Dual MI50 benchmarks show 32GB VRAM for under $300
tabletuser_blogspot · reddit · 2026-10-04
- Setup: two MI50 GPUs (16GB HBM2 each, 1TB/s bandwidth, overclockable to 1200MHz), power capped at 145W, one card's miniDP activated via firmware + xrandr custom mode. Benchmarks run with Ubuntu Vulkan build of llama.cpp (b11325).
- Key results: dense 27-31B models hit 120-167 t/s prompt processing / 12-18 t/s generation; MoE models are far faster — Qwen3.6 35B-A3B Q4KM reached 983 t/s pp512 and 47 t/s tg128, Nemotron 30B-A3B hit 864 t/s pp512.
- Cost angle: 16GB MI50s average under $150, so 32GB total VRAM at this performance level costs under $300; author used RADVPERFTEST=nogttspill with -fa on -ngl 99 and notes cooling is still unsolved. A third MI50 will be added to cover the VRAM gap.
More from Infra
- Investors want to fund "insurance for AI" startups wrapping data center leases — katieruthmishra · 2026-10-04
- Data tower waste heat could warm 100,000 homes: 134 MW at 85°C, ~1 TWh/year — IgorCarron · 2026-10-04
- Ex-OpenAI researcher flags: SPV shills from six months ago are now selling you data centers — suchenzang · 2026-10-04
- MTPLX Ships New Version: Cache Copies Cut to One, Much Faster Decoding Past 140k Tokens — HankYeomans · 2026-10-04
- Open-source marin project pushes single-rack training MFU from 28.2% to 31.1% — dlwh · 2026-10-04
- Samsung: HBM to consume 30% of DRAM wafer capacity by 2027, up from ~20% — Beth_Kindig · 2026-10-04