B200
NVIDIA's flagship data center GPU, based on the Blackwell architecture and intended for large-scale AI training and inference. It represents the highest performance tier currently offered.
Timeline (last 90 days)
The B200 belongs to the Blackwell architecture and uses a dual-die package.
As demand for large-model training and inference continues to swell while professional-grade accelerator cards (such as the H100 and B200 series) have limited production capacity and high prices, buyers turn their attention to consumer-grade flagship cards.
Products such as the H100 and B200 typically have unit prices in the tens of thousands of dollars, and due to production capacity and export controls, lead times are long and channel markups are severe.
SM100 corresponds to GPUs with compute capability 10.0 in NVIDIA's Blackwell architecture, such as B200 and GB200; it is an architecture identifier rather than a specific graphics card model.
Official data shows that on a single B200, LFM2.5-1.2B's p95 latency on RIPL is 14.4 to 28.4 times lower than Qwen3.5-0.8B's on vLLM.
DeepInfra's kernels are written for NVIDIA data center Blackwell chips (such as the B200) and cannot be used out of the box on devices using different Blackwell chips, such as the DGX Spark.
A single cluster processes 155 trillion Tokens per day, with average revenue of $38 per card per hour.
The A100 wastes more than 97% of HBM bandwidth during 4K-token prefill; HBM accounts for 18% of the A100's manufacturing cost, and as much as 45% on the B200.
The TTT-Discover recipe cited by Reddit discussants uses Qwen3-8B with thinking mode enabled, running on two B200s.
The capacity of the same advanced process node must be allocated among data center GPUs (such as the H100/B200 series), consumer flagship and mid-range models.
32 more timeline events
All Facts (20)
英伟达最新的B200芯片单卡热设计功耗达1000瓦
90%Unverified装满B200芯片的机架功率可达120千瓦，是传统服务器机架的十倍以上
80%UnverifiedImage models can maintain INT8 or FP8 precision when offloaded, and Z-Image generation time on a B200 with a 10 GiB VRAM limit dropped from 38.4 seconds to 2.3 seconds
70%UnverifiedOn the B200, the per-step time for LoRA SFT training of Qwen3.5-9B decreased from 0.84 seconds to 0.77 seconds.
70%UnverifiedA single cluster processes 155 trillion Tokens per day, with average revenue of $38 per card per hour.
60%UnverifiedNVIDIA's A100, H100, and the latest B200 GPUs have become the de facto standard for AI training and inference.
60%UnverifiedProducts such as the H100 and B200 typically have unit prices in the tens of thousands of dollars, and due to production capacity and export controls, lead times are long and channel markups are severe.
50%UnverifiedThe B200 belongs to the Blackwell architecture and uses a dual-die package.
50%UnverifiedAs demand for large-model training and inference continues to swell while professional-grade accelerator cards (such as the H100 and B200 series) have limited production capacity and high prices, buyers turn their attention to consumer-grade flagship cards.
50%UnverifiedSM100 corresponds to GPUs with compute capability 10.0 in NVIDIA's Blackwell architecture, such as B200 and GB200; it is an architecture identifier rather than a specific graphics card model.
50%UnverifiedOfficial data shows that on a single B200, LFM2.5-1.2B's p95 latency on RIPL is 14.4 to 28.4 times lower than Qwen3.5-0.8B's on vLLM.
50%UnverifiedDeepInfra's kernels are written for NVIDIA data center Blackwell chips (such as the B200) and cannot be used out of the box on devices using different Blackwell chips, such as the DGX Spark.
50%UnverifiedThe A100 wastes more than 97% of HBM bandwidth during 4K-token prefill; HBM accounts for 18% of the A100's manufacturing cost, and as much as 45% on the B200.
50%UnverifiedThe TTT-Discover recipe cited by Reddit discussants uses Qwen3-8B with thinking mode enabled, running on two B200s.
50%UnverifiedThe capacity of the same advanced process node must be allocated among data center GPUs (such as the H100/B200 series), consumer flagship and mid-range models.
50%UnverifiedThe high-end GPUs required for AI training, such as NVIDIA H100 and B200, all depend on TSMC's 5-nanometer, 3-nanometer, or even more advanced nodes for production.
50%UnverifiedOn 8×B200, DiffusionGemma 26B MoE + ORL reaches 7,708 tokens/s at NVFP4 precision (single-request decoding).
50%UnverifiedRadixArk/Qwen3.8-Flash-Next-NVFP4 mentioned in the LMSYS blog is an SGLang checkpoint that uses post-training quantization, quantizing only the routed experts, calibrated with 128 CNN/DailyMail articles truncated to 512 tokens, validated on GB300 and B300, and reported to reach 540 tok/s at batch size 1 with B200 TP4 plus MTP.
50%UnverifiedIn a Kimi Linear, 8-card B200, concurrency-8 test, cached-prefix throughput improved by approximately 10.9x.
50%UnverifiedKimi-K2.7-Code can be trained on 6 B200 GPUs with packed INT4 experts, whereas using 16-bit copies would require approximately 2 TB of VRAM.
50%