Compute-cost ranking: How much compute you get for the same budget

Value-for-money = currently monitored lower hourly rate ÷ FP16 Tensor dense peak compute · 22 GPU models included · Data updated: 2026-11-06

⚡ Direct Answer: The most cost-efficient GPU per TFLOPS is currently MI300X 192GB (¥0.003/TFLOPS·hour, converted from RunPod (US)’s lowest hourly spot rate), costing just 1/21.3 of the priciest RTX 6000 Ada 48GB on this list (¥0.059). Price ranking by hourly rate differs significantly from value ranking by TFLOPS—choose from this table for training workloads, and from the hourly-rate table for inference/light tasks.

GPU model	FP16 nominal compute	VRAM	Bandwidth	Internet	Currently lower hourly rate	¥/TFLOPS·hour	Lowest-Price Platform
MI300X 192GB Flagship Data Center	1307 TFLOPS	192GB	5300 GB/s	Infinity Fabric 896GB/s	¥3.60/hourMost cost-effective compute per unit Spot	¥0.003/TFLOPS·hour	RunPod (USA)
RTX 4090 Premium Spending	165 TFLOPS	24GB	1008 GB/s	PCIe	¥0.87/hour Spot	¥0.005/TFLOPS·hour	Vast.ai (USA)
H100 80GB Flagship Data Center	989 TFLOPS	80GB	3350 GB/s	NVLink 900GB/s	¥8.28/hour Pay-as-you-go	¥0.008/TFLOPS·hour	Paperspace (USA)
RTX 3090 Premium Spending	71 TFLOPS	24GB	936 GB/s	PCIe	¥0.63/hour Spot	¥0.009/TFLOPS·hour	Vast.ai (USA)
H800 80GB Flagship Data Center	989 TFLOPS	80GB	3350 GB/s	NVLink 400GB/s	¥10.50/hour Pay-as-you-go	¥0.011/TFLOPS·hour	AutoDL
V100 16GB Entry-level training	125 TFLOPS	16GB	900 GB/s	NVLink 300GB/s	¥1.37/hour Spot	¥0.011/TFLOPS·hour	RunPod (USA)
A100 40GB Data Center	312 TFLOPS	40GB	1555 GB/s	NVLink 600GB/s	¥3.45/hour Pay-as-you-go	¥0.011/TFLOPS·hour	AutoDL
Ascend 910B2 64GB Domestic GPU	376 TFLOPS	64GB	None GB/s	HCCS	¥4.20/hour Pay-as-you-go	¥0.011/TFLOPS·hour	Gongji Compute
L40S 48GB Data Center	362 TFLOPS	48GB	864 GB/s	PCIe	¥4.97/hour Pay-as-you-go	¥0.014/TFLOPS·hour	Cloudrift (USA)
V100 32GB Entry-level training	125 TFLOPS	32GB	900 GB/s	NVLink 300GB/s	¥1.98/hour Pay-as-you-go	¥0.016/TFLOPS·hour	AutoDL
A800 80GB Data Center	312 TFLOPS	80GB	2039 GB/s	NVLink 400GB/s	¥5.15/hour Pay-as-you-go	¥0.017/TFLOPS·hour	MornAI
H200 141GB Flagship Data Center	989 TFLOPS	141GB	4800 GB/s	NVLink 900GB/s	¥16.49/hour Spot	¥0.017/TFLOPS·hour	Jarvislabs (India)
A40 48GB Data Center	150 TFLOPS	48GB	696 GB/s	PCIe	¥2.52/hour Spot	¥0.017/TFLOPS·hour	RunPod (USA)
L4 24GB Beginner Inference	121 TFLOPS	24GB	300 GB/s	PCIe	¥2.09/hour Spot	¥0.017/TFLOPS·hour	Jarvislabs (India)
B200 180GB Flagship Data Center	2250 TFLOPS	180GB	8000 GB/s	NVLink 1.8TB/s	¥39.11/hour Pay-as-you-go	¥0.017/TFLOPS·hour	Massed Compute (USA)
A100 80GB Data Center	312 TFLOPS	80GB	2039 GB/s	NVLink 600GB/s	¥5.47/hour Pay-as-you-go	¥0.018/TFLOPS·hour	Paperspace (USA)
T4 16GB Beginner Inference	65 TFLOPS	16GB	320 GB/s	PCIe	¥1.35/hour Pay-as-you-go	¥0.021/TFLOPS·hour	MatPool
H20 96GB Data Center	148 TFLOPS	96GB	4000 GB/s	NVLink 900GB/s	¥3.40/hour Spot	¥0.023/TFLOPS·hour	PPIO Paiou Cloud
L40 48GB Data Center	181 TFLOPS	48GB	864 GB/s	PCIe	¥4.50/hour Pay-as-you-go	¥0.025/TFLOPS·hour	Gongji Compute
A6000 48GB Professional GPU	77 TFLOPS	48GB	768 GB/s	PCIe	¥2.38/hour Spot	¥0.031/TFLOPS·hour	RunPod (USA)
A10 24GB Beginner Inference	125 TFLOPS	24GB	600 GB/s	PCIe	¥6.12/hour Pay-as-you-go	¥0.049/TFLOPS·hour	Paratera
RTX 6000 Ada 48GB Professional GPU	91 TFLOPS	48GB	576 GB/s	PCIe	¥5.33/hour Spot	¥0.059/TFLOPS·hour	RunPod (USA)

Definition: Compute refers to vendor-published FP16/BF16 Tensor dense peak spec (sparse acceleration excluded);\n Variants of the same GPU model (SXM/PCIe/OAM) differ in performance; actual platform configurations follow spec pages; value-for-money is calculated using currently monitored lower hourly rate;\n Spot/low-usage pricing fluctuations may shift rankings. Peak spec ≠ actual utilization — real throughput depends on workload, driver, interconnect, and memory bottlenecks.\n Use this table as a reference for “maximum compute-per-budget ceiling”; pair it with selection tool to choose GPUs by workload.\n 10 GPU models (including some domestic and new releases) lack comparable published specs from vendors and are excluded from this ranking; their price rankings appear in homepage price comparison table.

How to use this table?

① For training/fine-tuning, compare compute per ¥/TFLOPS·hour — but exclude GPUs with insufficient VRAM (VRAM specs listed on GPU model pages);

② For inference, choose the lowest hourly rental rate among GPUs with sufficient VRAM; Homepagesorted by hourly rental rate;

③ For multi-GPU training, prioritize interconnect bandwidth (NVLink/HCCS); PCIe-based multi-GPU setups suffer high communication overhead.