B300
An NVIDIA server-grade AI accelerator based on the Blackwell architecture. Compared with the H100/H800, it offers higher video memory capacity, memory bandwidth, and FP8 compute throughput, making it suitable for large-scale MoE model inference scenarios.
Timeline (last 90 days)
IDC did not specify which Nvidia products were counted in the 34% share, nor whether the data includes restricted chips such as the B200 and B300 that reportedly entered China through smuggling channels.
The unified Radix Cache added an Agentic-Aware tail-optimized LRU eviction strategy, which reduced p99 ITL by 43.9% at concurrency 32 in an agentic trajectory test on DeepSeek-V4-Pro FP4 with 4-card B300.
The B300 is an NVIDIA Blackwell architecture server-grade accelerator card that offers significant improvements in memory capacity, memory bandwidth, and FP8 compute throughput compared to the previous generation H100/H800.
Kernel and scheduling strategies tuned for the B300 may not yield the same magnitude of gains on other hardware.
The 2.2–2.8× improvement was achieved on specific B300 hardware under specific benchmark workloads, and actual gains will vary depending on model configuration, request distribution, sequence length, and hardware model.
On the B300 benchmark, Kimi K3's new optimization solution on vLLM achieved a 2.2 to 2.8× throughput improvement compared to version v0.27.1.
In the B300 EP8 decoding scenario, Chord achieves a 2.15× speedup compared to the untuned default implementation.
The benchmark comparison target on H200 is the tuned public Humming, while on B300 it is the untuned default implementation.
All Facts (8)
The B300 is an NVIDIA Blackwell architecture server-grade accelerator card that offers significant improvements in memory capacity, memory bandwidth, and FP8 compute throughput compared to the previous generation H100/H800.
70%UnverifiedIDC did not specify which Nvidia products were counted in the 34% share, nor whether the data includes restricted chips such as the B200 and B300 that reportedly entered China through smuggling channels.
50%UnverifiedThe unified Radix Cache added an Agentic-Aware tail-optimized LRU eviction strategy, which reduced p99 ITL by 43.9% at concurrency 32 in an agentic trajectory test on DeepSeek-V4-Pro FP4 with 4-card B300.
50%UnverifiedThe 2.2–2.8× improvement was achieved on specific B300 hardware under specific benchmark workloads, and actual gains will vary depending on model configuration, request distribution, sequence length, and hardware model.
50%UnverifiedOn the B300 benchmark, Kimi K3's new optimization solution on vLLM achieved a 2.2 to 2.8× throughput improvement compared to version v0.27.1.
50%UnverifiedKernel and scheduling strategies tuned for the B300 may not yield the same magnitude of gains on other hardware.
50%UnverifiedIn the B300 EP8 decoding scenario, Chord achieves a 2.15× speedup compared to the untuned default implementation.
50%UnverifiedThe benchmark comparison target on H200 is the tuned public Humming, while on B300 it is the untuned default implementation.
50%