CPUs are still often used for graph algorithms and other workloads that either don't map very well to the gpu model of parallelism or contain nontrivial control flow.

They still value raw CPU grunt for legacy software that can't easily be ported to use GPU. There's a lot of old software that has received billions in verification and validation work (especially in the nuclear area, which is important for DoE's mission) and porting it to GPU would invalidate that. This software is typically being replaced by more modern stuff, but it is often validated against its predecessor, so you still need to have the resources to run the older software.

My understanding was that one of POWER's biggest selling points is the huge number of memory channels each socket can support and the way that the CPU connects to the accelerators, which other platforms don't enable, using interfaces like OpenCAPI and BlueLink. Nvidia has been working with IBM closely on this, I gather.

Info: https://www.nextplatform.com/2016/10/17/opening-server-bus-c...

> I'm curious why they don't target having simpler/cheaper CPUs that can be moved closer to the GPUs.

You can certainly buy silicon that's set up that way from Nvidia.

Can you provide a link? Is it an ARM system?

The Parker SoC that comes with the Jetson TX2 came to mind immediately, since I've got one:

https://www.hotchips.org/wp-content/uploads/hc_archives/hc28... https://elinux.org/Jetson_TX2#Processing_Components

Six different ARM CPUs (two different kinds!) on the same silicon as a fairly powerful GPU.

Because there is latency and overhead involved when shuffling things to the GPU, and a lot of the focus on POWER these days is on OpenCAPI/NVLink. If there is an ARM chip fast enough to make OpenCAPI support worthwhile, I would be interested to hear about it.