How to read this page. Each section starts Simple, then goes to a Deep dive: stop wherever you have what you need. The small numbers are sources: click one to open the original document. Where the maker has not published a figure -- a price, a die size, a factory address -- this page says so rather than estimate it.
1.At a glance
SimpleStart here
The H100 is the NVIDIA chip most associated with the generative-AI boom that followed ChatGPT. Announced 22 March 2022 at GTC 1, NVIDIA declared it “in full production” on 20 September 2022, shipping from October 2022 2.
Deep diveThe technical detail
Codename GH100, architecture Hopper. Ships as SXM5 (80GB HBM3) and PCIe (80GB HBM2e) variants, plus the dual-GPU H100 NVL PCIe card 34. By SC22 (14 November 2022), NVIDIA announced worldwide partner rollout: 50+ server models by end-2022, more in H1 2023 5.
2.Launch and history
SimpleStart here
H100 launched as the successor to A100 at a moment when demand for AI training hardware was about to explode with the release of ChatGPT eight months later. NVIDIA’s own materials frame it around one new capability above all others: the Transformer Engine, built for exactly the kind of model that would define the following two years.
Deep diveThe technical detail
NVIDIA announced Hopper at GTC on 22 March 2022 alongside DGX H100 16. Full production was declared 20 September 2022, with shipping starting the following month 2. By SC22 in November 2022, over 50 server models from OEM partners were promised by year-end 5. H100 went on to become the primary chip used to train the large language models released through 2023–2024.
3.What’s inside it
SimpleStart here
H100’s headline feature is the Transformer Engine, which automatically switches between 8-bit and 16-bit number formats to speed up exactly the kind of neural network (a “transformer”) that powers ChatGPT and similar systems. NVIDIA also built H100 as the first GPU it calls a native confidential-computing device.
Deep diveThe technical detail
Three claims from NVIDIA’s own architecture blog 3:
- Transformer Engine — “intelligently manages and dynamically chooses between FP8 and 16-bit calculations,” delivering “up to 9x faster AI training and up to 30x faster AI inference speedups on large language models.”
- FP8 number formats (E4M3/E5M2) — “halving data storage requirements and doubling throughput compared to FP16 or BF16.”
- Confidential Computing — NVIDIA states “H100 implements the world’s first native Confidential Computing GPU,” with hardware isolation supported down to individual MIG-instance granularity.
4.Full spec table
SimpleStart here
16,896 CUDA cores (SXM5 variant), 80GB of high-bandwidth memory, and NVLink connecting up to 8 chips at 900 GB/s 3.
Deep diveThe technical detail
| Spec | H100 SXM5 | H100 PCIe |
|---|
| Process node | TSMC 4N (custom) 3 |
|---|
| Transistors | 80 billion 3 |
|---|
| Die size | 814 mm² 3 |
|---|
| CUDA cores | 16,896 (132 SMs) | 14,592 (114 SMs) |
|---|
| Memory | HBM3, 80GB | HBM2e, 80GB |
|---|
| Memory bandwidth | 3.35 TB/s 7 | 2 TB/s 7 |
|---|
| TDP | Up to 700 W (configurable) 7 | 300–350 W (configurable) 7 |
|---|
| FP64 / FP64 Tensor | 34 / 67 TFLOPS | 26 / 51 TFLOPS |
|---|
| TF32 Tensor (sparse) | 989 TFLOPS | 756 TFLOPS |
|---|
| BF16/FP16 Tensor (sparse) | 1,979 TFLOPS | 1,513 TFLOPS |
|---|
| FP8 Tensor (sparse) | 3,958 TFLOPS | 3,026 TFLOPS |
|---|
| INT8 Tensor (sparse) | 3,958 TOPS | 3,026 TOPS |
|---|
| NVLink | 4th gen, 18 links, 900 GB/s total | 3 bridges, 600 GB/s bidirectional (2-GPU) |
|---|
| Form factor | SXM5 | PCIe Gen5 (also dual-GPU H100 NVL variant) |
|---|
Figures per NVIDIA’s own H100 Tensor Core GPU datasheet 7, which supersedes the rounded, pre-launch figures NVIDIA gave in its original architecture announcement. Sparse figures are with structural sparsity; NVIDIA’s dense figures are roughly half of each sparse number.
5.Where it’s made
SimpleStart here
H100 is made by TSMC, on a custom 4-nanometer-class process NVIDIA calls 4N 3. NVIDIA does not operate its own chip factories.
Deep diveThe technical detail
NVIDIA’s own Hopper architecture material names TSMC and the custom 4N process 3. Neither NVIDIA nor TSMC publishes which individual TSMC fab site produces H100 wafers.
6.Which systems use it
SimpleStart here
NVIDIA’s own system is DGX H100: 8 chips connected by NVLink, rated at 32 petaflops of FP8 AI performance 6. Amazon’s EC2 P5 instances offer up to 8 H100s per instance with 640GB of total HBM3 memory 8.
Deep diveThe technical detail
DGX H100 combines 8× H100 chips for 32 petaflops of FP8 AI performance 6. AWS EC2 P5 instances provide up to 8× H100 per instance, 640GB total HBM3 and 900 GB/s NVSwitch bandwidth 8. OEM partners named by NVIDIA for H100 systems include Dell, HPE, Lenovo, Supermicro, Cisco, Atos, Fujitsu and GIGABYTE 5. NVIDIA also named the Barcelona Supercomputing Center, Los Alamos National Laboratory, the Swiss National Supercomputing Centre, TACC and the University of Tsukuba as adopters 2.
7.Official pricing
SimpleStart here
NVIDIA has not published a standalone H100 unit price, and DGX H100 pricing was not disclosed in the March 2022 announcement 6. Cloud providers publish their own hourly rates on their own pricing pages, which this page does not reproduce because they change over time.
Deep diveThe technical detail
NVIDIA’s DGX H100 release 6 does not state a price. AWS (P5), Microsoft Azure (ND H100 v5) and Google Cloud (A3) each publish official on-demand hourly rates for H100 instances on their own pricing pages; those rates change frequently enough that this page links to AWS’s own instance page 8 rather than quoting a figure that would go stale.
9.What came before, what came next
SimpleStart here
Came before: A100 (Ampere). Came after: H200, a memory upgrade on the same die, announced November 2023; then B200 (Blackwell architecture), announced March 2024.
Deep diveThe technical detail
H100 succeeded A100 (Ampere). NVIDIA’s next Hopper-family announcement was H200 — “NVIDIA Supercharges Hopper, the World’s Leading AI Computing Platform”, November 2023 — which reuses the same GH100 die with upgraded HBM3e memory rather than changing the compute architecture. The next full architectural generation is Blackwell (B200), announced at GTC in March 2024.
10.Hidden in plain sight
SimpleStart here
NVIDIA’s flagship AI-training chip doubles as a security appliance. H100 is officially “the world’s first native Confidential Computing GPU” 3, meaning a cloud tenant’s model weights can be hardware-isolated from the cloud provider running the machine underneath them.
Deep diveThe technical detail
NVIDIA’s own Hopper architecture blog states H100 “implements the world’s first native Confidential Computing GPU,” with trusted-execution-environment support down to individual MIG-instance granularity 3. This is a hardware capability, not a software add-on — and it shipped on the same chip that most people only ever hear about in the context of AI training compute.