- Last Active
blastdoor said:ecarlseen said:I'm happy that Apple is still addressing the pro market, but it's unfortunate they missed so badly using Intel Xeon CPUs instead of AMD Epyc CPUs for this generation. The additional cores and PCIe bandwidth would be massively useful, especially if they could shave some serious money off the price at the same time.
But, I can think of three reasons why the path they chose might be ok:
1. AVX 512 --- the AltiVec of our day. Xeon has it, Epyc doesn't.
2. AMD is historically an unreliable supplier on the CPU side of things
3. Intel has already announced that Cascade Lake Xeon Ws will be much cheaper than their Skylake predecessors. That could mitigate the Epyc price advantage.
"Pretty good" is a bit of an understatement. Epyc Rome thrashes Xeon Cascade Lake more badly than Apple's A-series CPUs beat Qualcomm's Snapdragon. It's all just flat-out annihilation with the exception of single-threaded loads on low core-count CPUs, where Intel maintains a very slight edge. On the top end, Epyc's multi-core performance is about 260% of Skylake at less than half the price. I don't expect this to change until 2022 unless Intel gets incredibly lucky. Their 10nm performance is still utter crap - stuff coming off the line right now is slower than 14nm unless they bump the TDP, and who wants to use more power for the same speed? Intel is still excellent at core and uncore design, but they can't manufacture at a competitive level so it doesn't matter. Ultimately, this is a leadership issue. AMD has a fairly brilliant engineer (Lisa Su) running things, and Intel has been trying to cost-cut itself to death (turns out that dumping engineering resources while bringing up 10nm cost them several orders of magnitude more that it "saved"). Anyway, if early-silicon Ice Lake vs. early-silicon Milan is anything to go by, it only gets worse in the near term. Much, much worse.
AVX512 is a distinct advantage in single-threaded loads (and how many of those are there for AVX512?), but I don't think it overcomes the massive overall performance differential (assuming one doesn't dump those loads to a GPU anyway).
With regards to reliability - AMD had their shot back in that Athlon64 / Opteron days and promptly fumbled the opportunity because they couldn't match Intel in IPC or performance-per-watt. Both of these issues have been overcome with Rome. AMD's Naples was a solid performer for the money (I bought a few of their servers to play with), Rome is awesome, and Milan is looking great. In the meantime, Intel has basically screwed things up fairly consistently for the past seven years (14nm was a mess and 10nm is still deeply problematic unless magic happens). Furthermore, AMD gambled heavily on their "chiplet" solution for scaling core count in a single package and it payed off perfectly. Intel doesn't have anything to compete with that in the near term, and even if they started their own project in the last year or two then they're still likely looking at a 2022 shipment timeframe (in line with when they expect 7nm to start volume production, assuming all goes well).
Intel will have to cut Cascade Lake prices to stay in the game, but it doesn't matter because except in a very few edge cases the performance will not be close. The 56-core Cascade Lake part is a joke - it requires water cooling, offers only one DIMM per channel, and nobody other than Intel is even bothering to make motherboards for it. AMD's 64-core Rome beats it soundly with air cooling, a full load of memory, massively more PCIe bandwidth, and for far less money. Even if Intel matches the price, they can't do anything about the performance (or lack thereof).
The only reason I can see for Apple to stay with Xeon for their Mac Pros is if macOS so heavily optimized for Intel's architecture that adjusting for AMD's Zen2 / Zen3 cores is too problematic to be worth the effort. AMD still isn't terribly competitive in the mobile / laptop space, which is where Apple sells most of their Intel-based machines (although there are some interesting APU possibilities there).
Ultimately, I don't care whether Intel or AMD supplies the CPUs - I just need some serious performance bumps. We've been treading water in the industry for over half a decade now, and it's getting very stale.
I'm happy that Apple is still addressing the pro market, but it's unfortunate they missed so badly using Intel Xeon CPUs instead of AMD Epyc CPUs for this generation. The additional cores and PCIe bandwidth would be massively useful, especially if they could shave some serious money off the price at the same time.