NVIDIASXMInquiry only
NVIDIA H100 SXM
Image usage rights confirmedRights holder: RTCompute
Detailed description
SXM vs PCIe: SXM — 900 GB/s NVLink and NVSwitch scaling up to 256 GPUs for distributed training; PCIe — lower interconnect limits, 350 W TDP. Memory: SXM — HBM3 at 3.35 TB/s; PCIe — HBM2e at ~2.0 TB/s. Lifecycle: NVIDIA phased out original DGX H100 end-of-sale orders in early 2025, transitioning the ecosystem toward H200.
NVIDIA H100 SXM (SXM5) is a high-performance data center and AI accelerator built on the NVIDIA Hopper Architecture, designed for massive-scale foundation model training and cluster workloads.
Key specifications
ArchitectureNVIDIA Hopper (GH100, 5 nm)
Form factorSXM5 (HGX/DGX, 8 GPUs per node)
Memory80 GB HBM3
Memory bandwidthup to 3.35 TB/s
NVLink900 GB/s (4th gen)
TDPup to 700 W
CUDA / Tensor cores16 896 / 528 (4th gen)
Packing list
Dimensions
320 × 220 × 90 mm
Net weight
1.4 kg
Gross weight
2.1 kg
Typical applications
AI trainingHPC