Nvidia Crypto Cards Now Used for Inference
Nvidia Crypto Cards Now Used for Inference
Nvidia's crypto cards have been repurposed for inference after a vulnerability in their drivers was exploited, significantly enhancing their performance.
Nvidia Crypto Cards Now Used for Inference
Back in 2021, Nvidia released a series of heavily cut-down CMP HX170 cards based on the A100 chip for mining. Nominally, they had either 8 or 10 gigabytes of memory with 32 times reduced computational power, but with full memory bandwidth needed for mining. After Ethereum transitioned to a proof of stake algorithm, these cards became nearly useless.
However, in late June, a paper emerged detailing the removal of some restrictions through a vulnerability in the new Nvidia drivers. The authors did not publish the code, but within a few weeks, the vulnerability was independently replicated, leading to the creation of cmpunlocker.
The 10-gigabyte cards became 40 (or even 80) gigabytes, while the 8-gigabyte cards turned into 64-gigabyte cards. The issue was that removing the HBM chips without reducing bandwidth was impossible, so Nvidia simply blocked 80% of the memory through software. Hackers also managed to remove the computational limits. The main remaining problem is that PCIe is limited to version 2.0 and physically only x4, but enthusiasts are already trying to fix this.
As a result, lucky buyers who managed to purchase the cards before the price surge for $100-200 got a nearly full-fledged A100 for a very low price, especially considering the current crisis. Now prices have skyrocketed, and buying them at 5-10 times the original price is hardly worth it.
Are there any owners of such cards among the channel's subscribers? What are your impressions?
Why it matters
AnalysisThis news is significant as it illustrates how vulnerabilities in software can dramatically alter the graphics card market. It also highlights the importance of security in technologies related to cryptocurrencies.
Discuss in community
Share your questions and insights with developers
Related in Ecosystem
Practical guides, core terminology, and agent skills
Nvidia's Monopoly and the CUDA Platform
An analysis of Nvidia's technological and economic dominance in the AI market. The CUDA (Compute Unified Device Architecture) platform, created in 2006, transformed ordinary gaming GPUs into the planet's primary computational tool, making it difficult for competitors like AMD and Intel to break this monopoly.
TensorRT-LLM & SGLang High-Speed Engines
Deeply compiled computational engines for extreme optimization of language model inference on NVIDIA servers, utilizing graph optimization, FlashAttention-3, and advanced routing.
Anthropic releases FlashProve in open source, a tool for provable inference
Anthropic published a prototype of an open cryptographic protocol called FlashProve that lets anyone verify a model response without access to model weights.
Anthropic introduces FlashProve prototype for provable AI inference
Anthropic has presented FlashProve, a prototype system that lets developers mathematically verify that an answer came from a specific AI model without exposing the model weights.