DeepSeek on RDNA2/3 + DDR4: 19 t/s peaks, 9 t/s average decode on 56GB VRAM
EmPips · reddit · 2026-10-23
A Reddit user ran a 350GB iq2-quantized DeepSeek model (Q6 ngrams, Dspark on) on consumer AMD hardware: 3950x, 96GB DDR4-3000, RX 7900xtx 24GB + W6800 32GB (56GB combined), PCIe Gen3 NVMe.
- Highly variable performance: peaks of 19 t/s, usually 13 t/s, dropping to 3 t/s during I/O; averaged 9 t/s decode after a long prompt, with 150 t/s prefill.
- Verdict: the model is "crazy smart" — clearly better than their iq4-xs quant of Qwen3.8-Flash-Next — but it thinks a lot, and at 9 t/s average it's only practical for chatting.
More from Infra
- TokenGate e2e-learns to drop tokens on the fly, opening a new frontier for LLM prefill — CShorten30 · 2026-10-23
- $126B could carpet the entire 494-mile California high-speed rail route in $100 bills — john__allard · 2026-10-23
- TTR-Compressor: test-time regression with closed-form solutions beats pooling for sequence compression — SphereLab · 2026-10-23
- Dragonfly files CFTC comment letter urging standardized compute derivatives markets — AccBalanced · 2026-10-23
- Liquid Inference launches token Reservations: lock in fixed prices for 1-12 months with competing providers — AccBalanced · 2026-10-23
- MiniMax H3 video generation launches on Japan's ai& platform from $0.0125/second — DavidBennett__ · 2026-10-23