这是史上最快GPU!我们测了四张H100!价值120万元!
Independent Chinese-language hands-on video whose page identifies a four-H100 test. Any results belong to its stated setup; the link does not establish a Museum benchmark claim.
极客湾Geekerwan · zh-CN · 2023-06-14NVIDIA H100 family — GH100 die photograph
| Identity & fabrication | |
|---|---|
| Record year | Manufacturer / primary sourceNVIDIA Scope: GPU design Source date: 2022-03-22 Year basis: announced. This is distinct from source publication and record update. |
| Foundry process | Manufacturer / primary sourceNVIDIA Scope: GPU design Source date: 2022-03-22 |
| Process label | Manufacturer / primary sourceNVIDIA Technical Blog Scope: Full GH100 GPU silicon design Source date: 2022-03-22 |
| Packaging, power & physical analysis | |
| Transistors | Manufacturer / primary sourceNVIDIA Scope: GPU design Source date: 2022-03-22 |
| Die area | Manufacturer / primary sourceNVIDIA Scope: One full GH100 silicon die powering H100 Source date: 2022-03-22 Excludes HBM dies, interposer, substrate and SXM/PCIe board. Full silicon die area is not an area proportional to enabled SM count. |
| Transistors | Manufacturer / primary sourceNVIDIA Technical Blog Scope: Full GH100 GPU silicon design Source date: 2022-03-22 |
| Die area | Manufacturer / primary sourceNVIDIA Technical Blog Scope: Full GH100 GPU silicon design; excludes HBM stacks, interposer, substrate and board Source date: 2022-03-22 |
| Additional specifications | |
| Specialized engine | Manufacturer / primary sourceNVIDIA Scope: GPU design Source date: 2022-03-22 |
| Full-design SM count | Manufacturer / primary sourceNVIDIA Technical Blog Scope: Full GH100 GPU design; not an enabled H100 product count Source date: 2022-03-22 |
Die photographs shared by their authors on X. Open a post to explore its images and discussion.
HiEQ · @HiEq_2005 · View the post above ↑TECHNICAL ARTICLE
NVIDIA discloses a full 144-SM GH100 design and separately documents the enabled resources and HBM3 configuration of H100 SXM5.
NVIDIA's Hopper technical article explicitly labels the full GH100 GPU as a TSMC 4N design with 80 billion transistors and an 814 mm² die. Its full organization lists eight GPCs, 72 TPCs and 144 SMs. These are silicon-design facts: the 814 mm² value does not include external HBM stacks, an interposer, substrate or an SXM board, and the 144-SM figure is not an enabled count for every product. NVIDIA then gives a separate H100 SXM5 board-form-factor list: eight GPCs, 66 TPCs, 132 SMs, 80 GB HBM3 and five HBM3 stacks. The same source distinguishes PCIe H100 with 114 SMs, demonstrating why the product label alone cannot supply a universal enabled-SM count or memory capacity. This intake models the full design as GH100 and the SXM5 configuration separately. It preserves their relationship without borrowing SXM5 memory or enabled resources as attributes of the complete die. NVIDIA's launch announcement independently describes H100 as built with 80 billion transistors on TSMC 4N, consistent with the architecture article. Neither source here is used as a measured application-performance result.
Eight H100 GPUs; system score is not inherited by the GH100 die.
NVIDIA · EvidenceThe relationship identifies the disclosed GH100 silicon basis. The SXM5 enabled-SM and HBM3 fields remain SKU-scoped.
NVIDIA Technical Blog · Evidencesilicon design · NVIDIA GH100 full GPU design