B300
Next-generation data center GPU based on NVIDIA Blackwell Ultra architecture with 288 GB HBM3e memory, delivering 1.5x FP4 performance and 2x attention performance over the B200 for AI reasoning and real-time inference workloads.
GPU Benchmarks
B300 vs RTX 6000 Ada
Compare performance across LLMs and image models to find the best GPU for your workload.
Estimated B300 performance (not measured yet)
We have not run the Runpod benchmark suite on the B300 yet, so the charts below show the RTX 6000 Ada and every other measured GPU without a B300 bar. The table gives our estimate instead: the measured B200 result for each workload, scaled by the one specification that governs it. The B300 (Blackwell Ultra) matches the B200 on memory bandwidth (8 TB/s), dense FP16 tensor throughput (2.25 PFLOPS) and dense FP8 (4.5 PFLOPS), so for BF16 and FP8 inference on 7B to 8B models the defensible estimate is parity with the B200. Where the B300 pulls ahead is 288 GB of HBM3e against 180 GB (1.6x, so larger models and longer contexts fit on one GPU) and 1.5x dense FP4 (15 vs 10 PFLOPS), and neither shows up in these BF16/FP8 tests on small models.
| Workload | RTX 6000 Ada measured | B200 measured | B300 estimated | What sets the estimate |
|---|---|---|---|---|
| Llama 3.1 8B, 1,024 in / 1,024 out, batch 1 (tok/s) | 52 | 196 | 196 | Decode is memory-bandwidth-bound at this batch size: 8 TB/s on both parts |
| Llama 3.1 8B, 1,024 in / 1,024 out, batch 32 (tok/s) | 956 | 2,621 | 2,621 | Compute-bound at this batch size: dense FP8/FP16 tensor throughput is unchanged (4.5/2.25 PFLOPS) |
| Llama 3.1 8B, 1,024 in / 1,024 out, batch 128 (tok/s) | 1,397 | 3,928 | 3,928 | Compute-bound at this batch size: dense FP8/FP16 tensor throughput is unchanged (4.5/2.25 PFLOPS) |
| Qwen2.5 7B, 512 in / 512 out, batch 256 (tok/s) | 2,422 | 13,764 | 13,764 | Compute-bound at this batch size: dense FP8/FP16 tensor throughput is unchanged (4.5/2.25 PFLOPS) |
| Flux.1-dev, 512 px, 10 steps (s/image) | 2.4 | 1.1 | 1.1 | Diffusion is compute-bound in BF16: dense FP16 tensor throughput is unchanged (2.25 PFLOPS) |
| Flux.1-dev, 1,024 px, 50 steps (s/image) | 33.1 | 10.7 | 10.7 | Diffusion is compute-bound in BF16: dense FP16 tensor throughput is unchanged (2.25 PFLOPS) |
| SDXL 1.0, 1,024 px, 30 steps (s/image) | 4.3 | 4.4 | 4.4 | Diffusion is compute-bound in BF16: dense FP16 tensor throughput is unchanged (2.25 PFLOPS) |
How to read the estimates: expect measured B300 results within roughly 0.95x to 1.10x of the B200 figures. The low end is run-to-run noise; the high end is what the 1,400 W power envelope (against 1,000 W) and Blackwell Ultra's faster attention softmax could add on long-context runs. Values are output tokens per second for LLM rows (higher is better) and seconds per image for diffusion rows (lower is better). Sources: Runpod benchmark service measurements for the B200 and the comparison GPU; NVIDIA Blackwell and Blackwell Ultra specifications. Measured B300 results replace this table as soon as the runs land.
Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
LLM inference benchmarks.
Benchmarks were run using vLLM in May 2025 with Runpod GPUs
RTX 6000 Ada
Professional workstation GPU based on Ada Lovelace architecture with 48GB GDDR6 memory and 18,176 CUDA cores for advanced AI workloads.
H100 PCIe
High-efficiency LLM processing at 90.98 tok/s.
Image generation benchmarks.
Benchmarks were run using Hugging Face Diffusers in May 2025 on Runpod GPUs.
H100 SXM
Unmatched image gen speed with 49.9 images per minute.
RTX 6000 Ada
AI image processing at 40.3 images per minute.
H100 PCIe
Pro-grade performance with 36 images per minute.
Related Comparisons
- B300 vs RTX A4000
- B300 vs RTX 3090
- B300 vs H100 PCIe
- B300 vs RTX 4090
- B300 vs B200
- B300 vs L4
- B300 vs H100 SXM
- B300 vs A40
- B300 vs H100 NVL
Case Studies
Real-world GPU performance in action.
See how teams optimize cost and performance with the right GPU for their workloads.
.png){kind=small}
"All of these projects, the renders for AMD, the Coca-Cola builds, that has to do with scalability. If we can't scale, we can't deliver. Runpod makes that possible."
"Runpod has changed the way we ship because we no longer have to wonder if we have access to GPUs. We've saved probably 90% on our infrastructure bill, mainly because we can use bursty compute whenever we need it."
"Runpod has allowed the team to focus more on the features that are core to our product and that are within our skill set, rather than spending time focusing on infrastructure, which can sometimes be a bit of a distraction.”
"Runpod helped us scale the part of our platform that drives creation. That’s what fuels the rest, image generation, sharing, remixing. It starts with training."
"Runpod allowed us to reliably handle scaling from zero to over 1,000 requests per second in our live application."
"Runpod has allowed us to focus entirely on growth and product development without us having to worry about the GPU infrastructure at all."
"We could stop worrying about infrastructure and go back to building. That’s the real win.”
“The main value proposition for us was the flexibility Runpod offered. We were able to scale up effortlessly to meet the demand at launch.”
"After migration, we were able to cut down our server costs from thousands of dollars per day to only hundreds."
Runpod’s scalable GPU infrastructure gave us the flexibility we needed to match customer traffic and model complexity, without overpaying for idle resources.
