Download gpus.csv from thanatchon36/one-gpu-capacity: direct link, hf CLI and curl.
- Browser
- Download file 2.74 kB
-
https://huggingface.co/thanatchon36/one-gpu-capacity/resolve/main/gpus.csv
- Command line
-
hf download hf://thanatchon36/one-gpu-capacity/gpus.csv
-
curl -L -o gpus.csv https://huggingface.co/thanatchon36/one-gpu-capacity/resolve/main/gpus.csv
2.74 kB
| gpu,label,tier,vram_gb,bw_gbs,fp16_tflops,tdp_w,msrp_usd,price_central_usd,price_low_usd,price_high_usd,host_usd,form,source_specs,notes | |
| rtx3060,RTX 3060 12GB,budget,12,360,25.5,170,329,300,250,350,900,pcie,gpu.json:NVIDIA_GA106-300-A1,FP16 tensor halved for FP32 accumulate (GeForce) | |
| rtx4060ti,RTX 4060 Ti 16GB,budget,16,288,44.0,165,499,450,400,500,900,pcie,gpu.json:NVIDIA_AD106-351,FP16 tensor halved for FP32 accumulate (GeForce) | |
| rtx5060ti,RTX 5060 Ti 16GB,budget,16,448,47.5,180,429,450,429,550,900,pcie,gpu.json:NVIDIA_GB206-300_2,FP16 tensor halved for FP32 accumulate (GeForce; assumed as Ada) | |
| rtx3090,RTX 3090 24GB,enthusiast,24,936,71.2,350,1499,800,650,1000,1200,pcie,gpu.json:NVIDIA_GA102-300-A1,price is used-market assumption; FP16 halved for FP32 accumulate | |
| rtx4090,RTX 4090 24GB,enthusiast,24,1008,165.2,450,1599,1800,1599,2500,1200,pcie,gpu.json:NVIDIA_AD102-300,FP16 halved for FP32 accumulate | |
| rtx5090,RTX 5090 32GB,enthusiast,32,1792,209.5,575,1999,2500,1999,3500,1500,pcie,gpu.json:NVIDIA_GB202-300,FP16 halved for FP32 accumulate (assumed as Ada) | |
| a6000,RTX A6000 48GB,workstation,48,768,154.8,300,4649,4500,3500,5000,2500,pcie,gpu.json:NVIDIA_GA102-875-A1,dense = sparse/2 | |
| rtx6000ada,RTX 6000 Ada 48GB,workstation,48,960,364.2,300,6799,7000,6000,7500,2500,pcie,gpu.json:NVIDIA_AD102-870-A1 + manual FP16,gpu.json FP16 field is FP8-sparse; FP16 dense from vendor datasheet (verify) | |
| rtxpro6000,RTX PRO 6000 Blackwell 96GB,workstation,96,1792,312.5,600,8565,8565,7500,9500,2500,pcie,gpu.json:NVIDIA_GB202_3,dense = listed sparse/2 (verify) | |
| t4,T4 16GB,dc-inference,16,320,64.8,70,,1500,1000,2500,4000,pcie,gpu.json:NVIDIA_nan_56,no MSRP; price assumed | |
| l4,L4 24GB,dc-inference,24,300,121.0,72,,2500,2000,3500,4000,pcie,gpu.json:NVIDIA_nan_69,no MSRP; price assumed | |
| a10,A10/A10G 24GB,dc-inference,24,600,125.0,150,,3000,2000,4000,4000,pcie,gpu.json:NVIDIA_nan_58,A10G (AWS) treated as A10; price assumed | |
| l40s,L40S 48GB,dc-inference,48,864,362.0,350,,8000,7000,10000,4000,pcie,manual (vendor datasheet; verify),not in gpu.json; price assumed | |
| a100,A100 SXM 80GB,flagship,80,2039,312.0,400,,12000,9000,17000,8000,sxm,manual bandwidth for 80GB SXM + gpu.json:NVIDIA_nan_62 compute,gpu.json lists 40GB PCIe (1555 GB/s); price assumed | |
| h100,H100 SXM 80GB,flagship,80,3352,989.4,700,,25000,20000,35000,8000,sxm,gpu.json:NVIDIA_nan_64,price assumed | |
| h200,H200 SXM 141GB,flagship,141,4800,989.4,700,,32000,28000,40000,8000,sxm,gpu.json:NVIDIA_nan_66 + manual bandwidth,gpu.json lists 3360 GB/s; vendor datasheet 4800 GB/s used (verify) | |
| b200,B200 180GB,flagship,180,8000,2250.0,1000,,40000,30000,50000,8000,sxm,gpu.json:NVIDIA_nan_71 + manual,gpu.json lists 2x96 GB / 2x4100 GB/s; 180 GB usable and 8 TB/s per vendor DGX spec (verify) | |