Exploration of Cryptocurrency Mining-Specific GPUs in AI Applications: A Case Study of CMP 170HX

TL;DR

This paper demonstrates that modifying NVIDIA CMP 170HX GPUs by disabling specific instructions can significantly enhance their AI computational performance, offering a cost-effective and environmentally friendly hardware reuse strategy.

Contribution

It introduces a novel method of reusing mining GPUs for AI tasks by disabling certain instructions, validated through extensive benchmarking and hardware analysis.

Findings

FP32 performance exceeds 15 times original capability

Inference performance in large language models triples

Recycling GPUs reduces electronic waste and costs

Abstract

This study systematically tests a computational power reuse scheme proposed by the open source community disabling specific instruction sets (Fused Multiply Add instructions) through CUDA source code modifications on the NVIDIA CMP 170HX platform. Experimental results validate the effectiveness of this approach, partially restoring the GPU's computational capabilities in artificial intelligence (AI) tasks. Performance evaluations using open-source GPU benchmarks (OpenCL benchmark, mixbench) and AI benchmarks (LLAMA-benchmark) reveal that its FP32 floating-point performance exceeds 15 times the original capability, while inference performance for certain precision levels in large language models surpasses threefold improvements. Furthermore, based on hardware architecture analysis, this paper proposes theoretical conjectures for further improving computational utilization through alternative adaptation pathways.Combining energy efficiency ratios and cost models, the recycling value of such obsolete GPUs in edge computing and lightweight AI inference scenarios is evaluated. The findings demonstrate that rationally reusing residual computational power from mining GPUs can significantly mitigate the environmental burden of electronic waste while offering cost-effective hardware solutions for low-budget computing scenarios.