Microsoft debuts new AI-optimized Azure instances

Microsoft Corp. is extending its Azure cloud platform with the addition of a brand new occasion household designed to run synthetic intelligence fashions.

The occasion household, referred to as the ND H100 v5 sequence, made its debut at this time. 

“Delivering on the promise of superior AI for our clients requires supercomputing infrastructure, companies, and experience to handle the exponentially rising dimension and complexity of the newest fashions,” Matt Vegas, a principal undertaking supervisor at Azure’s high-performance computing and AI group, wrote in a blog post. “At Microsoft, we’re assembly this problem by making use of a decade of expertise in supercomputing and supporting the most important AI coaching workloads.”

Every ND H100 v5 occasion options eight of Nvidia Corp.’s H100 graphics processing items. Launched final March, the H100 is Nvidia’s most superior information heart GPU. It will possibly practice AI fashions 9 instances sooner than the corporate’s earlier flagship chip and performs inference as much as 30 instances sooner. 

The H100 options 80 billion transistors produced utilizing a four-nanometer course of. It features a specialised module, referred to as the Transformer Engine, that’s designed to hurry up AI fashions primarily based on the Transformer neural community structure. The structure powers many superior AI fashions together with OpenAI LLC’s ChatGPT chatbot.

Nvidia has additionally outfitted the H100 with different enhancements. The chip provides, amongst different capabilities, a built-in confidential computing characteristic. The characteristic can isolate an AI mannequin in a approach that blocks unauthorized entry requests, together with from the working system and hypervisor on which it runs.

Superior AI fashions are often deployed on not one however a number of graphics playing cards. GPUs utilized in such a way should repeatedly alternate information with one other to coordinate their work. To hurry up the circulate of information between their GPUs, firms usually hyperlink them collectively utilizing high-speed community connections. 

The eight H100 chips in Microsoft’s new ND H100 v5 cases are related to at least one one other utilizing an Nvidia know-how known as NVLink. In response to Nvidia, the know-how is seven instances sooner than PCIe 5.0, a preferred networking commonplace. Microsoft says that NVLink offers 3.6 terabits per second of bandwidth between the eight GPUs in its new cases. 

The occasion sequence additionally helps one other Nvidia networking know-how known as NVSwitch. Whereas NVLink is designed to hyperlink collectively the GPUs inside a single server, NVSwitch connects a number of GPU servers with each other. This makes it simpler to run complicated AI fashions that must be distributed throughout a number of machines in an information heart. 

Microsoft’s ND H100 v5 cases mix the H100 graphics playing cards with Intel Corp. central processing items. The CPUs are sourced from Intel’s new 4th Gen Xeon Scalable Processor sequence. The chip sequence, which is often known as Sapphire Rapids, made its debut in January.

Sapphire Rapids is predicated on an enhanced model of Intel’s 10-nanometer course of. Every CPU within the sequence contains a number of onboard accelerators, computing modules optimized for particular duties. Because of the  built-in accelerators, Intel says that Sapphire Rapids offers as much as ten instances higher efficiency for some AI functions than its previous-generation silicon.

The ND H100 v5 occasion sequence is at the moment obtainable in preview. 

Picture: efes/Pixabay