will it local
GB203, 10752 CUDA, 256-bit GDDR7 at 30 Gbps = 960 GB/s (TechPowerUp, live-pinned 0905). Blackwell halo card with mid-bandwidth memory config; decode is memory-bound so it lands close to the cheaper 5070 Ti. Pick order for pure LLM: 5070 Ti < 5080 < 5090.
computed band 93.0-150.0 tok/s for 8B Q4_K_M (roofline, 960.0 GB/s VRAM); community: 140 tok/s Meta Llama 3.1 8B Instruct Q4_K_M (https://www.kunalganglani.com/llm-benchmarks); community: 26.1 tok/s Meta Llama 3.1 8B Instruct Q4_K_M (https://www.localscore.ai/accelerator/1119)
Effective decode window: 0.50–0.80 of 960 GB/s nominal → ~480–768 GB/s effective (llama.cpp decode, Q4_K_M basis; card-specific window from cited community benches)
Community check — localmaxxing.com medians (1 bucket): localmaxxing.com
→ ~130 tok/s · 8-9B Q4_0/Q4_K_S · 6 runs · inside our computed band
→ 13b-q4 tight 14.6 of 16 GB usable — keep context modest
→ 3b-q4 full 2.8 GB needed of 16 GB usable — headroom for context
→ 4b-q4 full 3.7 GB needed of 16 GB usable — headroom for context
→ 8b-q4 full 6.1 GB needed of 16 GB usable — headroom for context
→ 140 tok/s · Meta Llama 3.1 8B Instruct Q4_K_M Kunal Ganglani bench DB (reference, llama.cpp) · source
→ 26.1 tok/s · Meta Llama 3.1 8B Instruct Q4_K_M LocalScore median (early sample, skewed low) · source
Nominal bandwidth: 960.0 GB/s — see the effective decode window above
→ qwen3:0.6b on Tesla M60 (16GB, used) ($35) 0.6B · 0.6GB · needs ~1GB (4k ctx)
→ qwen3:1.7b on Tesla M60 (16GB, used) ($35) 1.7B · 1.4GB · needs ~1.8GB (4k ctx)
→ qwen3:4b on Tesla M60 (16GB, used) ($35) 4B · 2.6GB · needs ~3.2GB (4k ctx)
→ llama3.2:1b on Tesla M60 (16GB, used) ($35) 1B · 1.3GB · needs ~1.4GB (4k ctx)
→ llama3.2:3b on Tesla M60 (16GB, used) ($35) 3B · 2.0GB · needs ~2.4GB (4k ctx)
→ gemma3:1b on Tesla M60 (16GB, used) ($35) 1B · 1.0GB · needs ~1.1GB (4k ctx)
→ phi4:mini on Tesla M60 (16GB, used) ($35) mini 3.8B · 2.5GB · needs ~3GB (4k ctx)
→ moondream:2b on Tesla M60 (16GB, used) ($35) 2B · 1.7GB · needs ~3.7GB (4k ctx)
→ whisper:tiny on Tesla M60 (16GB, used) ($35) tiny · 0.1GB · needs ~2.1GB (4k ctx)
→ whisper on Tesla M60 (16GB, used) ($35) base · 0.3GB · needs ~2.3GB (4k ctx)
→ whisper:small on Tesla M60 (16GB, used) ($35) small · 0.9GB · needs ~2.9GB (4k ctx)
→ whisper:medium on Tesla M60 (16GB, used) ($35) medium · 3.1GB · needs ~5.1GB (4k ctx)
→ kokoro:82m on Tesla M60 (16GB, used) ($35) 82M · 0.3GB · needs ~2.3GB (4k ctx)
→ gemma3:4b on Tesla M60 (16GB, used) ($35) 4B · 3.3GB · needs ~3.8GB (4k ctx)
→ deepseek-r1:7b on Tesla M60 (16GB, used) ($35) distill 7B · 4.7GB · needs ~4.9GB (4k ctx)
→ mistral:7b on Tesla M60 (16GB, used) ($35) 7B · 4.1GB · needs ~4.6GB (4k ctx)
→ llava:7b on Tesla M60 (16GB, used) ($35) 7B · 4.7GB · needs ~6.7GB (4k ctx)
→ qwen3:8b on Tesla M60 (16GB, used) ($35) 8B · 5.2GB · needs ~5.8GB (4k ctx)
→ qwen2.5vl:7b on Tesla M60 (16GB, used) ($35) 7B · 5.6GB · needs ~5.8GB (4k ctx)
→ llama3.1:8b on Tesla M60 (16GB, used) ($35) 8B · 4.9GB · needs ~5.4GB (4k ctx)
→ deepseek-r1:8b on Tesla M60 (16GB, used) ($35) distill 8B · 4.9GB · needs ~5.4GB (4k ctx)
→ whisper:large-v3 on Tesla M60 (16GB, used) ($35) large-v3 · 6.2GB · needs ~8.2GB (4k ctx)
→ mistral-nemo:12b on Tesla M60 (16GB, used) ($35) 12B · 7.1GB · needs ~7.7GB (4k ctx)
→ sdxl on Tesla M60 (16GB, used) ($35) SDXL base · 6.9GB · needs ~8.9GB (4k ctx)
→ llava:13b on Tesla M60 (16GB, used) ($35) 13B · 8.0GB · needs ~11.1GB (4k ctx)
→ gemma3:12b on Tesla M60 (16GB, used) ($35) 12B · 8.1GB · needs ~9.6GB (4k ctx)
→ qwen3:14b on Tesla M60 (16GB, used) ($35) 14B · 9.3GB · needs ~9.9GB (4k ctx)
→ deepseek-r1:14b on Tesla M60 (16GB, used) ($35) distill 14B · 9.0GB · needs ~9.8GB (4k ctx)
→ phi4:14b on Tesla M60 (16GB, used) ($35) 14B · 9.1GB · needs ~9.9GB (4k ctx)
→ gpt-oss:20b on Tesla M40 (24GB, used) ($50) 20B MoE · 14GB · needs ~14.2GB (4k ctx)
→ devstral:24b on Tesla M40 (24GB, used) ($50) 24B · 14.6GB · needs ~15.2GB (4k ctx)
→ qwen3.8:27b 27B — 18.5 GB vs 16 GB usable — partial CPU offload, expect large speed loss
→ qwen3:30b-a3b 30B-A3B MoE — 19.0 GB vs 16 GB usable — partial CPU offload, expect large speed loss
→ qwen3-coder:30b-a3b 30B-A3B MoE — 19.0 GB vs 16 GB usable — partial CPU offload, expect large speed loss
→ gemma3:27b 27B — 18.9 GB vs 16 GB usable — partial CPU offload, expect large speed loss
→ flux:schnell schnell fp8 — 19.0 GB vs 16 GB usable — partial CPU offload, expect large speed loss
⚠ The awkward flagship: 5070 Ti bandwidth (896) in a $999-MSRP card — for LLM duty the 5070 Ti gives ~93% of the decode speed for ~25% less money; pay 5080 money only for gaming+LLM dual duty
⚠ 16GB wall is the same as every 16GB card: 14B Q4 comfortable, 32B Q4 does not fit — this is NOT a 'flagship VRAM' card despite the halo badge
⚠ LocalScore median (26.1 tok/s 8B) is a low-skewed early sample; community llama.cpp rows sit 94-140 — band is to contain both
⚠ 360W on 16-pin, 750W PSU suggested
$5000 — the only real in-family step: 32GB + 1792 GB/s doubles both fit and speed
used-market alternative: 24GB beats 16GB for 32B Q4 fits at similar decode speed