NVIDIA introduced the Blackwell Ultra B300 Data Center GPU today at CEO Jensen Huang’s GTC 2025 keynote in San Jose, CA. The new GPU has 50% more memory and FP4 compute than the B200, pushing the competition for faster, more advanced AI models even further. NVIDIA describes it as built for the age of reasoning, pointing to advanced AI LLMs like DeepSeek R1 that can do more than regurgitate information they’ve already processed.
The Blackwell Ultra B300 is far more than a single GPU. In addition to the base B300, NVIDIA is releasing the new B300 NVL16 server racks, a GB300 DGX station, and GB300 NV72L full rack solutions. Combining 8 NV72L racks creates the Blackwell Ultra DGX SuperPod, which includes:
- 288 GRES CPUs
- 576 Blackwell Ultra GPUs
- 300 TB of HBM3e memory
- 11.5 EXA flops of FP4
NVIDIA refers to these large systems as AI factories.
NVIDIA claims the Blackwell Ultra will have 1.5 times the FP4 compute density, but it is unclear whether other compute types have increased by the same amount. We expect similar improvements, but NVIDIA may have made changes beyond just adding more SMs, such as raising clock speeds or expanding HBM3e memory. For example, clock speeds are lower in FP8 or FP16 modes. Here are the main specs we know so far, with some inferred data indicated by question marks.
NVIDIA Blackwell Ultra B300 vs Blackwell B200
| Platform | B300 | B200 | B100 |
| Configuration | Blackwell GPU | Blackwell GPU | Blackwell GPU |
| FP4 Tensor Dense/Sparse | 15/30 Petaflops | 10/20 Petaflops | 7/14 Petaflops |
| FB6/FB8 tensorDense/Sparse | 7.5/15 petaflops? | 5/10 petaflops | 3.5/7 petaflops |
| INT 8 Tensor Dense/Sparse | 7.5/15 petaops? | 5/10 petaops | 3.5/7 petaops |
| FP16/BF16 Tensor Dense/ sparse | 3.75/7.5 Petaflops? | 2.5/5 Petaflops | 1.8/3.5 petaflops |
| TF32 Tensor Dense/Sparse | 1.88/3.75 petaflops? | 1.25/2.5 Petaflops | 0.9/1.8 petaflops |
| FP-64 Tensor Dense | 68 TeraFLOPS? | 45 TeraFLOPs. | 30 teraflops |
| Memory | 288 GB (8x 36 GB) | 192 GB (8x 24 GB) | 192 GB (8x 24 GB) |
| Bandwidth | 8 TB/s? | 8 TB/s | 8 TB/s |
| Power | ? | 1300 W | 700 W |
Asked NVIDIA for more details about the Blackwell Ultra B300’s performance and received this response: “Blackwell Ultra GPUs in GB300 and B300 are different chips than Blackwell GPUs in GB200. Blackwell Ultra GPUs are designed to meet the demand for test-time scaling inference with a 1.5x increase in FP4 compute. This suggests that the B300 might be a physically larger chip to accommodate more tensor scores, but we are still waiting to verify.”
The new B300 GPUs will deliver much higher computing throughput than the B200, with 50% more on-package memory. They can support even larger AI models with more parameters, and the extra compute power will be a big advantage.
NVIDIA shared some performance examples, but compared the B300 to Hopper, which makes the results less clear. It would be more helpful to see direct comparisons between the B200 and B300 using the same number of GPUs, but that information isn’t available yet.
With FP4 instructions and the new Dynamo software library, the B300 can serve reasoning models like DeepSeek much more efficiently. NVIDIA claims an NV72L rack can deliver 30 times the inference performance of a similar Hopper setup. This improvement comes from faster NVLink, more memory, increased compute, and the use of FP4.
For example, Blackwell Ultra can process up to 1,000 tokens per second with the DeepSeek R1671B model, whereas Hopper can process only 100 tokens per second. This means throughput is ten times higher, reducing the time to handle a large query from 1.5 minutes to just 10 seconds.
300 products are expected to start shipping in the second half of the year before year-end, and NVIDIA hopes to avoid any packaging issues or delays this time. Last fiscal year, the company made $11 billion from Blackwell B200/B100, and it’s probably targeting an even higher number this year.
