GPU Specifications / NVIDIA
NVIDIA B300 SXM
Blackwell Ultra datacenter GPU with 288 GB of HBM3e memory, 8000 GB/s of bandwidth and up to 5,000 TFLOPS of FP16 tensor compute.
Key Specifications
Memory
288 GB HBM3e
Memory Bandwidth
8,000 GB/s
TDP
1400 W
Architecture
Blackwell Ultra
Interconnect
NVLink 5.0 · 1800 GB/s
Est. On-demand Price
~$18.00/h
Hourly rates are indicative on-demand estimates; actual pricing varies by provider and commitment.
Compute Performance
| Precision | Peak throughput |
|---|---|
| FP64 (double precision) | 1.4 TFLOPS |
| FP32 (single precision) | 80 TFLOPS |
| FP24 | 160 TFLOPS |
| FP16 (tensor) | 5,000 TFLOPS |
| INT8 (tensor) | 10,000 TFLOPS |
| INT4 (tensor) | 20,000 TFLOPS |
Tensor figures use the vendor's peak numbers (with structured sparsity where supported).
System Requirements
Recommended CPU
Intel Xeon 6 or NVIDIA Grace (GB300 NVL72)
Max VRAM per node (8 GPUs)
2,304 GB
System RAM (min / recommended)
1024 / 2048 GB
Minimum PSU
2600 W
LLMs on the B300 SXM
Number of B300 SXM GPUs needed to serve popular models at 8-bit quantization (including 20% overhead for activations and KV cache).
| Model | Params | VRAM (8-bit) | GPUs needed |
|---|---|---|---|
| GPT-5.6 Sol | 2400B | 2682 GB | 10x B300 SXM |
| GPT-5 Flagship | 2100B | 2347 GB | 9x B300 SXM |
| GPT-5.6 Luna | 1400B | 1565 GB | 6x B300 SXM |
| Kimi K3 (1.2T) | 1200B | 1341 GB | 5x B300 SXM |
| Kimi K2.6 (1T) | 1000B | 1118 GB | 4x B300 SXM |
| GPT-5.6 Terra | 800B | 894 GB | 4x B300 SXM |
| DeepSeek V4 Pro (671B) | 671B | 750 GB | 3x B300 SXM |
| Llama 4 Behemoth (500B) | 500B | 559 GB | 2x B300 SXM |
| Claude 5 Fable (480B) | 480B | 536 GB | 2x B300 SXM |
| GLM 5.2 (400B) | 400B | 447 GB | 2x B300 SXM |
| Grok 4.5 | 350B | 391 GB | 2x B300 SXM |
| Claude 4.8 Opus (300B) | 300B | 335 GB | 2x B300 SXM |
| Muse Spark 1.1 | 300B | 335 GB | 2x B300 SXM |
| Grok 4 | 270B | 302 GB | 2x B300 SXM |
| Gemini 3.1 Pro | 250B | 279 GB | 1x B300 SXM |
| Qwen 3.7 Max (235B) | 235B | 263 GB | 1x B300 SXM |
| Mistral Large 3 (200B) | 200B | 224 GB | 1x B300 SXM |
| Grok 3 Mini | 190B | 212 GB | 1x B300 SXM |
| Claude 5 Sonnet (175B) | 175B | 196 GB | 1x B300 SXM |
| Gemini 3.5 Flash | 150B | 168 GB | 1x B300 SXM |
| Gemini 2.5 Flash | 140B | 156 GB | 1x B300 SXM |
| Llama 4 Maverick (128B) | 128B | 143 GB | 1x B300 SXM |
| DeepSeek V4 Flash (120B) | 120B | 134 GB | 1x B300 SXM |
| Qwen 3.6 Plus (110B) | 110B | 123 GB | 1x B300 SXM |
| Nova Premier (80B) | 80B | 89 GB | 1x B300 SXM |
| Qwen 3 Coder-Next (80B) | 80B | 89 GB | 1x B300 SXM |
| Claude 4.5 Haiku (70B) | 70B | 78 GB | 1x B300 SXM |
| Llama 3.3 Instruct (70B) | 70B | 78 GB | 1x B300 SXM |
| Mistral Medium 3.5 (70B) | 70B | 78 GB | 1x B300 SXM |
| Yi 1.5 (40B) | 40B | 45 GB | 1x B300 SXM |
| Nova Core (34B) | 34B | 38 GB | 1x B300 SXM |
| DeepSeek V3.1 (32B) | 32B | 36 GB | 1x B300 SXM |
| Gemma 3 (27B) | 27B | 30 GB | 1x B300 SXM |
| Mistral Small 4 (24B) | 24B | 27 GB | 1x B300 SXM |
| Yi 1.5 (15B) | 15B | 17 GB | 1x B300 SXM |
| Phi 4 (14B) | 14B | 16 GB | 1x B300 SXM |
| Nova Lite (12B) | 12B | 13 GB | 1x B300 SXM |
| Llama 3.2 Instruct (11B) | 11B | 12 GB | 1x B300 SXM |
| Gemma 3 (9B) | 9B | 10 GB | 1x B300 SXM |
| Yi 1.5 Lite (9B) | 9B | 10 GB | 1x B300 SXM |
| Phi 4 Mini (7B) | 7B | 8 GB | 1x B300 SXM |
| Phi 3.5 (3.8B) | 3.8B | 4 GB | 1x B300 SXM |
Frequently Asked Questions
How much VRAM does the NVIDIA B300 SXM have?
The NVIDIA B300 SXM has 288 GB of HBM3e memory with 8000 GB/s of memory bandwidth.
Which LLMs can run on a single B300 SXM?
At 8-bit quantization, a single B300 SXM (288 GB) can serve models up to roughly 250B parameters, such as Gemini 3.1 Pro. Larger models require multiple GPUs or more aggressive quantization.
How much does it cost to rent a NVIDIA B300 SXM?
On-demand cloud pricing for the B300 SXM is around $18.00/hour, i.e. about $13,140/month running 24/7. Actual prices vary by provider, region, and commitment.
What are the power and system requirements of the NVIDIA B300 SXM?
The B300 SXM has a TDP of 1400W. A power supply of at least 2600W per GPU is recommended. Recommended host CPUs: Intel Xeon 6 or NVIDIA Grace (GB300 NVL72).
Deploy on a GPU cloud
Rent the NVIDIA B300 SXM by the hour instead of buying hardware.
