Now Hiring: Are you a driven and motivated 1st Line IT Support Engineer?
Unmatched End-to-End Accelerated Computing Platform
NVIDIA HGX Blackwell B200
The NVIDIA HGX B200 integrate NVIDIA Blackwell Tensor Core GPUs with high-speed interconnects to propel the data center into a new era of accelerating computing and generative AI. As a premier accelerated scale-up platform with up to 15X more inference performance than the previous generation, Blackwell-based HGX systems are designed for the most demanding generative AI, data analytics, and HPC workloads.
The NVIDIA B200 is Revolutionary Performance Backed by Evolutionary Innovation
B200 delivers leading-edge performance, offering 3X the training performance and 15X the inference performance of previous generations. Leveraging the NVIDIA Blackwell GPU architecture,
Optimize performance with our flexible infrastructure.
SIAM.AI CLOUD is a distinct cloud platform built on Kubernetes and Slurm, ensuring you enjoy bare-metal advantages without the complexity of managing infrastructure. We take care of all the complex tasks, such as handling dependencies, managing drivers, and scaling the control plane, allowing your workloads to seamlessly function without any intervention.
networking architecture powered by NVIDIA InfiniBand technology.
Our distributed training clusters featuring the B200 are designed with rail optimization and utilize NVIDIA Quantum InfiniBand networking. This configuration supports in-network computing with NVIDIA SHARP, delivering a staggering 3.2Tbps of GPUDirect bandwidth per node.
Migrate your workloads seamlessly.
Specifically tailored for NVIDIA GPU-accelerated workloads, providing seamless compatibility to run your current workloads with minimal adjustments or none at all. Whether your operations are SLURM-based or focused on containerization, our straightforward deployment solutions empower you to achieve more with less effort.
What’s inside a SIAM.AI CLOUD's DGX B200
8X
NVIDIA HGX B200 Blackwell GPUs
144
peta FLOPS inference
72
peta FLOPS training
112
Cores total of Intel® Xeon® Platinum 8570 Processors
1440GB
GPU Memory
4TB
System RAM
3200
Gbps of GPUDirect InfiniBand Networking (8x 400 Gbps InfiniBand NDR Adapters)
200
Gbps Ethernet Networking (10Gb/s onboard NIC with RJ45 or 100Gb/s dual-port ethernet NIC)
HERE TO HELP
DGX B200
Proven Infrastructure Standard
The DGX B200 is a fully integrated hardware and software solution, featuring the complete NVIDIA AI software stack. This includes NVIDIA Base Command, NVIDIA AI Enterprise software, an extensive ecosystem of third-party support, and access to expert guidance from NVIDIA professional services.
DGX B200 Performance
Next-Generation Performance Powered
Real-Time Large Language Model Inference
Performance metrics are subject to change. Token-to-token latency (TTL) is measured at 50ms in real-time, with a first token latency (FTL) of 5 seconds. Testing involved an input sequence length of 32,768 and an output sequence length of 1,028, comparing 8x eight-way DGX H100 GPUs (air-cooled) to a single eight-way DGX B200 (air-cooled) in terms of per-GPU performance.
Enhanced AI Training Performance
Performance figures are subject to adjustment. Tests were conducted using a 32,768 GPU configuration with a 4,096-node eight-way DGX H100 air-cooled cluster connected via a 400G IB network. For comparison, a 4,096-node eight-way DGX B200 air-cooled cluster using the same 400G IB network was also evaluated.
HERE TO HELP
HERE TO HELP
B200 Architecture
NVIDIA Blackwell Architecture
Blackwell-architecture GPUs integrate an impressive 208 billion transistors and are produced using a specialized TSMC 4NP process. Each Blackwell GPU consists of two reticle-limited dies, seamlessly connected via a 10 terabytes per second (TB/s) chip-to-chip interconnect, forming a unified single GPU.
Second-Generation Transformer Engine
The second-generation Transformer Engine combines custom Blackwell Tensor Core technology with innovations from NVIDIA® TensorRT™-LLM and the NeMo™ Framework. This design significantly accelerates the inference and training of large language models (LLMs) and Mixture-of-Experts (MoE) models.
To optimize MoE model inference, Blackwell Tensor Cores introduce new precisions, including community-defined microscaling formats. These formats deliver high accuracy while simplifying the transition from larger precisions. The Blackwell Transformer Engine leverages fine-grain techniques, such as micro-tensor scaling, to enhance both performance and precision, supporting 4-bit floating point (FP4) AI. This approach effectively doubles the performance and capacity of next-generation models, all while preserving exceptional accuracy.