CARDINALNVIDIA H100 specifications ↗
SXM and NVL specifications; memory is per GPU
HTML · Source reviewed 2026-09-08 · Opens at the manufacturer
PRIMARY SOURCE LIBRARY
GO STRAIGHT
TO THE SOURCE
Manufacturer documents used in our exact-variant references. Check the product, revision and footnotes before applying a specification to an offer.
36 source records · Files open at the original publisher. Source availability and revisions can change.
NVIDIA
NVIDIA Hopper architecture in depth ↗
Confirms H100 SXM5 HBM3 and PCIe HBM2e
HTML · Source reviewed 2026-09-08 · Opens at the manufacturer
NVIDIA H100 PCIe product brief ↗
Document identifies itself PB-11133-001_v02. Table 6 gives 600 GB/s NVLink; introduction says 900 GB/s. Table 6 used
PDF · Source reviewed 2026-09-08 · Opens at the manufacturer
NVIDIA H100 NVL product brief ↗
Per-card identity and bridge topology
PDF · Source reviewed 2026-09-08 · Opens at the manufacturer
NVIDIA H100 datasheet ↗
Resolved from NVIDIA resource portal
PDF · Source reviewed 2026-09-08 · Opens at the manufacturer
NVIDIA H200 specifications ↗
Manufacturer labels specification table preliminary
HTML · Source reviewed 2026-09-08 · Opens at the manufacturer
NVIDIA H200 datasheet ↗
Resolved from manufacturer product page
PDF · Source reviewed 2026-09-08 · Opens at the manufacturer
NVIDIA L40S specifications ↗
Confirms no NVLink or MIG
HTML · Source reviewed 2026-09-08 · Opens at the manufacturer
NVIDIA A100 specifications ↗
80GB PCIe and SXM; CTS power footnote
HTML · Source reviewed 2026-09-08 · Opens at the manufacturer
NVIDIA A100 datasheet ↗
Current linked datasheet
PDF · Source reviewed 2026-09-08 · Opens at the manufacturer
NVIDIA A100 40GB PCIe product brief ↗
PB-10137-001_v03; memory table and board power
PDF · Source reviewed 2026-09-08 · Opens at the manufacturer
NVIDIA Ampere architecture whitepaper ↗
Pages 35-37 document original A100 40GB SXM4
PDF · Source reviewed 2026-09-08 · Opens at the manufacturer
NVIDIA A40 datasheet ↗
Manufacturer specifications; historical publication does not establish current software qualification
PDF · Source reviewed 2026-09-08 · Opens at the manufacturer
NVIDIA A10 product brief ↗
Manufacturer specifications; historical publication does not establish current software qualification
PDF · Source reviewed 2026-09-08 · Opens at the manufacturer
NVIDIA A30 product brief ↗
Manufacturer specifications; historical publication does not establish current software qualification
PDF · Source reviewed 2026-09-08 · Opens at the manufacturer
NVIDIA T4 product brief ↗
Manufacturer specifications; historical publication does not establish current software qualification
PDF · Source reviewed 2026-09-08 · Opens at the manufacturer
NVIDIA V100 family datasheet ↗
January 2020 table separates V100 PCIe, V100 SXM2 and V100S; no V100S specifications transferred to V100
PDF · Source reviewed 2026-09-08 · Opens at the manufacturer
NVIDIA CUDA architecture support guidance ↗
Pre-Turing offline compilation removal in CUDA 13; legacy deployment qualification needed
HTML · Source reviewed 2026-09-08 · Opens at the manufacturer
NVIDIA L40 specifications ↗
Corroborates memory, bandwidth, 300W rating, passive cooling and dimensions
HTML · Source reviewed 2026-09-08 · Opens at the manufacturer
NVIDIA L40 product brief ↗
January 2023 v03 marked preliminary; physical card figure on printed page 1, dimension drawing printed page 11
PDF · Source reviewed 2026-09-08 · Opens at the manufacturer
NVIDIA HGX reference architecture: GPU and node specifications ↗
Table 1 explicitly separates per-GPU and per-node memory/bandwidth; unrelated CPU table cells contain apparent editorial errors and are not used
HTML · Source reviewed 2026-09-08 · Opens at the manufacturer
NVIDIA HGX B200 PCF summary specifications ↗
Page 1 explicitly states individual GPU maximum configurable power and SXM6; platform performance and weight are not per-GPU fields
PDF · Source reviewed 2026-09-08 · Opens at the manufacturer
NVIDIA DGX B300 power-capping guidance ↗
An example power-limit response is not treated as a universal B300 TDP specification
HTML · Source reviewed 2026-09-08 · Opens at the manufacturer
NVIDIA A100 four-variant datasheet, June 2021 ↗
Page 1 explicit shared compute rows for PCIe/SXM 40GB/80GB with separately printed dense and sparse values
PDF · Source reviewed 2026-09-08 · Opens at the manufacturer
NVIDIA H100 PCIe theoretical arithmetic rates in HPC system article ↗
September 28 2023 body explicitly states PCIe vector FP32/FP64 and tensor FP64 peaks; benchmark results and ambiguously qualified BF16 headline are not reused
HTML · Source reviewed 2026-09-08 · Opens at the manufacturer
AMD
AMD Instinct MI300X specifications ↗
Infinity Fabric bandwidth is per link; eight links listed
HTML · Source reviewed 2026-09-08 · Opens at the manufacturer
AMD Instinct MI300X datasheet ↗
Manufacturer linked two-page datasheet
PDF · Source reviewed 2026-09-08 · Opens at the manufacturer
AMD Instinct MI325X specifications ↗
Peak board power, not measured system consumption
HTML · Source reviewed 2026-09-08 · Opens at the manufacturer
AMD Instinct MI325X datasheet ↗
Manufacturer linked two-page datasheet
PDF · Source reviewed 2026-09-08 · Opens at the manufacturer
AMD Instinct MI210 product brochure ↗
Reference geometry on page 1; three links are per-card capabilities, not a complete-server topology
PDF · Source reviewed 2026-09-08 · Opens at the manufacturer
AMD Instinct MI350X specifications ↗
Per-accelerator memory, interface and typical board power
HTML · Source reviewed 2026-09-08 · Opens at the manufacturer
AMD Instinct MI355X specifications ↗
Per-accelerator memory, interface and typical board power
HTML · Source reviewed 2026-09-08 · Opens at the manufacturer
AMD CDNA 4 architecture white paper ↗
Printed page 15 distinguishes air-cooled MI350X and direct-liquid-cooled MI355X platform configurations
PDF · Source reviewed 2026-09-08 · Opens at the manufacturer
AMD Instinct MI350X GPU brochure ↗
Page 1 explicitly labels maximum TBP and specifies seven scale-up links at 153.6 GB/s each. Product webpage uses typical board power terminology.
PDF · Source reviewed 2026-09-08 · Opens at the manufacturer
AMD Instinct MI355X GPU brochure ↗
Page 1 explicitly labels maximum TBP and specifies seven scale-up links at 153.6 GB/s each. Product webpage uses typical board power terminology.
PDF · Source reviewed 2026-09-08 · Opens at the manufacturer
INTEL
Intel Gaudi 3 accelerator product overview ↗
Pages 9, 18 and 22 distinguish OAM from HL-338 PCIe
PDF · Source reviewed 2026-09-08 · Opens at the manufacturer
MAKE THE SPECIFICATION USEFUL
Learn how memory, bandwidth and platform requirements affect the hardware you should request.
Read the specification guide ↗