NVIDIA B300 Ultra (Blackwell Ultra)
Product Overview
NVIDIA B300 / B300 Ultra (codename Miranda / GB300) is the mid-cycle upgrade of the Blackwell architecture, shipping January 2026. The biggest upgrade is memory — 192GB → 288GB HBM3e, with FP4 sparse compute reaching 14 PFLOPS and TDP at 1,400W (liquid cooling required).
Purpose-built for the era of ultra-large LLM inference — 288GB memory can load a 70B model in FP16 on a single GPU, leaving 100GB+ for KV Cache. In DeepSeek R1 benchmarks, prefill throughput hits 22,476 TGS, an 8× improvement over H200.
Core Specifications
| Parameter | Value |
|---|
| Architecture | Blackwell Ultra (GB300) |
| Process Node | TSMC 4NP |
| GPU Chips | 2 Blackwell dies (CoWoS-L packaging) |
| Memory | 288 GB HBM3e (12-Hi stack) |
| Memory Bandwidth | 8 TB/s |
| FP8 Tensor Core (dense) | 7 PFLOPS |
| FP8 Tensor Core (sparse) | 14 PFLOPS |
| FP4 Tensor Core (dense) | 7 PFLOPS |
| FP4 Tensor Core (sparse) | 14 PFLOPS |
| FP16 Tensor Core (dense) | 3.5 PFLOPS |
| INT8 | 7,000 TOPS |
| TDP | 1,400 W (liquid cooling required) |
| NVLink Bandwidth | 1.8 TB/s (5th Gen) |
| PCIe | Gen 6 (first generation) |
| DC Networking | ConnectX-8, 1.6 Tbps |
| Release | January 2026 — shipping |
B200 vs B300 Ultra Upgrade Comparison
| Metric | B200 | B300 Ultra | Improvement |
|---|
| Architecture | Blackwell | Blackwell Ultra | Mid-cycle upgrade |
| Memory | 192 GB HBM3e | 288 GB HBM3e | +50% |
| Memory Bandwidth | 8 TB/s | 8 TB/s | Flat |
| FP8 Dense | 4.5 PFLOPS | 7 PFLOPS | +56% |
| FP4 Sparse | ~9 PFLOPS | 14 PFLOPS | +56% |
| TDP | 1,000 W | 1,400 W | +40% |
| PCIe | Gen 5 | Gen 6 | 2× |
| DC Networking | ConnectX-7 (400G) | ConnectX-8 (1.6T) | 4× |
| Release | 2024-Q4 | 2026-01 | — |
Key: FP4 is Blackwell Ultra's new precision tier (between FP8 and INT4), reducing memory footprint by another 50% compared to FP8.
| Metric | H100 | H200 | B300 | Improvement |
|---|
| Architecture | Hopper | Hopper | Blackwell Ultra | — |
| Memory | 80GB HBM3 | 141GB HBM3e | 288GB HBM3e | 3.6× |
| Memory Bandwidth | 3.35 TB/s | 4.8 TB/s | 8 TB/s | 2.4× |
| FP8 Dense | 989 TFLOPS | 989 TFLOPS | 7 PFLOPS | 7× |
| TDP | 700W | 700W | 1,400W | 2× |
| NVLink | 900 GB/s | 900 GB/s | 1,800 GB/s | 2× |
| Release | 2023-03 | 2024-Q4 | 2026-01 | — |
DeepSeek Inference Benchmarks (vLLM, Feb 2026 Report)
DeepSeek-V3.2 (GB300)
Test config: NVFP4 quantization + TP2 (Tensor Parallel 2)
| Scenario | Throughput (TGS) |
|---|
| Prefill-only (ISL=1) | 7,360 |
| Mixed context (ISL=2k, OSL=1k) | 2,816 |
ISL = Input Sequence Length, OSL = Output Sequence Length
DeepSeek-R1 (B300)
| Scenario | Throughput (TGS) |
|---|
| Prefill-only (ISL=2k, batch=256) | 22,476 |
| Mixed context (ISL=2k, OSL=1k) | 3,072 |
R1 prefill throughput is approximately 3× that of V3.2, benefiting from R1's chain-of-thought architecture optimizations.
