Skip to main content

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

ParameterValue
ArchitectureBlackwell Ultra (GB300)
Process NodeTSMC 4NP
GPU Chips2 Blackwell dies (CoWoS-L packaging)
Memory288 GB HBM3e (12-Hi stack)
Memory Bandwidth8 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
INT87,000 TOPS
TDP1,400 W (liquid cooling required)
NVLink Bandwidth1.8 TB/s (5th Gen)
PCIeGen 6 (first generation)
DC NetworkingConnectX-8, 1.6 Tbps
ReleaseJanuary 2026 — shipping

B200 vs B300 Ultra Upgrade Comparison

MetricB200B300 UltraImprovement
ArchitectureBlackwellBlackwell UltraMid-cycle upgrade
Memory192 GB HBM3e288 GB HBM3e+50%
Memory Bandwidth8 TB/s8 TB/sFlat
FP8 Dense4.5 PFLOPS7 PFLOPS+56%
FP4 Sparse~9 PFLOPS14 PFLOPS+56%
TDP1,000 W1,400 W+40%
PCIeGen 5Gen 6
DC NetworkingConnectX-7 (400G)ConnectX-8 (1.6T)
Release2024-Q42026-01

Key: FP4 is Blackwell Ultra's new precision tier (between FP8 and INT4), reducing memory footprint by another 50% compared to FP8.

H100 / H200 / B300 Generational Performance

MetricH100H200B300Improvement
ArchitectureHopperHopperBlackwell Ultra
Memory80GB HBM3141GB HBM3e288GB HBM3e3.6×
Memory Bandwidth3.35 TB/s4.8 TB/s8 TB/s2.4×
FP8 Dense989 TFLOPS989 TFLOPS7 PFLOPS
TDP700W700W1,400W
NVLink900 GB/s900 GB/s1,800 GB/s
Release2023-032024-Q42026-01

DeepSeek Inference Benchmarks (vLLM, Feb 2026 Report)

DeepSeek-V3.2 (GB300)

Test config: NVFP4 quantization + TP2 (Tensor Parallel 2)

ScenarioThroughput (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)

ScenarioThroughput (TGS)
Prefill-only (ISL=2k, batch=256)22,476
Mixed context (ISL=2k, OSL=1k)3,072

R1 prefill throughput is approximately that of V3.2, benefiting from R1's chain-of-thought architecture optimizations.

FP4 vs FP8 Quantization (DeepSeek-R1)

Quantization SchemePrefill ImprovementMixed Context Improvement
NVFP4 + TP2 vs FP81.8×

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 .

B300 vs H200 Generational Performance

MetricB300 vs H200
Prefill Throughput (ISL=2k)
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.

ScenarioRecommended Config
DeepSeek R1 online servingB300 + NVFP4 + EP2 (Expert Parallel)
DeepSeek V3 inference + trainingB300 + NVFP4 + TP2 (Tensor Parallel)
Long-context document understandingB300 (full use of 288GB memory)
Cost-sensitive inferenceB300 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

ParameterValue
Total GPU Memory2.3 TB HBM3e (288GB × 8)
GPU InterconnectNVLink 5.0 + ConnectX-8
Peak Power~14 kW (2× H100 DGX)
Supported ModelsFull loading of 400B+ parameter models
CoolingLiquid cooling required (DLC)

Cloud Pricing Comparison (March 2026)

ProviderInstance TypePer GPU/Hour Price
AWSp6-b200.48xlarge (8× B300)$11.70
DigitalOceanB300 GPU Droplet (coming soon)~$8.00 (estimated)
Oracle CloudOCI 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)

GPUThroughput (tok/s)Per GPU/HourToken Cost (Relative)
H100 SXM~21,800$2.001.0× (baseline)
H200 SXM~31,700$3.500.83× (17% savings)
B300 (FP8)~100,000+~$8.000.58× (42% savings)
B300 (FP4)~150,000+~$8.000.39× (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 1,400Wliquid 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)

Vendor Information

ParameterValue
VendorNVIDIA Corporation
Product Pagehttps://www.nvidia.com/en-us/data-center/blackwell/
LaunchJanuary 2026 — shipping
Cloud DeploymentAWS / DigitalOcean / Oracle Cloud
OEM PartnersDell / HPE / Supermicro / Lenovo