For Supercomputing Intel is giving their usual broad but shallow offerings of new toys and CPUs. This time SemiAccurate has bits of CPUs, Knights *, software, and a change in direction for Intel.
That change of direction is the big one, did you know Intel is now an AI focused company? They are going to have a day about it later in the week, but everything is now deep learning/AI/HPC focused. There are two prevailing ways people look at the new orientation from the company, that it is just a PR stunt or that it is a genuine sea change. SemiAccurate takes a hard left once again and has come to the conclusion that there really is no change, it is business as usual, but it isn’t a stunt either.
Why do we say this? If you look at how much production AI is done on Intel CPUs vs anything else, you quickly come to the conclusion that Intel is the, pardon the pun, core of AI. There are several new parts coming from Intel that we will talk about later on which will be AI focused in future revisions. Cynics will say they are adding a popular tag to existing roadmap items but SemiAccurate thinks there are enough real changes under the hood to make it much more than a PR stunt.
In short this change of hardware has been coming for a while. Intel is well aware of the need for AI focused hardware and it will be out in less than a year. Hardware has a three-year development cycle so you can do the math. What is changing is that Intel has finally noticed that Wall Street seems to care about AI, they think it is a big deal. Intel has it, has had it, and will keep having it, they are just now banging the drum loudly. It isn’t a drumbeat without backing, nor is it cynical, it is just a conservative company starting to shout about what they already are doing.
What about the hardware? The first big news is a new Broadwell-SP/Xeon E5v4 CPU with the traditional Intel awful naming schemes. The top line part used to be the E5 2699, a 22-core 2.2GHz part with 55MB of L3 cache. The new one is the E5 2699A which bumps the clock up to 2.4GHz but everything else remains the same. Unfortunately both of these seem to have dumb caches because if you look at ARK, there is an unannounced E5 2699R with a smartcache, not just a dumb one. Again win/win on the naming Intel.
But wait there’s more, specifically on Knights *, in this first case * = Landing. If you recall from the launch in June, there was a general shortage of a lot of KNL SKUs. Today we get the breathless news that by mid-2017, Intel will have caught up with the backlog and there will be ‘general availability’. I’ll give you a moment to cool down from that huge news, catch your breath people. From there the hits keep coming with the news that the news * will be Mill or Knights Mill in full, and it will be AI focused. Whoa. For real news, the integrated fabric SKU variants will be green-lighted within weeks, that is actual news.
From there, what do you get if you take an Altera Arria 10 FPGA and put it on a PCIe card? The new Intel Deep Learning Inference Accelerator. To be fair it comes pre-loaded with software and tools for deep learning and related AI tasks. It is not just a simple FPGA-on-a-stick, it has CNN algorithms built-in, tools for the host as well, and is aimed at ease of use for AI programmers. Think kick the tires, not push the boundaries.
It is complex, that is why it is useful
From here we move on to software, specifically Intel’s HPC Orchestrator. This is a suite of tools and configurators built on OpenHPC. It is similar in nature to the older Scalable Systems Framework but now with a bent on HPC. It lets you configure a system from small to very large with pre-tested and pre-configured blocks all built on the OpenHPC framework. It is basically an HPC cluster configuration in a box or on a memory stick as the case may be. If you are not one of the biggest datacenters out there, the HPC Orchestrator is going to make your life a lot easier and therefore cheaper.
Last up we will end on a shiny note, Lustre. Intel has had their version of Lustre for a while with Seagate having the only real competing version. Seagate has now put their effort behind Intel’s variant and so the world of Lustre is now almost a shiny singularity. Ending on that bright note we have new hardware, software, and lots of non-category specific updates once again. Nothing earth-shaking once but there is a lot of useful stuff for AI and HPC, exactly what Supercomputing is all about.S|A
Latest posts by Charlie Demerjian (see all)
- HyperX ships it’s 60 millionth enthusiast memory module - Oct 15, 2018
- Bittware/Nallatech water cools 300W of Xilinx FPGA - Oct 12, 2018
- More on Intel’s 10nm process problems - Sep 17, 2018
- Intel puts out another 14nm 2020 server platform - Sep 11, 2018
- Why Can’t Intel Supply Enough 14nm Xeons? - Sep 10, 2018