GPU specs chart

Dense and with sparsity are different measurements. NVIDIA's datasheets print tensor throughput with a footnote that reads "with sparsity". That figure assumes a model pruned to a 2:4 pattern and is exactly twice what the same GPU does on an ordinary dense model. Almost nobody serves pruned models, so the dense column is the one to compare. For the H100, the datasheet headline is 1,979 TFLOPS at FP16 with sparsity; the dense figure is 989 TFLOPS.

A blank cell is information. FP8 arrived with Ada Lovelace and Hopper, and FP4 with Blackwell, so older GPUs have nothing to publish at those precisions. A model quantised to FP8 still runs on them, but through a slower software path. Consumer GeForce pages leave out FP64, and AMD and Intel publish fewer tensor figures than NVIDIA. We never fill a gap with an estimate.

For LLM inference, read memory first. VRAM decides whether the model fits at all, and memory bandwidth sets the ceiling on tokens per second for a single request, because every generated token reads the weights once. Throughput in TFLOPS matters most for training and for serving many requests at once. To turn a model into a VRAM number, use the LLM VRAM calculator or look the model up in LLM GPU requirements.

TDP is the power budget of the GPU alone. A server adds CPUs, memory, fans and network cards on top. The rent vs buy calculator uses this column to cost the electricity of owning the card.

NVIDIA data center parts are named for their architecture: the letter in front is the architecture's initial, and a bigger number within a letter is the bigger or later part. A leading G means the GPU is packaged with a Grace CPU. GeForce and RTX PRO cards count generations in the first digits and the tier in the last two. Suffixes name the form factor or link: PCIe cards fit any server, SXM modules sit on an HGX baseboard with NVLink between all GPUs, and NVL parts are PCIe cards bridged in pairs.

• Maxwell (NVIDIA, from 2015): Quadro M4000
• Pascal (NVIDIA, from 2016): GTX 1070, GTX 1080, P100, Quadro P4000, Quadro P5000, Quadro P6000, TITAN Xp
• Volta (NVIDIA, from 2017): TITAN V, V100
• Turing (NVIDIA, from 2018): Quadro RTX 4000, Quadro RTX 5000, Quadro RTX 6000, Quadro RTX 8000, RTX 2060, RTX 2070, RTX 2080, T4
• Ampere (NVIDIA, from 2020): A10, A100, A16, A30, A40, RTX 3060, RTX 3070, RTX 3080, RTX 3090, RTX A2000, RTX A4000, RTX A5000, RTX A6000
• CDNA 2 (AMD, from 2021): MI250X
• Ada Lovelace (NVIDIA, from 2022): L4, L40, L40S, RTX 2000 Ada, RTX 4000 Ada, RTX 4060, RTX 4070, RTX 4080, RTX 4090, RTX 4500 Ada, RTX 5000 Ada, RTX 5880 Ada, RTX 6000 Ada
• Gaudi (Intel, from 2022): Gaudi 2
• Hopper (NVIDIA, from 2022): GH200, H100, H200
• CDNA 3 (AMD, from 2023): MI300X, MI325X
• Blackwell (NVIDIA, from 2024): B200, RTX 5060, RTX 5070, RTX 5080, RTX 5090, RTX PRO 6000
• Blackwell Ultra (NVIDIA, from 2025): B300, GB300
• CDNA 4 (AMD, from 2025): MI355X

The full story, generation by generation, is in NVIDIA GPU generations explained.

The figures come from GPUPerHour's spec database, one row per GPU family, served by our public /api/gpu-specs endpoint. Each row was checked by hand against the vendor's datasheet or product page, which is linked in the last column with the date of the check (this release: 15 Sep 2026 and 13 Sep 2026). Where a vendor prints only a with-sparsity figure, the dense figure is that number halved, which is how the vendor defines it. Where a vendor prints nothing, the cell says so.

Prices are not part of the spec data. They are read from live provider listings when the page renders: the lowest on-demand price per GPU-hour among offers that are in stock, from secure providers, seen in the last 15 minutes. MIG slices, fractional vGPU plans and laptop parts are left out. If the price feed is down the column says so instead of showing an old number.

The data is free to reuse under CC BY 4.0; please cite GPUPerHour and link to this page. Found a wrong figure? Tell us through the contact page and we will check it against the datasheet.