Aiserver Manufacturer
![Dell New Spot [New Product Launch] GB200 NVL72 Rack Liquid-cooled with 36 Grace CPUs and 72 Blackwell GPUs in Stock](https://site.cdnfile.io/www.aiservermanufacturer.com/8a4c63bac9f460e9da70c96c9a5c0e93.jpeg)
![Dell New Spot [New Product Launch] GB200 NVL72 Rack Liquid-cooled with 36 Grace CPUs and 72 Blackwell GPUs in Stock](https://site.cdnfile.io/www.aiservermanufacturer.com/a97efe7b7c04cb3104b94be0c21589ae.jpg)
![Dell New Spot [New Product Launch] GB200 NVL72 Rack Liquid-cooled with 36 Grace CPUs and 72 Blackwell GPUs in Stock](https://site.cdnfile.io/www.aiservermanufacturer.com/058687d3a7e79a7cf7c3d3e65492b1e0.jpg)
![Dell New Spot [New Product Launch] GB200 NVL72 Rack Liquid-cooled with 36 Grace CPUs and 72 Blackwell GPUs in Stock](https://site.cdnfile.io/www.aiservermanufacturer.com/f777eec6988fad4b583adebf64907be3.jpeg)
![Dell New Spot [New Product Launch] GB200 NVL72 Rack Liquid-cooled with 36 Grace CPUs and 72 Blackwell GPUs in Stock](https://site.cdnfile.io/www.aiservermanufacturer.com/08c4579d5a5e6a1b915a9ce809431cea.jpg)
1 / 5
|
GB200 Grace Blackwell Superchip
|
A key component of the GB200 NVL72, using NVLink-C2C to connect two high-performance Blackwell Tensor Core GPUs and an Grace CPU, delivering 900GB/s of bidirectional bandwidth, applications can consistently access unified memory space through NVLink-C2C to meet large memory demands.
|
|
GB200 Compute Tray
|
Based on the new MGX design, it contains two Grace CPUs and four Blackwell GPUs, equipped with a cooling plate and liquid cooling interface, supports PCIe Gen 6 high-speed networking, and an NVLink interface for NVLink cable boxes, providing 80 petaflop of AI performance and 1.7TB of fast memory per compute tray.
|
|
NVLink Switch System
|
The GB200 NVL72 rack-level system uses an NVLink Switch system with 9 NVLink switch trays and a cable box for interconnecting GPUs and switches to improve parallel model efficiency across 18 compute nodes. Each NVLink switch tray provides 144 100GB of NVLink ports, and 9 switches can fully connect 72 NVLink ports on 18 Blackwell GPUs.
|
|
Powerful Computing Power
|
The GB200 NVL72 introduces advanced features and a second-generation Transformer engine that supports FP4 AI, and when combined with the fifth-generation NVLink, it can provide 30 times faster real-time LLM inference performance for trillion-parameter language models. Its second-generation Transformer engine has FP8 accuracy, providing 4x the training performance for large language models compared to the same number of H100 GPUs.
|
|
Efficient Communication
|
The NVLink switch system features 36 GB200 interconnected by the largest NVLink domain ever interconnected to deliver low-latency GPU communications at 130 terabytes per second for AI and high-performance computing (HPC) workloads. With a revolutionary 1.8TB/s bidirectional throughput per GPU, more than 14 times the bandwidth of PCIe 5.0, it delivers seamless, high-speed communication for today's most complex large models.
|
|
Superior Data Processing
|
The Blackwell architecture introduces a hardware decompression engine with native support for decompressing data using LZ4, Deflate, and Snappy compression formats, delivering up to 800GB/s of performance, enabling Grace Blackwell to execute up to 18x faster than CPUs (Sapphire Rapids) and up to 100x faster than H6 Tensor Core GPUs in query benchmarks. With 8TB/s of high memory bandwidth and Grace CPU high-speed NVlink chip-to-chip (C2C), the engine speeds up the entire process of database querying.
|
|
Energy Saving Advantages
|
The liquid-cooled GB200 NVL72 rack reduces the carbon footprint and energy consumption of data centers, delivering 25x more performance at the same power consumption compared to the H100 air-cooled infrastructure, while also reducing water consumption.
|
