Modern model pre-training often calls for larger cluster deployment to reduce time and cost. At the server level, such training workloads demand faster compute and increased memory allocation. As models grow to hundreds of billions of …
Read More
In this post, we describe an integration of Open OnDemand with AWS ParallelCluster so admins can provide web-based access to HPC resources beyond what they have at their site, by using the AWS cloud to add new capabilities and extend …
Read More
We’ve come a long way since the launch of the m1.small instance in 2006, adding instances with additional memory, compute power, and your choice of Intel, AMD, or Graviton processors. The original general-purpose “one size fits all” …
Read More
Steve Litster leads the HCLS compute specialists team, and he’s well known to most of that community because he’s worked in research, and at a pharmaceutical company, and in the tech industry. But just speaking to him makes you …
Read More
Have you ever heard about error correction? Without it, we could not obtain awe-inspiring pictures of Jupiter and its moons, conduct intelligible mobile phone calls, or have reliable computers. In this blog post, we explain the basic ideas …
Read More
GPUdirect RDMA was invented to massively speed up the ML frameworks that so many customers use to train their models using our large fleets of NVIDIA GPUs. Amr Ragab from EC2 Engineering took a few minutes at SC'22 to explain to us how it …
Read More
In AWS ParallelCluster 3.4, you can now build HPC clusters that span multiple Amazon EC2 Availability Zones. In this post, we describe how the new feature works, how to use it, and some implications for cluster design that it raises. Read …
Read More
Tom Deakin (and his colleague Simon McIntosh Smith) work on the hard problem of performance portability - the task of getting great performance out of software running on a range of computer architectures. This means they’re often …
Read More