~/best-gpu-for/llama-3-3-70b
Meta provider

Best GPU for Llama 3.3 70B

Ranked recommendations from the 42-GPU catalog for running Llama 3.3 70B (70.6B params). Top picks by quality, then by cost.

Top picks.

$ ./rank --by quality
Tightest budget
RTX A6000
48GB · runs at Q3_K_M
30 GB weights · 8.2 GB headroom
Best quality
2× H100 NVL
188GB · runs at FP16/BF16
smallest card supporting the best available quant
Apple Silicon
M2 Max 64
64GB · runs at Q5_K_M
smallest M-series that fits
Datacenter (FP16 ambition)
B200
192GB · runs at FP16/BF16
best-quality datacenter card that fits

Full ranked list.

$ ./vrambudget --rank llama-3-3-70b
GPUVRAMBest quantWeights at 15% safetyFit
nvidia2× H100 NVL188GBFP16/BF16
141fitsappleM2 Ultra 192192GBFP16/BF16
141fitsnvidiaB200192GBFP16/BF16
141fitsamdMI300X192GBFP16/BF16
141fitsappleM3 Ultra 512512GBFP16/BF16
141fitsappleM3 Max 9696GBQ8_0
75fitsnvidiaRTX Pro 600096GBQ8_0
75fitsappleM4 Max 128128GBQ8_0
75fitsappleM5 Max 128128GBQ8_0
75fitsnvidiaDGX Spark128GBQ8_0
75fitsintelGaudi 3128GBQ8_0
75fitsnvidiaH200141GBQ8_0
75fitsnvidiaH100 80GB80GBQ6_K
58fitsappleM2 Max 6464GBQ5_K_M
49fitsappleM3 Max 6464GBQ5_K_M
49fitsappleM4 Pro 6464GBQ5_K_M
49fitsappleM5 Pro 6464GBQ5_K_M
49fitsnvidiaRTX A600048GBQ3_K_M
30fits

Next steps.

$ ./next
$ checkCan I run Llama 3.3 70B?see all quants →full quant matrix$ tuneTune the mathopen calculator →/calc?model=llama-3-3-70b

Discussion.

$ gh discussion list

// sign in with github to leave a comment. threads live in the repo's discussions tab.