OSDev.org

Solar wrote:Apple has been in the CPU business, and then pulled their plug on PowerPC in favor of x86... so why should they even want to go back to "rolling their own"?

Am I missing something here? Apple has never been in the business of making CPUs. They make computers. For this they choose parts, like all others who make computers. The PowerPC had been developed by IBM at the start. Then they lost interest in the 32-bit part and sold it to Freescale, who subsequently lost interest and sold it to NXP, who are currently not developing it any further. Apparently they're waiting for a sucker to hoist this onto.

PowerPC 64 OTOH is still going strong in data centers. And still developed by IBM.

But that's beside the point: Apple has never been in the CPU business. Unless you think I'm in the CPU business because I recently purchased one.

You are missing something. The PowerPC was developed by a consortium of Apple, IBM, and Motorola. This is not the first, nor the last, CPU that Apple had a hand in.

nullplan wrote: A server CPU however doesn't care about power use (as much) and is more interested in throughput than latency.

Power consumption is very much a concern for servers; hell, a standard metric for server performance is 'cycles per watt per dollar'.

It is also one of the reasons why server CPUs are clocked somewhat lower than high-performance computing CPUs (the two are not at all the same topic, mind you) or Workstation CPUs (which are still another matter, though they generally clock around the same range as HPC chips), which in turn are clocked lower than most consumer CPUs - your typical server-class Xeon or Epyc chip has a base clock somewhere under 3GHz, with the ones aimed at HPC or (for Xeon, since AMD's Threadripper line is a separate badge) workstations are usually 200-500 MHz higher. While the Coffee Lake and Epyc 3 lines have been pushing that up towards 3.5 GHz or even 4GHz for some SKUs, they are still clocked lower than all but the lowest-end contemporary Core, Ryzen, or Threadripper chips meant for the desktop.

While the other main reasons for this - longevity and stability - are in many ways the more serious concerns, both of them are tightly tied to cooling, and to the CPU's TDP - which naturally enough is determined in part by (and is a limiting factor to) power draw.

Oh, and cooling itself is a significant cost center when you have hundreds or thousands of CPUs in tightly-packed blade racks, and have to dump the waste heat not only out of the systems but out of the whole room, if not out of the entire building through a specialized HVAC system.

Simply put, power costs are a key part of a datacenter's bottom line.

Server CPUs do not only clock lower than consumer CPUs because of power consumption but also because they generally have (a lot) more cores. Their TDP is not really lower (the opposite is true), though it might be lower per core.