Intel Xeon Scalable Processor Family (Skylake)

Broadwell vs Skylake Processors

Everybody talks about what the Broadwell processors are and what the Skylake processors do, as well as describe how Intel does the Tick-Tock releases. For HPC, there are a few differences that really make a difference when looking at Broadwell vs Skylake processors (Intel Xeon Processor Scalable Family).

HPL Performance

First, a lot of people talk about HPL performance, and this benchmark is still what is used for ranking on the Top 500 list. Compared to the 16 Double Precision Floating Point Operations per second (DP FLOPs) per cycle for Broadwell, the Gold and Platinum Skylake processors can do 32 DP FLOPs per cycle. This means that for every CPU core, you get twice the theoretical performance. It is important to note that the Bronze and Silver CPUs can only do 16 DP FLOPs per cycle. And not all of your applications are going to run twice as fast as it did with Broadwell processors. There are a lot of optimizations in code which is needed to take advantage of all of the features, such as AVX-512.

Memory Speed and Capacity

Another important factor when looking at HPC performance is memory speed, and capacity. The Broadwell CPUs can accept 2133MHz and up to 2400 MHz for memory speed. For Skylake, the Bronze CPUs still take 2133MHz memory, and the Silver CPUs take 2400MHz memory, but the Gold and Platinum can take 2666MHz memory DIMMs. Not only do you get higher memory speeds, but you also get six (6) memory channels (path from the processor to memory) per CPU, meaning with one memory DIMM per channel (DPC), you get twelve memory DIMMs in a dual socket server. This is an upgrade from Broadwell which only had four (4) memory channels per CPU. With a dual-socket Skylake server, if you populate out two (2) DIMMs per channel (share each memory to CPU path with two memory DIMMs), you can get 24 DIMMs and using 64GB LRDIMMs, you can easily get to 1.5TB of memory.

Multi-processor Servers

There are times when researchers want large core count multi-processor servers. The Gold processors can be used in a 4-socket server, and the Platinum processors can be used in 8 (or larger) socket servers. This means you can get 88 Gold processor cores in a single four-way server, or 228 Platinum processor cores in a single eight-way server. With hyperthreading, you can get 456 CPU threads in a single server!

Parallel Processing

Finally, with parallel processing, networking becomes very important. Intel has put their Omni-Path Architecture (OPA) fabric, right only the CPU chip. These CPUs have a “-F” after the CPU model. This means that network traffic no longer has to go through the motherboard and through the PCIe slot, but a cable comes right out of the CPU for network connectivity. And you not only get one OPA connection, but two per CPU. In a dual-socket system, you can get two “-F” CPUs in the system, or both CPUs can share the OPA connection through the UPI connection between the two CPUs.

The Intel Xeon Processor Scalable Family has been created for performance, and with HPC users in mind. There are a lot of new features available as described above, and in general, most code can take advantage of some of these features without any changes to the code. For best performance, the code would need to be updated and optimized to take advantage of all that these new CPUs offer. Still have questions? An Aspen Engineer would be happy to answer your questions.