Hopper
Nvidia's datacentre GPU architecture, first shipping in the H100 accelerator
Hopper is a graphics processing unit microarchitecture from Nvidia, named after computer scientist and US Navy rear admiral Grace Hopper. It was revealed in March 2022 and built on TSMC's N4 process, powering the Tesla H series of datacentre accelerators with a new streaming multiprocessor, faster memory and a transformer engine.
- Revealed March 2022, named after Grace Hopper
- H100 built on TSMC N4 with 80 billion transistors
- Up to 144 streaming multiprocessors on the H100
- HBM3 memory bandwidth of 3 TB/s, 50% above A100
- First Nvidia architecture with a transformer engine
What it is
Hopper is a GPU microarchitecture designed by Nvidia for datacentres, announced at the company's GTC event in March 2022. It is named after Grace Hopper, the computer scientist and US Navy rear admiral. The design improves on the earlier Turing and Ampere generations with a rebuilt streaming multiprocessor, a quicker memory subsystem and hardware aimed at transformer models. Its best known product is the H100 accelerator, made on TSMC's N4 process with 80 billion transistors and up to 144 streaming multiprocessors. Hopper runs alongside the Ada Lovelace architecture, which covers consumer and professional graphics, and it was succeeded by Blackwell.
How it came about
The name first surfaced in November 2019, when a well-known Twitter account posted that the generation after Ampere would be called Hopper and would use a multi-chip module approach to raise yields and cut waste. Nvidia formally unveiled the architecture at its March 2022 GTC. Demand for the H100 surged during the 2023 AI boom; Oracle's Larry Ellison described dinner conversations with Jensen Huang, Elon Musk and himself as an hour of sushi and begging for the chips. Export rules later reshaped the line, with the H800 built for China in late 2022 and the H20 following in late 2023.
Inside the chip
Each streaming multiprocessor keeps Ampere's limit of 64 concurrent warps but adds a Tensor Memory Accelerator that moves tensors of up to five dimensions between shared and global memory without touching registers. Thread block clusters let blocks run atomics against each other's shared memory, with a portable cluster size of 8 and up to 16 on the H100. FP32 throughput per SM doubles versus the previous generation, and new instructions cover the Smith-Waterman and Needleman-Wunsch algorithms. A transformer engine shifts work between FP16 and FP8 at runtime, rebalancing mantissa and exponent bits to hold precision. Cache and memory grow too: 256 KB of combined L1, texture cache and shared memory per SM, 50 MB of L2, and HBM3 reaching 3 TB/s.
How it was made
The H100 is built on TSMC's N4 process and carries 80 billion transistors across up to 144 streaming multiprocessors. It ships in both SXM5 and PCIe forms, with the SXM5 socket delivering more memory bandwidth and therefore better performance. The SXM5 variant is rated at 700 W. The GH200 module pairs a Hopper H100 with a 72-core Grace CPU over NVLink, drawing up to 1000 W in total.
Where it went
Hopper targets datacentres, appearing in Nvidia's Tesla H series and in DGX systems. The H100 supports HBM3 and HBM2e memory up to 80 GB, with HBM3 bandwidth of 3 TB/s, half again the 2 TB/s of the Ampere A100. NVLink gains a faster generation for chip-to-chip traffic. The GH200 superchip combines the GPU with a Grace CPU on one module, keeping CPU and GPU memory coherent over NVLink. Export-controlled derivatives, the H800 and H20, serve the Chinese market.
Versions and family
The H100 is the flagship part, offered in SXM5 and PCIe configurations. The H800, created in late 2022 for China under US export limits, has lower bandwidth than the standard H100. After further restrictions in late 2023, Nvidia produced the H20, a trimmed H100 variant that became the leading Nvidia chip in China by 2025. The H200 is another Hopper-based accelerator. On the consumer and professional side, Nvidia runs the related Ada Lovelace architecture, while Blackwell replaces Hopper at the top end.
Why it matters
Hopper arrived just as demand for AI training hardware exploded. Analysts at Raymond James Financial put the H100's selling price at $25,000 to $30,000 in January 2024, with individual units above $40,000 on eBay, and by February 2024 shipments were reportedly moving to data centres in armoured cars. The architecture also became a bargaining chip in trade policy: the H200 drew US export approvals and Chinese import friction through 2026, and its sale to named Chinese firms featured in a presidential state visit that May.
Problems and reception
Reviewers and programmers focused on the memory and scheduling changes, particularly the Tensor Memory Accelerator, distributed shared memory and the transformer engine. Memory synchronisation domains were added to stop fence and flush operations from stalling on writes that only appeared visible by chance timing. The main controversy has been trade rather than silicon: US export controls forced the H800 and H20 reworkings, and in early 2026 Chinese customs reportedly held up H200 shipments even after Washington cleared them, leaving suppliers in limbo.
Questions people ask about Hopper
What is Hopper?
Hopper is Nvidia's GPU microarchitecture for datacentres, announced in March 2022 and named after Grace Hopper. It underpins the Tesla H series of accelerators, most famously the H100, and improves on Turing and Ampere with a new streaming multiprocessor, faster memory and transformer acceleration.
Who makes the chips?
Nvidia designs them and TSMC manufactures the H100 on its N4 process. The die carries 80 billion transistors and up to 144 streaming multiprocessors, and it is sold in SXM5 and PCIe forms.
What does the transformer engine do?
It speeds up transformer model work by dropping calculations from higher precision such as FP16 to faster lower precision such as FP8 whenever the loss in accuracy is acceptable. It can also shift bits between mantissa and exponent at runtime to keep as much precision as possible.
Why are there several Hopper chips for China?
US export rules limited what Nvidia could sell there. The H800 arrived in late 2022 with reduced bandwidth, and after tighter rules in late 2023 the H20 followed as a modified H100. By 2025 the H20 was the most prominent Nvidia chip in the Chinese market.
How much memory does an H100 have?
It supports HBM3 and HBM2e memory up to 80 GB. The HBM3 system runs at 3 TB/s, which is 50% more bandwidth than the 2 TB/s of the Ampere A100, and each streaming multiprocessor gets 256 KB of combined L1, texture cache and shared memory.
Good reading and sources
- CUDA Programming Model for Hopper Architecture A41095 | GTC Digital September 2022 | NVIDIA On-DemandNvidia, Article
- Nvidia reveals H100 GPU for AI and teases ‘world’s fastest AI supercomputer’The Verge, Article
- Nvidia’s Next GPU Shows That Transformers Are Transforming AIIEEE Spectrum, Article
- Deep Dive Into Nvidia’s “Hopper” GPU ArchitectureThe Next Platform, Article
The facts on this page come from the Wikipedia article "Hopper (microarchitecture)", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryChip rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.
Like Hopper
All graphics processorsMore from Nvidia
All of them
🇺🇸 United States and 1 moreApple A16 Bionic
ARM-based system on a chip for iPhone and iPadSoC · Arm · 2022