CMP 170HX ECC Is Now Unlocked. What Actually Changes?

ECC can correct certain transient errors in memory data during a long computation, helping keep those errors from affecting the result. v0.5 enables DRAM and SRAM ECC and increases active SMs from 70 to 74. SM count relates to parallel compute resources, but this update alone does not provide a measured speed change for any particular model.

Earlier card notes that said ECC was unavailable describe the state before v0.5. The patch record and installation notes below show the release's scope and published test details.

The repository contains ecc-enable.patch, ecc-fbpa-static.patch, and ecc-reporting.patch. The first two are used to enable ECC behavior in the memory-control path; the reporting patch adds a compatibility path for how the driver handles ECC status and error counters. Even though all three patches were introduced together, the behavior change and display compatibility should be read separately.

In the maintainer's published test with driver 615.71.09, ECC_CTRL was 0x41, memory register MR4 was 0x33, the readout was 0x00010233, and nvidia-smi showed ECC Enabled. The configuration also retained 64GB of memory. This record demonstrates the v0.5 path, but does not guarantee the same result for other driver, card, or kernel combinations.

The ECC reporting compatibility patch synthesizes statuses such as Enabled, Supported, scrubComplete, and InfoROM 6.16, and initializes some counters to zero. Therefore, an Enabled display and zero counters in nvidia-smi alone cannot show whether an error was corrected when one occurred or whether memory remains stable after a long load. That is why you should record the patch version and review memory-load tests alongside real workload results.

cmpunlocker v0.5 requires the NVIDIA open driver 610.xx.xx or later on Linux x86-64. It also needs matching driver libraries and firmware, kernel headers for the running kernel, and Secure Boot disabled. A network connection is needed for the first build. The repository's current support list includes 615.78.08, so choose a driver branch listed as supported rather than assuming the newest release is suitable.

Installation uses the repository's sudo ./install.sh command followed by a reboot. The default profile detects the card configuration. If you need to select a profile explicitly, first identify the hardware, then use sudo ./install.sh --profile=8gb or sudo ./install.sh --profile=10gb. The 8GB profile can expose 64GB on dual-rank cards and 32GB on some single-rank cards; the 10GB profile targets a 40GB configuration. A profile changes the memory-unlock configuration as well as the ECC setup, so it must match the card.

The installer patches Linux kernel modules and may use unsigned modules. A kernel update can require rebuilding the matching module, and these changes do not permanently flash the card's firmware. To roll back, follow the repository instructions with sudo ./remove.sh --yes, then reboot. Run the procedure on the Linux PC that uses the GPU, and prepare a way to recover its operating system and driver before installing.

After rebooting, check whether nvidia-smi reports ECC Enabled and record the cmpunlocker and driver versions. Also confirm that the card is detected and its memory capacity matches the selected profile.

Next, run a memory-intensive fixed-input workload or a numerical calculation with a known answer, and record its result alongside CUDA² and kernel error logs. If checking model answers, use temperature 0 and the same prompt³ and settings. Also watch for errors or interruptions during extended jobs.

If errors persist, review the recorded driver and kernel combination and its logs. If removal is needed, follow the repository's removal procedure. Recheck with the model, context, and concurrent work you plan to use in production to assess whether the patch fits your workload.