GPU Models Dataset

Name: GPU models and specifications
Creator: GetDeploying
License: https://creativecommons.org/licenses/by/4.0/

Specifications for 107 GPU models.

Free with attribution · CC BY 4.0 · updated on change

GPU models: 107; from datasheets
Rows: 112; one per model and form factor
Fields: 36; CSV columns
Updated: 14 Sep 2026; on change

Download

One row per GPU model and form factor. Figures are as the vendor publishes them.

# Specs alone: datacenter cards with 80 GB or more, fastest memory first
import pandas as pd

gpus = pd.read_csv("https://getdeploying.com/dataset/gpu-models/gpu-models.csv")
big = gpus[(gpus.market_class == "DATACENTER") & (gpus.vram_gb >= 80)]
big.sort_values("mem_bandwidth_gbps", ascending=False)[["gpu_slug", "form_factor", "vram_gb", "mem_bandwidth_gbps", "tdp_w"]]

# With prices: what a GB of VRAM rents for, per model and week
import pandas as pd

gpus = pd.read_csv("https://getdeploying.com/dataset/gpu-models/gpu-models.csv")
prices = pd.read_csv("https://getdeploying.com/dataset/gpu-prices/weekly.csv")
joined = prices.merge(gpus, on="gpu_slug")
joined["usd_per_gb"] = joined.median_price / joined.vram_gb

Per-provider rows, daily data and longer history are in the API.

Every GPU model

One file, keyed by gpu_slug. 112 rows.

Download CSV · 33 KB JSON

Preview

gpu-models.csv · 6 of 112 rows

gpu_slug	name	vendor	release_date	form_factor	vram_gb	memory_type	mem_bandwidth_gbps	memory_bus_bits	memory_ecc	nvlink_bandwidth_gbps	pcie_gen	pcie_lanes	fp4_dense_tflops	fp4_sparse_tflops	fp8_dense_tflops	fp8_sparse_tflops	fp16_dense_tflops	fp16_sparse_tflops	int8_dense_tops	int8_sparse_tops	fp32_dense_tflops	fp64_dense_tflops	architecture	process	transistors_b	shader_cores	matrix_cores	compute_units	tdp_w	cooling	mig_instances	market_class	supported_precisions	datasheet_url
nvidia-h100	Nvidia H100	Nvidia	2022-09-20	SXM	80	HBM3	3350	5120	true	900	5	16			1979	3958	989.5	1979	1979	3958	67	34	Hopper	TSMC 4N	80	16896	528	132	700		7	DATACENTER	FP8\|FP16\|BF16\|TF32\|FP32\|FP64\|INT8	https://resources.nvidia.com/en-us-gpu-resources/h100-datasheet-24306
nvidia-b300	Nvidia B300	Nvidia	2025-03-18		288	HBM3e	8000	8192	true	1800	6	16	14000		4500	9000	2250	4500	153.5	307	75	1.2	Blackwell Ultra	TSMC 4NP	208				1100	passive	7	DATACENTER	FP4\|FP6\|FP8\|FP16\|BF16\|TF32\|FP32\|FP64\|INT8	https://resources.nvidia.com/en-us-blackwell-architecture/blackwell-ultra-datasheet
nvidia-b200	Nvidia B200	Nvidia	2024-03-18		180	HBM3e	7700		true	1800	5	16	9000	18000	4500	9000	2250	4500	4500	9000	75	37	Blackwell	TSMC 4NP	208				1000	passive	7	DATACENTER	FP4\|FP6\|FP8\|FP16\|BF16\|TF32\|FP32\|FP64\|INT8	https://dam-cdn.nvd.orangelogic.com/AssetLink/y441155802qub41q118b2852i557jem5.pdf
nvidia-rtx-pro-6000	Nvidia RTX PRO 6000	Nvidia	2025-03-18	PCIe	96	GDDR7	1597	512	true		5	16	2000	4000	1000	2000	500	1000	1000	2000	120		Blackwell	TSMC 4N	92.2	24064	752	188	600	passive		WORKSTATION	FP4\|FP8\|FP16\|BF16\|TF32\|FP32\|INT8	https://resources.nvidia.com/en-us-rtx-pro-6000
nvidia-h200	Nvidia H200	Nvidia	2023-11-13		141	HBM3e	4800		true	900	5	16			1979	3958	989.5	1979	1979	3958	67	34	Hopper	TSMC 4N	80	16896	528	132	700		7	DATACENTER	FP8\|FP16\|BF16\|TF32\|FP32\|FP64\|INT8	https://resources.nvidia.com/en-us-hopper-architecture/hpc-datasheet-sc23
nvidia-rtx-5090	Nvidia RTX 5090	Nvidia	2025-01-30		32	GDDR7	1792	512	false		5		1676	3352	419	838	209.5	419	838	1676	104.8		Blackwell	TSMC 4N	92.2	21760	680	170	575	active		CONSUMER	FP4\|FP8\|FP16\|BF16\|TF32\|FP32\|INT8	https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5090/

Fields

The CSV is flat, one column per field. The JSON groups the same fields under the headings below.

