Custom Quantization Boosts R9700 by 5.8%
alphatrad · reddit · 2026-07-18
The author tuned a custom quantization format Q80ROCMFPX for three Radeon AI PRO R9700 cards, aiming to outperform generic Q8 on gfx1201.
Results
- Model size is 2.94% smaller than upstream Q80
- HumanEval: 140/164, tying with the original
- Full model decoding: 5.77%–5.87% faster
- Equal workload test: Median 5.82% faster, winning 17/18 paired groups, p=0.000965
- End-to-end agent total runtime: Only 2.21% faster, missing the author's 3% threshold
Constraints
- Currently cannot run directly on upstream llama.cpp, Ollama, LM Studio, or vLLM
- Requires the author's specific ROCmFPX fork and source compilation
- This optimization is primarily effective for gfx1201 / R9700, offering limited benefits for NVIDIA or other AMD cards
The author provided a full experimental repository and noted it would be helpful if more people reproduced the results and compared them against stock Q8.
More from Infra
- Engineer with thousands of fabricated devices pours cold water on photonic crystal modulator yields — jwt0625 · 2026-10-26
- Anthropic S-1 reveals 48 cents of compute cost per dollar of Q2 revenue — nmasc_ · 2026-10-26
- Persistent AI joins UR network to route private AI inference with encrypted, IP-hidden traffic — 0xSammy · 2026-10-26
- Anthropic's disclosed compute commitments: $114.5B Google, $161.2B Broadcom top the list — BenBajarin · 2026-10-26
- SpaceX strikes $40B deal with Apollo Global to lease AI chips and data center hardware — Polymarket · 2026-10-26
- SpaceX AI revenue up 830%, Claude designs 318K antivenom proteins in 64 hours — PeterDiamandis · 2026-10-26