B300 SXM vs V100 SXM2 32GB
NVIDIA B300 SXM (Blackwell Ultra, 288 GB) against NVIDIA V100 SXM2 32GB (Volta, 32 GB): memory, compute, power and rental price, compared for LLM inference and training.
Pick two GPUs to compare
Side-by-Side Specifications
| Spec | B300 SXM | V100 SXM2 32GB |
|---|---|---|
| Architecture | Blackwell Ultra | Volta |
| Memory | 288 GB HBM3e | 32 GB HBM2 |
| Memory bandwidth | 8,000 GB/s | 900 GB/s |
| FP16 tensor compute | 5,000 TFLOPS | 125 TFLOPS |
| INT8 tensor compute | 10,000 TOPS | 62.8 TOPS |
| Interconnect | NVLink 5.0 · 1800 GB/s | NVLink 2.0 · 300 GB/s |
| TDP | 1400 W | 300 W |
| Est. on-demand price | ~$18.00/h | ~$2.00/h |
| FP16 TFLOPS per $/h | 278 | 63 |
Highlighted values indicate the stronger spec. Hourly rates are indicative on-demand estimates.
Verdict
Raw performance: The B300 SXM leads on FP16 tensor compute (40.0x advantage), which translates directly into higher token throughput for inference and shorter training steps.
Memory: With 288 GB per card, the B300 SXM fits larger models on fewer GPUs - fewer cards means less inter-GPU communication and simpler deployments.
Value: At current on-demand rates, the B300 SXM delivers more compute per dollar (278 vs 63 FP16 TFLOPS per $/h). If your model fits in its VRAM budget, it is usually the more economical choice.
GPUs Needed for Popular LLMs
Cards required to serve each model at 8-bit quantization (with 20% overhead for activations and KV cache).
| Model | VRAM (8-bit) | B300 SXM | V100 SXM2 32GB |
|---|---|---|---|
| GPT-5.6 Sol | 2682 GB | 10x | 84x |
| GLM 5.3 (743B) | 830 GB | 3x | 26x |
| Claude 4.8 Opus (300B) | 335 GB | 2x | 11x |
| Grok 3 Mini | 212 GB | 1x | 7x |
| Qwen 3.6 Plus (110B) | 123 GB | 1x | 4x |
| Nova Core (34B) | 38 GB | 1x | 2x |
| Phi 4 (14B) | 16 GB | 1x | 1x |
| Mage VL | 5 GB | 1x | 1x |
| LFM 2.5 2.6B | 3 GB | 1x | 1x |
Frequently Asked Questions
Which is better for LLM inference: B300 SXM or V100 SXM2 32GB?
The B300 SXM delivers more raw FP16 compute (5,000 TFLOPS) and the B300 SXM offers the most memory per card (288 GB). For cost-efficiency, the B300 SXM currently gives more FP16 TFLOPS per dollar of on-demand rental (278 vs 63 TFLOPS per $/h).
How much more memory does the B300 SXM have?
The B300 SXM has 288 GB of HBM3e versus 32 GB of HBM2 for the V100 SXM2 32GB - a ratio of 9.00x in favor of the B300 SXM. More VRAM per card means fewer GPUs to fit a given model.
Is the B300 SXM or the V100 SXM2 32GB cheaper to rent?
Estimated on-demand rates are ~$18.00/h for the B300 SXM and ~$2.00/h for the V100 SXM2 32GB. Raw hourly price is only part of the story: normalize by throughput (TFLOPS per $/h) and by how many cards you need for your model's VRAM.
How do the B300 SXM and V100 SXM2 32GB compare on power?
The B300 SXM has a TDP of 1400W versus 300W for the V100 SXM2 32GB. FP16 compute per watt: 3.6 vs 0.4 TFLOPS/W.
Deploy on a GPU cloud
Rent the B300 SXM or V100 SXM2 32GB by the hour instead of buying hardware.
