NVIDIA · Hopper architecture ↗
March 22, 2022 · Figure 1 and H100 SXM5 configuration
One GH100 compute die, five enabled HBM3 stacks; schematic layers and connection paths.
HOPPER · 80 GB HBM3
NVIDIA H100 SXM
NVIDIA’s Hopper data center GPU from 2022: one large die with 80 GB of HBM3 memory, the GPU inside clusters such as Colossus.
More about H100
The H100 is NVIDIA’s ninth-generation data center GPU, built on the Hopper architecture. One GH100 die sits beside five HBM3 stacks holding 80 GB, on an SXM module that plugs into an HGX board.
When
NVIDIA introduced the H100 at its GTC conference in March 2022. Blackwell, which followed, splits the work across two dies.
Why it matters
It added a Transformer Engine that switches between FP8 and 16-bit math for large language models. Comparing it with Blackwell Ultra and MI300X shows how GPUs grew from one die to several.
Key facts
- Compute die
- 1 GH100 (80 billion transistors)
- Memory
- 80 GB HBM3, five stacks, over 3 TB/s
- Made on
- TSMC 4N
- Introduced
- March 2022
NVIDIA · Hopper architecture in depth, March 2022NVIDIA · H100 specificationsSources reviewed October 28, 2026
TEACHING MODELDrag to rotate
Opening the 3D model…
Follow heat or data
CONNECT THE SCALES
What moves through the system?
Sources & model notes · reviewed October 28, 2026
Photo-informed SXM module with six physical memory sites and five enabled stacks. The muted inactive site’s location is illustrative; it adds no capacity. Layer heights are exaggerated.
NVIDIA · H100 SXM specifications ↗
Retrieved October 28, 2026 · Product specifications, H100 SXM column
80 GB HBM3. SXM, not H100 NVL or H200. Product rating, not measured Colossus performance.
Original authored geometry. No private CAD, exact facility floor plan, performance simulation, or current deployment total is implied. Cooling and networking lines show relationships, not surveyed routes.