GB200

The 'G' in GB200 stands for Grace CPU, and the 'B' stands for Blackwell GPU.

In the gpt-oss 120B test, Jalapeño achieved a peak throughput of approximately 85,448 per watt compared to GB200's 44,960, roughly 1.9×; end-to-end latency was reduced from 1.80 seconds to 1.03 seconds; and minimum TBT dropped from 1.87 ms to 0.69 ms.

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.

Megatron-Bridge receives the model's complete recipes for pretraining, SFT, and PEFT, targeting servers equipped with H100 and GB200.

GB200 is NVIDIA's next-generation GPU platform for large-scale AI inference and training, featuring higher memory bandwidth and denser compute units.

JetSpec++ reaches a generation speed of up to 1,361 tokens/second for a single request on NVIDIA GB200.

According to CoreWeave's comparison chart, Rubin holds approximately a 3x advantage over the GB200 at the slow end, climbing to more than 7x at the fast end

The attention forward pass of sigmoid attention improves by 7.4%, but the full GPU step improves by only 0.3%.

On the NVIDIA GB200 platform, full training step throughput improves by 2.7% for dense SiLU tasks and by 2.9% for tanh soft-capping attention.

On the NVIDIA GB200 platform, full training step throughput improves by 8.0% for routed-expert SwiGLU tasks, the highest gain among the four tasks.

Jalapeño has a rated power consumption of 700W with measured sustained power consumption not exceeding 550W; by comparison, the GB200 is rated at 1200W and the GB300 at 1400W.

The GB300 is the latest iteration of NVIDIA's Grace Blackwell architecture, with the previous generation being the GB200.

NVIDIA has launched DGX systems and rack-level solutions based on the GB200, selling complete AI computing infrastructure rather than individual chips

Compared to the H100, the GB200 achieves a 30× improvement in energy efficiency and a 25× reduction in latency for AI inference tasks.

The Blackwell architecture's B200/GB200 contains 208 billion transistors, uses a dual-chip package design, and achieves an inter-chip interconnect bandwidth of 10 TB/s.

The Blackwell architecture released in 2024 (represented by the B200/GB200) has a transistor count of 208 billion, uses a dual-chip packaging design, and is merged into a single logical GPU via 10 TB/s chip-to-chip interconnect.

Nvidia's Blackwell architecture released in 2024 (with representative products B200/GB200) features 208 billion transistors and adopts second-generation Transformer Engine and NVLink 5.0 interconnect technology.