PantheonGet Early Access

Hardware

Compare GPUs

Pick 2–4 data-center accelerators and compare them side by side — memory, bandwidth, dense FP16 / FP8 / FP4 compute, board power, perf-per-watt, and interconnect. Figures are drawn from published vendor datasheets; every column links through to matching nodes you can request a quote on. No pricing — compute is quoted per configuration on request.

SpecNVIDIANVIDIA H100 SXMNVIDIANVIDIA H200 SXMNVIDIANVIDIA B200 (HGX)
ArchitectureHopperHopperBlackwell
GPU memory80 GB141 GB180 GB
Memory typeHBM3HBM3eHBM3e
Memory bandwidth3.35 TB/s4.8 TB/s7.7 TB/s
FP16 dense989.5 TFLOPS989.5 TFLOPS2,250 TFLOPS
FP8 dense1,979 TFLOPS1,979 TFLOPS4,500 TFLOPS
FP4 dense9,000 TFLOPS
Board power (TDP)700 W700 W1000 W
Perf / watt (FP16)1.41 TFLOPS/W1.41 TFLOPS/W2.25 TFLOPS/W
InterconnectNVLink 4 · 900 GB/sNVLink 4 · 900 GB/sNVLink 5 · 1800 GB/s
Scale-up domain8 GPUs8 GPUs8 GPUs
Form factorSXM5SXM5SXM (HGX B200)
Power & cooling~5.6 kW / 8-GPU node
Liquid-cooled
Air-cooled
~5.6 kW / 8-GPU node
Air-cooled
~8.0 kW / 8-GPU node
Air-cooled
Liquid-cooled
AvailabilityView NVIDIA H100 SXM specsDatasheet ↗View NVIDIA H200 SXM specsDatasheet ↗View NVIDIA B200 (HGX) specsDatasheet ↗

Specs as of · sourced from published vendor datasheets. FLOPS are dense (no sparsity); “—” marks a precision the chip does not support. Power figures are approximate, accelerators only (excludes host CPUs, networking, and cooling overhead). Pricing on request — confirm the exact configuration when you request a quote.