NVIDIA B300 Tensor Core GPU

Overview#

B300 — marketed as 'Blackwell Ultra' — is the mid-cycle Blackwell refresh, much as H200 was for Hopper. The compute architecture is the same dual-die Blackwell silicon as B200, but the HBM3e stacks step up to 12-high, raising per-GPU capacity to 288 GB. NVIDIA also rebinned the part for higher sustained FP4 throughput.

Positioning is squarely at reasoning-model inference and long-context workloads. The 50 % memory uplift over B200 directly translates into longer contexts, larger batches, and the ability to host frontier-MoE inference replicas on fewer GPUs.

Specifications vs B200#

Why a Mid-Cycle Refresh#

The B300 refresh exists for the same reason H200 did: HBM density and bandwidth were the binding constraint on inference economics, not raw compute. Through 2024-2025, SK hynix and Micron 12-high HBM3e stacks reached production; B300 is the first SKU to package them at scale.

Reasoning models — models that emit long internal chains-of-thought before producing a user-visible answer — amplify the memory pressure. A frontier reasoning model serving 100K-token contexts at batch 64 can consume hundreds of gigabytes of KV cache per replica; the 288 GB-per-GPU envelope is the practical difference between one replica and four.

When to Pick B300#

• Inference of reasoning models with long internal chains-of-thought, where KV-cache pressure dominates.
• Frontier MoE inference where activated-parameter memory plus expert KV state exceeds the B200 envelope.
• Single-GPU replicas of 200B+ parameter dense models where avoiding tensor parallelism is worth the price premium.
• Pick B200 when supply or cost dominate and your KV-cache budget fits in 192 GB.
• Pick GB300 NVL72 when the unit of work is a full pod and the larger Grace coupling helps.

Operational Notes#

• Higher TDP than B200 (estimated ~1,400 W) — rack power budget should be revisited even on existing liquid-cooled designs.
• Drop-in compatibility with HGX-B200 baseboards is OEM-specific; confirm with your platform vendor before assuming a clean swap.
• Software stack is identical to B200 — CUDA 12.4+, driver R550+, TensorRT-LLM, vLLM Blackwell backend.
• HBM3e 12-high supply remains the primary availability constraint through 2026.

Software Ecosystem#

B300 inherits the B200 software stack without changes. Kernels tuned for Blackwell SMs run identically; the only consideration is making sure inference servers (vLLM, TensorRT-LLM, SGLang) are configured to use the full 288 GB rather than defaulting to 192 GB budgets carried over from B200 deployments.