H100
NVIDIA HGX H100 SXM, 8 GPU
Price on request
An HGX baseboard with eight NVIDIA H100 SXM GPUs, 80GB each. Per-GPU figures below. Available on request, tested before sale.
- GPUs
- 8 x H100 SXM
- GPU memory
- 80GB per GPU
- Memory bandwidth
- 3.35TB/s per GPU
- FP64
- 34 TFLOPS per GPU
- FP64 Tensor Core
- 67 TFLOPS per GPU
- FP32
- 67 TFLOPS per GPU
- TF32 Tensor Core
- 989 TFLOPS* per GPU
- BF16 / FP16 Tensor Core
- 1,979 TFLOPS* per GPU
- FP8 Tensor Core
- 3,958 TFLOPS* per GPU
- INT8 Tensor Core
- 3,958 TOPS* per GPU
- Max thermal design power
- Up to 700W per GPU, configurable
- Multi-Instance GPU
- Up to 7 MIGs at 10GB each
- Interconnect
- NVLink 900GB/s, PCIe Gen5 128GB/s
Source: NVIDIA H100 Tensor Core GPU product page. Tensor figures marked * are with sparsity.
Condition
Documented on the Known Good certificate when this unit is tested, with photos of the actual unit.
Provenance
Disclosed in general terms on the certificate. Supplier names are not published.
Warranty
- OEM warranty
- Looked up by serial at test
- Checked on
- At test
- Dead on arrival
- Per the Terms of Sale
Known Good certificate
This unit is tested to the published protocol before it ships. The certificate, with its verification link, is attached to this listing once testing is complete. See the sample certificate.
Serials
Serial numbers are disclosed once your deposit lands, and the unit is held for you from then on. The certificate shows only the last four characters.
Reserve this unit
Send the form and I reply within one business day with the deposit amount, the serials process and shipping options. Or write to sales@mulliganvalleycompute.com.
Terms in brief
- A deposit reserves the unit. The balance is due before shipping.
- Payment by bank wire or escrow.
- Ships insured from Milpitas, California, or pick up in person.
- Dead on arrival claims within the window in the Terms of Sale.
- Export controls apply. Some destinations and end uses cannot be served.