NSM Type 3 and NVIDIA SMBPBI may require the NVIDIA Partner Portal).
## References
- Vendor Advisory: NVIDIA guidance on defending GDDR6 GPUs against Rowhammer attacks.
- Relevant Links:
- bleepingcomputer com/news/security/nvidia-issues-guidance-to-defend-gddr6-gpus-against-rowhammer/
- dmtf org/standards/redfish (For BMC interface reference)