Edited by 
Short answer
The B300 (Blackwell Ultra) is an inference-focused update of the B200. It has 288 GB of HBM3e instead of 180 GB, 13.5 PFLOPS of dense FP4 instead of 9 (50% more), twice the throughput for the exponential math in attention, and 800 Gb/s networking with PCIe 6.0. FP8 and FP16 throughput are the same, memory bandwidth stays at 8 TB/s, and NVIDIA cut FP64 from 37 to about 1.25 TFLOPS and INT8 to a small fraction. Choose B300 for serving large models with FP4; B200 for FP64 or INT8 work, or when it is cheaper.
B200 vs B300 specifications
| Specification (per GPU, HGX) | B200 | B300 | Change |
|---|---|---|---|
| Architecture | Blackwell | Blackwell Ultra | |
| Transistors | 208 billion | 208 billion | same, two dies |
| Streaming multiprocessors | 148 | 160 | +8% |
| Memory | 180 GB HBM3e | 288 GB HBM3e (12-high stacks) | +60% |
| Memory bandwidth | 8 TB/s | 8 TB/s | same |
| Dense FP4 | 9 PFLOPS | 13.5 PFLOPS | +50% |
| Dense FP8 / FP6 | 4.5 PFLOPS | 4.5 PFLOPS | same |
| Dense FP16 / BF16 | 2.25 PFLOPS | 2.25 PFLOPS | same |
| FP64 | 37 TFLOPS | About 1.2 TFLOPS | cut |
| Dense INT8 | 4.5 POPS | About 0.15 POPS | cut |
| Attention exponentials (SFU) | Baseline | 2× | faster softmax |
| NVLink per GPU | 1.8 TB/s | 1.8 TB/s | same |
| Networking per GPU | 400 Gb/s (ConnectX-7) | 800 Gb/s (ConnectX-8) | 2× |
| Host interface | PCIe 5.0 | PCIe 6.0 | 2× |
| Power | About 1,000 W | About 1,100 W | +10% |
In GB300 NVL72 racks, each Blackwell Ultra GPU runs at up to 1,400 W and reaches 15 PFLOPS of dense FP4. See both in the GPU specs database.
More memory for bigger models and longer contexts
The jump from 180 to 288 GB comes from 12-high HBM3e stacks instead of 8-high. Bandwidth stays at 8 TB/s, so a model that fits on both runs at similar speed per user. The extra capacity lets one GPU hold larger models, longer KV caches and more users at once. An 8-GPU HGX B300 has 2.3 TB of memory against 1.44 TB for HGX B200, enough to hold a 1-trillion-parameter model at 8-bit on one server.
Faster FP4 and attention
Blackwell Ultra's Tensor Cores run NVFP4 50% faster than B200's, while FP8 and FP16 stay the same. NVIDIA also doubled the special function units that compute the exponentials in softmax, which helps attention because, as Tensor Cores get faster, softmax becomes a larger share of the time. Both changes target inference of large reasoning models, which generate long outputs.
What B300 gives up
To make room, NVIDIA cut FP64 to about 1.2 TFLOPS per GPU and INT8 to a small fraction of B200's. FP64 matters for scientific simulation; INT8 matters for older quantized inference pipelines, which are moving to FP8 and FP4. If you need either, B200 is the better buy, and NVIDIA's next GPU, Rubin, restores FP64 to 33 TFLOPS.
Which should you choose?
- B300 for serving large models, long contexts and reasoning workloads, especially with FP4, and when you need 800G networking.
- B200 for training with FP8 or BF16, where the two perform alike, for FP64 or INT8 work, or when B200 capacity is cheaper.
- Coming from Hopper? Read B200 vs H100 for the bigger jump.
Frequently asked questions
Is B300 faster than B200?
For FP4 inference, yes: 50% more dense FP4 throughput, faster attention and more memory. For FP8 and FP16 training, they have the same peak throughput and memory bandwidth, so performance is similar.
What is Blackwell Ultra?
Blackwell Ultra is NVIDIA's name for the 2025 update of the Blackwell architecture, used in B300 and GB300. It keeps the same two-die design and adds more memory and faster FP4 and attention.
What is the difference between B300 and GB300?
B300 is the GPU in 8-GPU HGX servers paired with x86 CPUs. GB300 pairs Blackwell Ultra GPUs with NVIDIA's Grace CPU in NVL72 racks, where each GPU can run at higher power, up to 15 PFLOPS of dense FP4.
Why did NVIDIA cut FP64 on B300?
To spend the chip's area and power on AI formats such as FP4 that its largest customers use most. Scientific users can choose B200 or Rubin.
