NVIDIA B200

1. Product Overview

The NVIDIA B200 (codename: GB200) is the flagship GPU of the first-generation Blackwell architecture unveiled at NVIDIA GTC 2024. It is positioned as the core hardware for large model training and high-concurrency inference for medium and large-scale LLMs. Manufactured via TSMC 4NP (approximately 4nm) process, it adopts a dual-die packaging design and is equipped with 192GB HBM3e high-bandwidth memory.

The supporting complete system DGX B200 is configured with 8× B200 SXM5 GPUs. It delivers 96 PFLOPS of FP4 inference performance and 72 PFLOPS of FP8 training performance, with a total video memory of 1.5TB. It is mainly optimized for 7B–130B parameter LLMs, long-context scenario applications, and large-scale training cluster deployment.

Form Factor: SXM5 liquid cooling version only (no PCIe air-cooled variant available)

Product Positioning: Cost-effective flagship GPU as the predecessor of B300, and a comprehensive upgraded replacement for H200

2. Core Specifications of Single B200 GPU

2.1 Architecture & Manufacturing Process

• Architecture: First-generation Blackwell (dual-die design)
• Process Technology: TSMC 4NP (≈4nm)
• Transistor Count: 20.8 billion
• Packaging: Dual-die packaging with NV-HBI inter-die interconnect (10 TB/s bandwidth)

2.2 Compute Cores

• SM Count: 148 Streaming Multiprocessors
• CUDA Cores: 20,480
• Tensor Cores: 640 fifth-generation Tensor Cores, supporting FP4, FP6, FP8 and FP16 precision computing
• Transformer Engine: Second-generation, optimized for LLM attention mechanism computing

2.3 Memory System

• Memory Capacity: 192GB HBM3e (8-stack configuration)
• Memory Bus Width: 8192-bit
• Memory Bandwidth: 8 TB/s
• ECC Support: Enabled

2.4 Dense Computing Performance (Single Card)

• FP4 (Inference): 12 PFLOPS (16 PFLOPS for sparse computing)
• FP8 (Training): 9 PFLOPS
• FP16/BF16: 4.5 PFLOPS

2.5 Interconnect & Interface

• NVLink 5: 1.8 TB/s bidirectional bandwidth
• PCIe Generation: Gen6, 256 GB/s bandwidth
• Physical Interface: SXM5 (liquid cooling)

2.6 Power Consumption & Thermal Solution

• Typical Power Draw: 1000W
• Cooling Solution: Cold-plate liquid cooling (standard configuration for DGX B200 system)

3. DGX B200 (8-GPU Complete System) Specifications

3.1 Hardware Configuration

• GPU Module: 8 × B200 SXM5 GPUs, total HBM3e memory up to 1.5TB
• CPU: 2 × Intel Xeon Platinum 6776P (64 cores / 128 threads per processor)
• System Memory: 2TB (expandable to maximum 4TB)
• NVSwitch: 2 × fifth-generation NVSwitch, aggregated bandwidth up to 14.4 TB/s
• Network Configuration: 8 × OSFP 400G InfiniBand ports; 2 × BlueField-3 400G DPU
• Storage: 8 × E1.S 3.84TB NVMe SSD (total storage: 30.72TB)
• Total System Power Consumption: Approximately 10.5kW
• Chassis Specification: 10U rack-mounted, liquid cooling design

3.2 System Computing Performance

• FP4 Inference Performance: 96 PFLOPS
• FP8 Training Performance: 72 PFLOPS
• FP16 Computing Performance: 36 PFLOPS

------------------------------------------

Many products have not been listed yet. Please contact us for more products

📧 sales5@xrjdcs.com |  📞 +86 18250705533（WhatsApp）丨Sales Manager： Jinny  

Main brand

ABB 丨 GE 丨Allen Bradley  丨  Honeywell 丨   Emerson  丨  Bently Nevada   丨 Westinghouse  

Triconex  丨Foxboro 丨   ICS Triplex  丨  Hima  丨  Schneider 丨  Yokogawa  丨 Woodward  & more

⇒The following are our main activities
Taken the ABB (AC800M/AC800F/AC800PEC/AC900F/excitation control)
★GE VMIVME/ VMIC series CPU mainboard
Relay protection SR469/SR369/SR489/SR750
IS200/IS215/IS220/ IS420/ DS200/DS215
/IC693/IC694/IC695/IC697/IC698
★ Invensys (3008, 3625, 3508E.3721, 3700A 4351B Safety system module)
★ Bentley (3500/42M, 3500/22M, 3500/95.3500/05)
★ Woodward 9907 series 505 governor, 8840 series controller, module
★ Foxboro (FBM230, FBM232, FCP270, H90/H92)
★ Emerson Westinghouse module 1C311/1C312, DELTAV module EPRO sensor
★ Dark Horse (F3330, F3236, F6217, F7126, F8621A)
★ Schneider Modicon 140 series CPU, communication module ELAU servo
★ Prosofte MVI46/56/69, AB, communication module