NVIDIA B200
The NVIDIA B200 is a high-performance data center GPU based on the Blackwell architecture, designed primarily for large-scale artificial intelligence training and inference workloads. The product employs advanced process technology and supports high-bandwidth memory (HBM), delivering large-scale parallel computing capability suited for compute-intensive applications such as large language models, scientific computing, and deep learning. It is one of NVIDIA's flagship accelerator products for enterprise-grade AI infrastructure.
Timeline (last 90 days)
B200 is an NVIDIA GPU product for large-scale AI workloads; it is a data center GPU of the Blackwell architecture, uses a dual-die package, and is equipped with HBM3e high-bandwidth memory.
Calculated over an eight-week window, if a team occupies one B200 for 10 hours a day, the cost is about $6,160 at the original price and about $5,040 at the new price, a saving of more than $1,100.
Together AI announced that the price of its post-training service based on NVIDIA B200 GPUs has been adjusted to $9 per GPU hour, applicable to the first eight weeks after activation, down from the original price of $11.
When sustained reasoning workloads exceed 40 million output tokens per day, migrating open-weight mixture-of-experts reasoning models to dedicated bare-metal NVIDIA H200/B200 clusters turns the variable per-token penalty into fixed hardware amortization.
Beam's roughly 280GB NVFP4 checkpoint can run on 4 B200s, while the full BF16 version needs about 1TB of VRAM.
The deployment configuration demonstrated by Hugging Face is: the BF16 checkpoint across 4 nodes with 8 H200 GPUs per node; the NVFP4 checkpoint running on a single B200 node.
Running Kimi K2.5 on an NVIDIA B200 with real coding agent trajectories (SWE-smith, EvalScope evaluation), TokenSpeed outperforms TensorRT-LLM across the entire latency/throughput curve; at batch size 1 it is about 9% faster, and at about 100 tokens/second per user its throughput is about 11% higher.
On seven benchmarks covering mathematics, code generation and dialogue, ReaDy using Qwen3.8-27B achieved an average speedup of 3.16x over autoregressive decoding on a single NVIDIA B200 GPU.
Executing a certain LoRA training task on a server CPU requires about 320 hours, while on an NVIDIA B200 GPU it takes only about 1 hour.
The NPU server's training throughput is about 54% of a single NVIDIA B200, while the hardware price is comparable.
3 more timeline events
All Facts (12)
NVIDIA B200 GPU基于Blackwell架构，相比H100在训练性能上提升约2.5倍，推理性能提升可达5倍
85%UnverifiedThe iterative advancement of dedicated inference chips such as the NVIDIA H200/B200 and Google TPU v5 continues to drive down the marginal cost per token.
60%UnverifiedB200 is an NVIDIA GPU product for large-scale AI workloads; it is a data center GPU of the Blackwell architecture, uses a dual-die package, and is equipped with HBM3e high-bandwidth memory.
50%UnverifiedCalculated over an eight-week window, if a team occupies one B200 for 10 hours a day, the cost is about $6,160 at the original price and about $5,040 at the new price, a saving of more than $1,100.
50%UnverifiedWhen sustained reasoning workloads exceed 40 million output tokens per day, migrating open-weight mixture-of-experts reasoning models to dedicated bare-metal NVIDIA H200/B200 clusters turns the variable per-token penalty into fixed hardware amortization.
50%UnverifiedBeam's roughly 280GB NVFP4 checkpoint can run on 4 B200s, while the full BF16 version needs about 1TB of VRAM.
50%UnverifiedThe deployment configuration demonstrated by Hugging Face is: the BF16 checkpoint across 4 nodes with 8 H200 GPUs per node; the NVFP4 checkpoint running on a single B200 node.
50%UnverifiedRunning Kimi K2.5 on an NVIDIA B200 with real coding agent trajectories (SWE-smith, EvalScope evaluation), TokenSpeed outperforms TensorRT-LLM across the entire latency/throughput curve; at batch size 1 it is about 9% faster, and at about 100 tokens/second per user its throughput is about 11% higher.
50%UnverifiedOn seven benchmarks covering mathematics, code generation and dialogue, ReaDy using Qwen3.8-27B achieved an average speedup of 3.16x over autoregressive decoding on a single NVIDIA B200 GPU.
50%UnverifiedThe core upgrade of the H200 over the previous-generation H100 lies in VRAM capacity and bandwidth.
50%UnverifiedA single inference pass of a trillion-parameter-scale AI model may require tens or even hundreds of high-end GPUs working in concert.
50%UnverifiedTogether AI announced that the price of its post-training service based on NVIDIA B200 GPUs has been adjusted to $9 per GPU hour, applicable to the first eight weeks after activation, down from the original price of $11.
50%