Our clients depend on Azure AI infrastructure to develop modern AI-driven options, which is why we’re delivering new cloud-based AI-supercomputing clusters constructed with Azure ND H200 v5 sequence digital machines (VMs) immediately.
The necessity for scalable and high-performance infrastructure continues to develop exponentially because the AI panorama advances. Our clients depend on Azure AI infrastructure to develop modern AI-driven options, which is why we’re delivering new cloud-based AI-supercomputing clusters constructed with Azure ND H200 v5 sequence digital machines (VMs) immediately. These VMs are actually typically out there and have been tailor-made to deal with the rising complexity of superior AI workloads, from foundational mannequin coaching to generative inferencing. The dimensions, effectivity and enhanced efficiency of our ND H200 v5 VMs are already driving adoption from clients and Microsoft AI providers similar to Azure Machine Studying and Azure OpenAI Service.
“We’re excited to undertake Azure’s new H200 VMs. We’ve seen that H200 affords improved efficiency with minimal porting effort, we’re trying ahead to utilizing these VMs to speed up our analysis, enhance the ChatGPT expertise, and additional our mission.” —Trevor Cai, head of infrastructure, OpenAI.
The Azure ND H200 v5 VMs are architected with Microsoft’s techniques method to boost effectivity and efficiency, and have eight NVIDIA H200 Tensor Core GPUs. Particularly, they deal with the hole as a result of GPUs rising in uncooked computational functionality at a a lot sooner charge than the hooked up reminiscence and reminiscence bandwidth. The Azure ND H200 v5 sequence VMs ship a 76% improve in Excessive Bandwidth Reminiscence (HBM) to 141GB and a 43% improve in HBM Bandwidth to 4.8 TB/s over the earlier era of Azure ND H100 v5 VMs. This improve in HBM bandwidth allows GPUs to entry mannequin parameters sooner, serving to cut back general utility latency, which is a vital metric for real-time functions similar to interactive brokers. The ND H200 V5 VMs also can accommodate extra complicated Massive Language Fashions (LLMs) throughout the reminiscence of a single VM, bettering efficiency by serving to customers keep away from the overhead of operating distributed jobs over a number of VMs.
The design of our H200 supercomputing clusters additionally allows extra environment friendly administration of GPU reminiscence for mannequin weights, key-value cache, and batch sizes, all of which instantly affect throughput, latency and cost-efficiency in LLM-based generative AI inference workloads. With its bigger HBM capability, the ND H200 v5 VM can assist greater batch sizes, driving higher GPU utilization and throughput in comparison with ND H100 v5 sequence for inference workloads on each small language fashions (SLMs) and LLMs. In early checks, we noticed as much as 35% throughput improve with ND H200 v5 VMs in comparison with the ND H100 v5 sequence for inference workloads operating the LLAMA 3.1 405B mannequin (with world dimension 8, enter size 128, output size 8, and most batch sizes – 32 for H100 and 96 for H200). For extra particulars on Azure’s excessive efficiency computing benchmarks, please learn extra right here or go to our AI Benchmarking Information on the Azure GitHub repository for extra particulars.
The ND H200 v5 VMs come pre-integrated with Azure Batch, Azure Kubernetes Service, Azure OpenAI Service and Azure Machine Studying to assist companies get began immediately. Please go to right here for extra detailed technical documentation of the brand new Azure ND H200 v5 VMs.