NVIDIA Blackwell B200 · key spec hierarchy

B200 specs by node, GPU, and die

This page summarizes the common 8× B200 HGX/DGX node, one B200 GPU, and one physical Blackwell die. Values are separated into directly sourced specs and derived per-die estimates to avoid double-counting the dual-die package.

Most important correction: a B200 GPU is dual-die, but the CUDA-visible GPU has 148 SMs total, not 148 SMs per die.

Per die ≈ 74 active SMs is derived from 148 ÷ 2, not an independent CUDA device.

8-GPU node
8× B200
Typical DGX/HGX B200 server/baseboard level
Blackwell dies
16
Active SMs
1,184
GPU memory
1,440 GB
NVLink aggregate
14.4 TB/s
1 GPU package
2 dies
Unified by NV-HBI; appears as one coherent GPU to software
Active SMs
148
Transistors
208B
Memory
180 GB*
HBM BW
7.7 TB/s
1 physical die
≈74 SMs
Derived split; not separately exposed as a CUDA GPU
Transistors
≈104B
HBM stacks
≈4
Memory
≈90 GB*
HBM BW
≈3.85 TB/s

Hierarchy: do not double count the two dies

Node
DGX/HGX B200

8 CUDA-visible B200 GPUs. NVIDIA DGX B200 lists 1,440 GB total GPU memory and 14.4 TB/s aggregate NVLink bandwidth.

GPU
1 B200

One B200 package contains two physical Blackwell dies but is presented as one coherent GPU device.

Die
½ B200

Useful for architecture reasoning only. CUDA kernels see the whole B200 device, not an individual die.

Node formulas: 8 GPUs × 2 dies/GPU = 16 Blackwell dies 8 GPUs × 148 SM/GPU = 1,184 active SMs 8 GPUs × 180 GB/GPU = 1,440 GB GPU memory 8 GPUs × 7.7 TB/s/GPU = 61.6 TB/s HBM bandwidth ≈ 62 TB/s

Compute peak: dense vs sparse, per node / per GPU / per die

Precision / path8× B200 node1× B200 GPU≈1 die, derivedMeaning
FP4 Tensor Core72 dense / 144 sparse PFLOPS9 dense / 18 sparse PFLOPS≈4.5 dense / 9 sparse PFLOPSAI inference-oriented low precision
FP8 / FP6 Tensor Core36 dense / 72 sparse PFLOPS4.5 dense / 9 sparse PFLOPS≈2.25 dense / 4.5 sparse PFLOPSTraining/inference Tensor Core formats
INT8 Tensor Core36 dense / 72 sparse POPS4.5 dense / 9 sparse POPS≈2.25 dense / 4.5 sparse POPSInteger Tensor Core path
FP16 / BF16 Tensor Core18 dense / 36 sparse PFLOPS2.25 dense / 4.5 sparse PFLOPS≈1.125 dense / 2.25 sparse PFLOPSCommon DL training/inference path
TF32 Tensor Core8.8 dense / 17.6 sparse PFLOPS1.1 dense / 2.2 sparse PFLOPS≈0.55 dense / 1.1 sparse PFLOPSFast Tensor Core path for FP32 tensors
FP32 CUDA-core/SIMT600 TFLOPS75 TFLOPS≈37.5 TFLOPSTrue FP32 path; not TF32 Tensor Core
FP64 / FP64 Tensor Core296 TFLOPS37 TFLOPS≈18.5 TFLOPSHPC-oriented double precision path
Dense = normal dense math Sparse = NVIDIA 2:4 structured sparsity peak where applicable Per-die compute is derived by ÷2

Memory and interconnect

Metric8× node1× GPU≈1 die
GPU memory1,440 GB total180 GB HBM3e*≈90 GB*
HBM bandwidth61.6 TB/s derived from per-GPU; NVIDIA pages also list ~62–64 TB/s7.7 TB/s≈3.85 TB/s
HBM stacks64 total, derived8 total≈4
GPU-to-GPU NVLink14.4 TB/s aggregate1.8 TB/s per GPUNot die-visible
Die-to-die linkInside each GPU packageNV-HBI coherent single-device link10 TB/s chip-to-chip link

*NVIDIA product datasheet for HGX B200 lists 180 GB HBM3e per GPU; the architecture paper describes a unified 192 GB HBM3e memory space. Treat 180 GB as the HGX/DGX product capacity exposed in current product specs.

Architecture facts

Metric1× B200 GPU≈1 die
Physical dies21
Transistors208 billion≈104 billion
Active SMs148 total≈74 if evenly split
GPCs8 total≈4 if evenly split
L2 partitions4 total≈2 if evenly split
MIG7 instances supportedMIG applies at GPU level
TDPConfigurable up to 1,000 W for HGX B200 GPUNot specified per die

Runtime rule for CUDA / PyTorch / vLLM

Use 148 SMs per B200 CUDA device. Do not multiply by two again for the two dies. If you have an 8-GPU node, the total active SM count is 8 × 148 = 1,184.

Performance rule for benchmarking

FP32 75 TFLOPS is the true FP32/SIMT product spec. TF32 1.1 / 2.2 PFLOPS is Tensor Core dense/sparse. A PyTorch float32 GEMM may use TF32 unless disabled.

What is directly sourced vs derived?

ValueStatusReason
8 B200 GPUs, 1,440 GB total memory, 14.4 TB/s NVLink aggregateDirect NVIDIA DGX B200 specNVIDIA DGX B200 page and user guide list these node-level values.
180 GB HBM3e, 7.7 TB/s, 75 TFLOPS FP32, 2.2 PFLOPS TF32 sparse, 18 PFLOPS FP4 sparseDirect NVIDIA Blackwell/HGX B200 datasheetThese are listed under individual HGX B200 GPU specifications.
2 dies, 208B transistors, 148 SMs, 8 GPCs, 8 HBM3e stacks, NV-HBI unified deviceDirect from the B200 microbenchmarking paperThe paper explicitly describes one B200 GPU this way.
≈74 SMs/die, ≈90 GB/die, ≈3.85 TB/s/die, ≈37.5 TFLOPS FP32/dieDerivedCalculated by dividing the GPU-level value by two. Useful for mental model, but not exposed as separate GPU devices.