{"id":357,"date":"2016-12-06T23:53:35","date_gmt":"2016-12-06T23:53:35","guid":{"rendered":"http:\/\/redmonk.com\/fryan\/?p=357"},"modified":"2016-12-07T00:01:39","modified_gmt":"2016-12-07T00:01:39","slug":"aws-reinvent-elastic-gpus-fpgas-deep-learning-compute","status":"publish","type":"post","link":"https:\/\/redmonk.com\/fryan\/2016\/12\/06\/aws-reinvent-elastic-gpus-fpgas-deep-learning-compute\/","title":{"rendered":"AWS re:Invent – Elastic GPUs, FPGAs, Deep Learning & Compute"},"content":{"rendered":"

Amazon Web Services annual conference, re:Invent<\/a>, was held in Las Vegas last week, and to say there is a lot to unpack would be somewhat of an understatement. As my colleague Stephen O\u2019Grady<\/a> put it, Amazon pack more announcements into a two-day window than most companies manage in three years.<\/p>\n

This is the first of several posts we will be doing on various aspects of re:Invent, and in this post we will primarily focus on Deep\u00a0Learning, Elastic GPUs and FPGAs, and touch on some of the interesting compute announcements at the end.<\/p>\n

Elastic GPUs<\/h2>\n

Earlier this year Amazon announced the availability of their P2 instances<\/a>, which featured up to 16GPUS for heavy machine learning. Given the comments\u00a0over the years on GPU performance in AWS<\/a>\u00a0 it is fair to say that the P2 instances were eagerly anticipated and went a long way towards resolving many of the previous complaints around virtualized GPUs.<\/p>\n

However, not everyone needs even a full GPU, never mind a dedicated instance. This is what makes the Elastic GPU concept so interesting. Attach as you need it, do the work, and then move on.<\/p>\n

The initial Elastic GPU offering<\/a> allows you to take as little as 1\/8th<\/sup> of a GPU and use it as you need, and then move on. The offering is limited to Windows for now, with plans for support in the Amazon Linux Deep Learning AMI<\/a> soon. There is one significant area of concern with the lack of CUDA support<\/a>.<\/p>\n

\n

1\/8 GPU to a full GPU that you can attach to compute #reinvent<\/a> << really can't say how much this changes things for lots of users<\/p>\n

— Fintan Ryan (@fintanr) November 30, 2016<\/a><\/p><\/blockquote>\n