NVIDIA B200 rental: early access

NVIDIA B200 specifications

Architecture

Blackwell

Announced

2024

Memory

180 GBUsable memory per NVIDIA. The chip’s maximum memory is 192 GB.

Memory type

HBM3e

Memory bandwidth

7.7 TB/sPer NVIDIA’s Blackwell Technical Brief (HGX B200); NVIDIA’s DGX B200 materials give up to 8 TB/s.

FP32

75 TFLOPS

FP16/BF16, dense

2,250 TFLOPS

FP8, dense

4,500 TFLOPS

FP4, dense

9,000 TFLOPS

FP64

37 TFLOPS

FP64 Tensor Core

37 TFLOPS

RT cores

no

NVENC/NVDEC

NVENC 0, NVDEC 7

MIG

up to 7 instances

GPU interconnect

NVLink 1.8 TB/s

Power (TDP/TGP)

up to 1,000 W

Form factor

SXM (being confirmed)

Memory is specified per GPU. Multiple GPUs have separate memory: using a larger model across them requires compatible software and model partitioning. NVLink does not make them one GPU.

The B200 form factor is still being confirmed. Specs from NVIDIA’s official materials, checked 26 September 2026.

The B200 by workload

• Inference: LLMs up to 123B in FP8 on one GPU
• Fine-tuning: QLoRA up to 235B, LoRA up to 32B on one GPU
• Training from scratch: models up to 8B on one GPU
• Rendering: no RT cores
• FP64 computing: 37 TFLOPS

An estimate from open formulas, not a benchmark or an invoice.

Servers with the NVIDIA B200

Server configuration (CPU, RAM, disk, network) is being confirmed.

Sizing assumption: up to 8 GPUs per server. We’ll confirm the actual configuration in our reply.

Sources

• NVIDIA Blackwell Technical Brief (opens in a new tab)
• NVIDIA HGX Platform (opens in a new tab)
• NVIDIA DGX B200 (opens in a new tab)
• NVIDIA Multi-Instance GPU User Guide (opens in a new tab)
• NVIDIA Video Encode and Decode Support Matrix (opens in a new tab)
• NVIDIA Omniverse Developer Guide: Technical Requirements (opens in a new tab)
• NVIDIA Newsroom, March 18, 2024 (opens in a new tab)

Similar GPUs

• NVIDIA H200

  141 GB HBM3e · 4.8 TB/s · NVLink

  B200 leads on 6 of 7 metrics, H200 on 1. B200 has more FP16/BF16 (2,250 vs 989.4 TFLOPS) and FP8 (4,500 vs 1,978.9 TFLOPS). But H200 has more FP64 Tensor Core throughput (67 vs 37 TFLOPS). Without Tensor Cores, B200 has more FP64: 37 vs 34 TFLOPS.

• NVIDIA B300

  270 GB HBM3e · 7.7 TB/s · NVLink

  B300 leads on 2 of 8 metrics, B200 on 1. B300 has more FP4 (14,000 vs 9,000 TFLOPS) and memory (270 vs 180 GB). But B200 has more FP64 (37 vs 1.2 TFLOPS). Equal: memory bandwidth, FP32, FP16/BF16, FP8, NVLink.

• NVIDIA RTX PRO 6000 Blackwell

  96 GB GDDR7 · 1.6 TB/s · PCIe 5

  B200 leads on 5 of 6 metrics, RTX PRO 6000 on 1. B200 has more memory bandwidth (7.7 vs 1.6 TB/⁠s) and FP16/BF16 (2,250 vs 467.8 TFLOPS). But RTX PRO 6000 has more FP32 (120 vs 75 TFLOPS). FP64: 37 TFLOPS on B200; not in NVIDIA’s card specs for RTX PRO 6000 (NVIDIA’s whitepaper gives 1/64 of FP32 for the GB202 chip).

Dense FP16/BF16 and FP8 from the H100 SXM5 column of NVIDIA’s whitepaper: for the H200, NVIDIA lists only with-sparsity figures, the same as the H100 SXM’s.

The RTX PRO 6000’s FP16, FP8 and FP4 are derived figures. Derived: NVIDIA’s whitepaper per-SM Tensor Core rate × the 2,430 MHz boost clock in the product brief. For the Server Edition, NVIDIA gives FP16/FP8/FP4 only as rounded 1/2/4 PFLOPS peaks, the level of sparse figures; it publishes no dense ones.