Field	Type	Definition
Model
`gpu_slug`	text	The GPU model's slug, as in its page URL (supported slugs)
`name`	text	The model's name, with its vendor
`vendor`	text	The card's vendor
`release_date`	date, ISO 8601	Launch date
`form_factor`	text	The card's form factor; a model has one row for each
Memory
`vram_gb` memory.capacity_gb	GB	Memory capacity
`memory_type` memory.type	text	Memory technology
`mem_bandwidth_gbps` memory.bandwidth_gbps	GB/s	Memory bandwidth
`memory_bus_bits` memory.bus_bits	bit	Memory bus width
`memory_ecc` memory.ecc	true / false	Whether the memory is error-correcting
Interconnect and host
`nvlink_bandwidth_gbps` interconnect.gpu_to_gpu_gbps	GB/s	GPU-to-GPU interconnect bandwidth (NVLink, Infinity Fabric); empty for a PCIe-only card
`pcie_gen` interconnect.pcie_gen	integer	PCIe generation of the host interface
`pcie_lanes` interconnect.pcie_lanes	integer	PCIe lanes of the host interface
Compute
`fp4_dense_tflops` compute.fp4.dense_tflops	TFLOPS	FP4 throughput, dense
`fp4_sparse_tflops` compute.fp4.sparse_tflops	TFLOPS	FP4 throughput with structured sparsity
`fp8_dense_tflops` compute.fp8.dense_tflops	TFLOPS	FP8 throughput, dense
`fp8_sparse_tflops` compute.fp8.sparse_tflops	TFLOPS	FP8 throughput with structured sparsity
`fp16_dense_tflops` compute.fp16.dense_tflops	TFLOPS	FP16 / BF16 throughput, dense
`fp16_sparse_tflops` compute.fp16.sparse_tflops	TFLOPS	FP16 / BF16 throughput with structured sparsity
`int8_dense_tops` compute.int8.dense_tops	TOPS	INT8 throughput, dense
`int8_sparse_tops` compute.int8.sparse_tops	TOPS	INT8 throughput with structured sparsity
`fp32_dense_tflops` compute.fp32.dense_tflops	TFLOPS	FP32 throughput, dense
`fp64_dense_tflops` compute.fp64.dense_tflops	TFLOPS	FP64 throughput, dense
Silicon
`architecture` silicon.architecture	text	As the vendor names it (Hopper, CDNA 3)
`process` silicon.process	text	Manufacturing process (TSMC 4N)
`transistors_b` silicon.transistors_b	billion	Transistor count
`shader_cores` silicon.shader_cores	integer	CUDA cores (Nvidia), stream processors (AMD) or Xe-cores (Intel)
`matrix_cores` silicon.matrix_cores	integer	Tensor cores (Nvidia) or matrix cores (AMD)
`compute_units` silicon.compute_units	integer	Streaming multiprocessors (Nvidia) or compute units (AMD)
Power
`tdp_w` power.tdp_w	W	Board power
`cooling` power.cooling	text	Cooling design
Platform
`mig_instances` platform.mig_instances	integer	Hardware partitions on an Nvidia card with MIG
`partitions` platform.partitions	integer	Hardware partitions under another partitioning scheme
Other
`market_class`	text	The product line
`supported_precisions`	text list	Number formats the silicon supports, pipe-separated in the CSV
`datasheet_url`	URL	The vendor's datasheet

License and attribution

CC BY 4.0: Credit GetDeploying and link to the license.

GetDeploying, GPU models and specifications, https://getdeploying.com/gpus, CC BY 4.0

GetDeploying (2026). GPU models and specifications [dataset]. https://getdeploying.com/dataset/gpu-models. Accessed September 21, 2026.

The license covers the published dataset files only and implies no endorsement. The site and the API stay under the terms.

Common questions

Where do the figures come from?

Each vendor's own datasheet or product page, read by hand into a fixed vocabulary so that any two cards line up in a column. The datasheet is linked on every row.

Why are dense and sparse throughput separate columns?

Nvidia's datasheets lead with the structured-sparsity figure and AMD's lead with dense, so a single throughput column would compare the two vendors on different bases. Dense is stated for every card that publishes a rate; sparse only where the datasheet states one.

Why does a model have several rows?

A card sold in several form factors, such as an H100 in SXM, PCIe and NVL, has one row per form factor, because memory capacity, bandwidth and interconnect differ between them while the silicon is the same. Most models are one row, with form_factor empty when the catalog does not state it.

Why is a value empty?

The vendor does not publish it, or it does not apply: a PCIe-only card has no interconnect bandwidth, and a card without tensor cores has no FP16 rate. Nothing is estimated to fill a gap.

How often does it update?

When a model is added or a figure is corrected, which is a few times a month. The file refreshes within a day of a catalog change.

How is this different from the API?

The API lists the same models with their live rental offerings and prices per provider, under a documented contract. This file is the specifications alone, with no stability promise: its shape can change at any time.

See also the GPU price dataset, which joins on gpu_slug, and the GPU price comparison. A provider's offering can differ from the headline card; its own listing applies.