Unlocking 80 GB Memory on NVIDIA CMP 170HX: 31× Performance Gain and 20× Price Surge
The article details how researchers bypassed OTP and firmware locks on the NVIDIA CMP 170HX mining GPU, expanding its memory from 10 GB to up to 80 GB, boosting FP32 performance from about 0.39 TFLOPS to roughly 12.2 TFLOPS (≈31×), while discussing benchmark results, stability limits, and the resulting price explosion.
The NVIDIA CMP 170HX, introduced five years ago as a dedicated mining card, originally shipped with 8 GB or 10 GB of memory, 70 SM groups (4 480 CUDA cores), and a PCIe Gen1 ×4 interface; its launch price was US$4 299 (≈CNY 30 000).
Hardware restrictions are enforced by a one‑time‑programmable (OTP) fuse that records the model code and by the Falcon security co‑processor firmware, which reads the OTP at boot and limits compute, memory size, and PCIe bandwidth.
In June, Arizona State University researcher Jon Pry published a paper showing that a DMA out‑of‑bounds and stack‑protection vulnerability in the Falcon firmware can be exploited to gain privileged register access, modify the lock registers, and effectively bypass the OTP seal.
Using this method, the 10 GB version can be “opened” to expose up to 80 GB of addressable memory (the 8 GB version can reach 64 GB). Benchmarks show FP32 SGEMM rising from 393 GFLOPS to 12.23–12.26 TFLOPS (≈31×), FP64 from 197 GFLOPS to 11.67 TFLOPS (≈59×), and TF32 Tensor performance from 6.24 TFLOPS to about 94.1 TFLOPS.
Stability tests reveal that the full 80 GB mapping produces 2 796 memory errors under default refresh; limiting to 40 GB eliminates errors, while increasing the refresh rate restores 80 GB at the cost of roughly 32 % performance loss. The 8 GB SK Hynix version, unlocked to 64 GB, is currently favored by the community.
Compared with a full‑spec A100, the unlocked CMP 170HX still lags in CUDA core count, PCIe bandwidth (only Gen2 ×4, with Gen3 still locked), and lacks ECC, HBM, and other features; restoring PCIe ×16 would require soldering back 24 missing AC‑coupling capacitors on the PCB.
Because the cards are retired mining units, long‑term stability is uncertain, and they often need rack‑mount cooling adaptations for desktop use. Recent user reports mention crashes after prolonged CUDA load on the 64 GB unlocked cards.
The market reaction has been dramatic: the card’s price jumped from under US$200 to at least €1 000 overseas, and from CNY 300‑500 to an average of CNY 6 000 in China. An open‑source project, cmpunlocker, hosted at
, provides the necessary steps for the community to perform the unlock.
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
IT Services Circle
Delivering cutting-edge internet insights and practical learning resources. We're a passionate and principled IT media platform.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