FP4 vs FP8 Quantization (DeepSeek-R1)
| Quantization Scheme | Prefill Improvement | Mixed Context Improvement |
|---|
| NVFP4 + TP2 vs FP8 | 1.8× | 8× |
NVFP4 (NVIDIA FP4) is a new 4-bit floating-point format introduced with Blackwell, reducing memory by another 50% vs FP8 and multiplying throughput several times over. While maintaining accuracy (FP4 + tensor parallelism), DeepSeek-R1 mixed-context inference improves by 8×.
| Metric | B300 vs H200 |
|---|
| Prefill Throughput (ISL=2k) | 8× |
| Short Output Throughput (ISL=2k, OSL=128) | 20× |
20× improvement on short-output scenarios — B300 + NVFP4 + TP2 is the optimal choice for high-concurrency production environments.
Recommended Deployment Configurations (DeepSeek)
| Scenario | Recommended Config |
|---|
| DeepSeek R1 online serving | B300 + NVFP4 + EP2 (Expert Parallel) |
| DeepSeek V3 inference + training | B300 + NVFP4 + TP2 (Tensor Parallel) |
| Long-context document understanding | B300 (full use of 288GB memory) |
| Cost-sensitive inference | B300 Spot + FP4 quantization |
EP2 = Expert Parallel 2, suited for MoE models (DeepSeek is MoE)
TP2 = Tensor Parallel 2, general-purpose acceleration
8-GPU DGX B300 System
| Parameter | Value |
|---|
| Total GPU Memory | 2.3 TB HBM3e (288GB × 8) |
| GPU Interconnect | NVLink 5.0 + ConnectX-8 |
| Peak Power | ~14 kW (2× H100 DGX) |
| Supported Models | Full loading of 400B+ parameter models |
| Cooling | Liquid cooling required (DLC) |
Cloud Pricing Comparison (March 2026)
| Provider | Instance Type | Per GPU/Hour Price |
|---|
| AWS | p6-b200.48xlarge (8× B300) | $11.70 |
| DigitalOcean | B300 GPU Droplet (coming soon) | ~$8.00 (estimated) |
| Oracle Cloud | OCI B300 | ~$10.00 (estimated) |
AWS p6-b200.48xlarge is one of the first 8-GPU B300 instances. DigitalOcean pricing is ~30% cheaper than AWS.
Mainstream GPU Inference Cost Comparison (Llama 70B)
| GPU | Throughput (tok/s) | Per GPU/Hour | Token Cost (Relative) |
|---|
| H100 SXM | ~21,800 | $2.00 | 1.0× (baseline) |
| H200 SXM | ~31,700 | $3.50 | 0.83× (17% savings) |
| B300 (FP8) | ~100,000+ | ~$8.00 | 0.58× (42% savings) |
| B300 (FP4) | ~150,000+ | ~$8.00 | 0.39× (61% savings) |
Key insight: B300 has a higher per-unit price, but per-token cost is actually 39–61% lower — making it the optimal choice for cloud inference.
Cooling & Infrastructure
- TDP 1,400W — liquid cooling required (Direct Liquid Cooling, DLC)
- Air cooling is not feasible (vs H100 700W air-cooled)
- 8-GPU DGX B300 = 14kW (= 2× H100 DGX)
- Data center power and cooling must be re-planned
Software Requirements
- CUDA 12.x
- cuDNN 9.x
- TensorRT-LLM 0.15+
- NVFP4 support (requires TensorRT 10+)
- vLLM 0.6+ (GB300 optimized)
Use Cases
- ✅ Large-scale inference serving (70B+ models, 100K+ tok/s)
- ✅ Inference-intensive workloads (DeepSeek R1, o1-class reasoning models)
- ✅ Long-context KV Cache (288GB fully retained)
- ✅ 400B+ parameter model deployment (8-GPU DGX B300 full loading)
- ✅ Multi-node training clusters (6.4 Tbps GPU interconnect)
- ❌ Small-to-medium inference (H200 more economical)
- ❌ No liquid-cooled facility (high infrastructure investment)