192GB HBM3e with Blackwell architecture FP4 Tensor Cores. 1800 GB/s NVLink for massive multi-GPU scaling. Significant inference throughput improvement over H100 with FP4 precision.

When is the B100 not a good fit?

New generation with limited provider availability. High cost per hour. For inference on models under 70B or workloads that don't benefit from FP4, the H100 or L40S is more cost-effective.

What size AI models can the B100 run?

With 192GB of VRAM, the B100 can usually run 70B-class models with headroom, and may handle much larger models in 4-bit quantized form depending on runtime overhead, KV cache, and context length.

How much VRAM does the B100 have?

The B100 has 192GB of VRAM. Multi-GPU setups increase total memory, but that memory is not automatically pooled across GPUs.

What is the B100's memory bandwidth?

The B100 has 8,000 GB/s of memory bandwidth. Higher bandwidth helps with faster data transfer between GPU memory and compute cores.

What data types does the B100 support?

The B100 supports 9 precision formats. Training: BF16, FP16, TF32, FP32. Inference: FP4, FP6, FP8, INT8. Scientific: FP64.

Does the B100 support NVLink?

Yes. The B100 supports NVLink with 1800 GB/s of bidirectional bandwidth. This helps accelerate multi-GPU communication.

Nvidia B100

Blackwell data center GPU for next-generation AI training and inference.

Compare vs other GPUs →

Key Specifications

Architecture

Blackwell

Memory

Up to 192GB HBM3e

Memory Bandwidth

8,000 GB/s

Release date

Q4 2024

We couldn't find any available B100 GPUs. Search for alternative GPUs.

Frequently Asked Questions

Why choose the B100?: 192GB HBM3e with Blackwell architecture FP4 Tensor Cores. 1800 GB/s NVLink for massive multi-GPU scaling. Significant inference throughput improvement over H100 with FP4 precision.
When is the B100 not a good fit?: New generation with limited provider availability. High cost per hour. For inference on models under 70B or workloads that don't benefit from FP4, the H100 or L40S is more cost-effective.
What size AI models can the B100 run?: With 192GB of VRAM, the B100 can usually run 70B-class models with headroom, and may handle much larger models in 4-bit quantized form depending on runtime overhead, KV cache, and context length.
How much VRAM does the B100 have?: The B100 has 192GB of VRAM. Multi-GPU setups increase total memory, but that memory is not automatically pooled across GPUs.
What is the B100's memory bandwidth?: The B100 has 8,000 GB/s of memory bandwidth. Higher bandwidth helps with faster data transfer between GPU memory and compute cores.
What data types does the B100 support?: The B100 supports 9 precision formats. Training: BF16, FP16, TF32, FP32. Inference: FP4, FP6, FP8, INT8. Scientific: FP64.
Does the B100 support NVLink?: Yes. The B100 supports NVLink with 1800 GB/s of bidirectional bandwidth. This helps accelerate multi-GPU communication.

Technical Specifications


Form Factor	8x NVIDIA B100 SXM
FP4 Tensor Core¹	112 PFLOPS
FP8/FP6 Tensor Core¹	56 PFLOPS
INT8 Tensor Core¹	56 POPS
FP16/BF16 Tensor Core¹	28 PFLOPS
TF32 Tensor Core¹	14 PFLOPS
FP32	480 TFLOPS
FP64	240 TFLOPS
FP64 Tensor Core	240 TFLOPS
Memory	Up to 1.5TB
NVLink	Fifth generation
NVIDIA NVSwitch™	Fourth generation
NVSwitch GPU-to-GPU Bandwidth	1.8TB/s
Total Aggregate Bandwidth	14.4TB/s