GPU-Visualiser

[This README was written by a human]

A tool to visualise GPU architectures.

Access here: https://madhavmenon10.github.io/gpu-visualiser/

Note

Every floorplan is built from the vendor's own published architecture. Units shown as hatched blocks are fused off i.e. permanently disabled at the factory by blowing tiny on-chip fuses, so a die with a manufacturing defect in that region can still be sold rather than scrapped. This is why a shipping part often has fewer active cores than the full die. This is also covered within the tool.

Screenshots

The die overview, with the whole floorplan in view and the legend mapping colours to block types:

Click a block to pin it and read its documentation. Here the GigaThread Engine on the NVIDIA GeForce RTX 4090 is selected:

Supported GPUs

Each entry cites the sources its data is drawn from. The numbers map to the References section.

NVIDIA

GPU	Die	Architecture	ISA	Process	Sources
H100 SXM5	GH100	Hopper	sm_90	TSMC 4N	[1], [9]
A100 SXM4 80GB	GA100	Ampere	sm_80	TSMC N7	[2], [9]
GeForce RTX 4090	AD102	Ada Lovelace	sm_89	TSMC 4N	[3], [9]
GeForce RTX 3090	GA102	Ampere (GA10x)	sm_86	Samsung 8N	[4], [9]
V100 SXM2	GV100	Volta	sm_70	TSMC 12FFN	[5], [9]

AMD

GPU	Die	Architecture	ISA	Process	Sources
Radeon RX 7900 XTX	Navi 31	RDNA 3	gfx1100	TSMC N5 + N6	[6], [9]
Instinct MI300X	Aqua Vanjaram	CDNA 3	gfx942	TSMC N5 + N6	[7], [9]

Intel

GPU	Die	Architecture	ISA	Process	Sources
Arc A770 16GB	ACM-G10	Xe-HPG (Alchemist)	Xe-HPG / DG2-512	TSMC N6	[8], [9]

Build and run

The project is a Vite + React + TypeScript app. You need Node.js 18 or newer.

# install dependencies
npm install

# start the dev server (hot reload) at http://localhost:5173
npm run dev

# type-check and produce a production build in dist/
npm run build

# serve the production build locally
npm run preview

You can view the current state of the project through the published GitHub pages link: https://madhavmenon10.github.io/gpu-visualiser/

Contributing

See CONTRIBUTING.md

References

Architecture descriptions come primarily from each vendor's own whitepaper or instruction-set documentation. Die area, transistor counts and clock speeds, where the whitepaper does not state them, are taken from the TechPowerUp GPU Database.

1. NVIDIA — NVIDIA H100 Tensor Core GPU Architecture (Hopper whitepaper). See also the Hopper Architecture In-Depth technical blog.
2. NVIDIA — NVIDIA A100 Tensor Core GPU Architecture (Ampere whitepaper).
3. NVIDIA — NVIDIA Ada GPU Architecture whitepaper.
4. NVIDIA — NVIDIA Ampere GA102 GPU Architecture whitepaper.
5. NVIDIA — NVIDIA Tesla V100 GPU Architecture (Volta whitepaper).
6. AMD — RDNA 3 Instruction Set Architecture Reference Guide and the Radeon RX 7900 XTX product page.
7. AMD — AMD CDNA 3 Architecture white paper and the AMD Instinct MI300 series microarchitecture (ROCm documentation).
8. Intel — Introduction to the Xe-HPG Architecture white paper and the Intel Arc A770 Graphics (16GB) product specifications.
9. TechPowerUp — GPU Database, secondary source for die area, transistor counts and clock speeds.