1. Product Overview

2. Core Specifications of B300 GPU (Single Chip)

2.1 Architecture & Manufacturing Process

• Architecture: Blackwell Ultra (upgraded version of B100, dual-reticle large-scale chip)
• Process Technology: TSMC 4NP (approximately 4nm)
• Transistor Count: 20.8 billion (about 2.6 times that of H100)
• Packaging: Dual-Reticle Dual-Die; inter-chip NV-HBI interconnect speed up to 10 TB/s

2.2 Compute Cores

• Streaming Multiprocessors (SM): 160 units
• CUDA Cores: 20,480 units (160 SM × 128 cores per SM)
• Tensor Cores: 640 units (5th-generation, supporting FP4/FP8/FP16 computing formats)
• Transformer Engine: 2nd-generation, optimized for LLM attention mechanism computation

2.3 Memory System (Key Highlight)

• Memory Capacity: 288GB HBM3e (12-layer stacking design)
• Memory Bus Width: 8192-bit
• Memory Bandwidth: 8 TB/s (2.4 times that of H100)
• ECC Support: Enabled

2.4 Computing Power (Single Chip, Dense Computing)

• FP4 (Inference): 18 PFLOPS (24 PFLOPS for sparse computing)
• FP8 (Training): 9 PFLOPS
• FP16: 4.5 PFLOPS

2.5 Interconnect & Interface

• NVLink 5: 1.8 TB/s bidirectional bandwidth per GPU card
• PCIe Gen6: 256 GB/s bandwidth
• Interface Specification: SXM5 dedicated high-speed interconnect interface

2.6 Power Consumption & Heat Dissipation

• Typical Power Consumption: 1400W
• Heat Dissipation Solution: Liquid cooling (cold-plate liquid cooling adopted for the entire DGX B300 system)