GPUDatacenterReleased —

NVIDIA B100

192 GB HBM3e · 8192-bit

Announced alongside the B200 at GTC 2024 as the 700 W, air-cooled member of the Blackwell family, intended as a drop-in upgrade for existing HGX H100 servers. Like the B200 it joins two reticle-limited dies into a single GPU over a 10 TB/s interconnect. Industry analysts reported later in 2024 that NVIDIA had effectively shelved it in favour of the B200 and B200A, and it does not appear in NVIDIA's current data center line-up.

General

Manufacturer: NVIDIA
Model: B100
Segment: Datacenter
GPU chip: Blackwell (2x GB100)
Architecture: Blackwell
Release date: —
Price: Search on Amazon

Memory

Memory size: 192 GB
Memory type: HBM3e
Bus width: 8192-bit
Bandwidth: 8,000 GB/s
Memory clock: —

Clocks & silicon

GPU clock: —
Boost clock: —
Shader units: —
Tensor / AI cores: —
RT cores: —
Transistors: 208 billion
Technology (nm): 4 nm
TSMC 4NP
TDP / TBP: 700 W

Interface & APIs

Bus interface: SXM
SLI / multi-GPU: —
DirectX: N/A (compute-only)
OpenGL: N/A (compute-only)
Vulkan: N/A (compute-only)
Other APIs: CUDA (compute capability 10.0)

AI support

Precisions: FP8FP16BF16INT8INT4TF32FP64
AI accelerator: Second-generation Transformer Engine with FP4 Tensor Cores
Peak AI TOPS: 14,000 TOPS
Notes: Second-generation Transformer Engine with FP4 and FP6 support. At its GTC 2024 announcement NVIDIA quoted HGX B100 throughput per GPU of FP4 14 PFLOPS with sparsity (7 dense), FP8 and FP6 7 PFLOPS, INT8 7 POPS, FP16 and BFLOAT16 3.5 PFLOPS and TF32 1.8 PFLOPS with sparsity, dense rates half of those, plus 30 TFLOPS of FP64. NVIDIA does not claim INT4 Tensor Core support on Blackwell.