Timeline for No performance increase when porting Python multiprocessing calculations from Apple M1 Pro to Dual CPU Xeon E5-2687W v4
Current License: CC BY-SA 4.0
13 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Sep 3, 2023 at 5:03 | vote | accept | albert | ||
| Sep 1, 2023 at 18:47 | answer | added | Mokubai♦ | timeline score: 2 | |
| Sep 1, 2023 at 14:35 | history | mod moved comments to chat | |||
| Sep 1, 2023 at 13:31 | comment | added | Mokubai♦ | Please note that Xeons have never really been about raw performance. They are used in servers where reliability trumps everything else. The Xeon is also on a 14 nanometer process while the M1 is on 5nm process so the 160W rating on the Xeon is deceptive, it may sound like it should be a more powerful processor but how many watts it takes doesn't really give a good indication of how much work it can do. That apple can give you an M1 that performs nearly identically at a fraction of the power is actually kindof impressive. | |
| Sep 1, 2023 at 13:16 | comment | added | Mokubai♦ | From experience hyperthreading can add 15-25% of a boost overall, which would somewhat bring those multicore passmark scores into line if you have it disabled. If youve disabled a feature used during a benchmark then that benchmark figure is no longer valid and has to be adjusted. | |
| Sep 1, 2023 at 13:14 | comment | added | Mokubai♦ | The xeon is far, far older. Look at the single core ratings on that comparison and you'll see that when scaled up the M1 should absolutely trounce the Xeon and for single core tasks it kind of does at ~3.1 seconds vs ~4 seconds. It's when you get to high core count stuff that the M1 falls behind. You should re-enable hyperthreading on the Xeon at the very least as that may let it make better use of the core resources at higher core counts which will be at least part of where it's better multicore score will come from. Without hyperthreading you are limiting the Xeon. | |
| Sep 1, 2023 at 13:06 | comment | added | albert | @Mokubai: What do you mean by "[...] but seems to perform identically otherwise." Based on my findings, the M1 Pro seems to perform identically to my Dual Xeon machine, but I would assume the latter should be (at least slightly) better in performance since I am using multiprocessing to run multicore calculations. | |
| Sep 1, 2023 at 12:51 | comment | added | Mokubai♦ | The M1 has odd mix of performance. It has a significantly higher single core rating but lower multicore performance cpubenchmark.net/compare/4580vs2765.2/… which suggests that there is something weird going on. Likely the M1 is hitting a power or thermal budget. Given that it is 6 years newer the single core performance being better is expected, but the multicore performance seems to be rather poor by comparison. The M1 may have better performance per watt, but seems to perform identically otherwise. | |
| Sep 1, 2023 at 12:04 | comment | added | albert | @DanielB: Thanks for that advice. I was thinking about asking that at SO, but was not sure if it suits there as it seemed to be more hardware or architecture related. Need to do some further investigations in order to get a small reproducible code snippet for SO. | |
| Sep 1, 2023 at 11:40 | comment | added | Daniel B | It is highly likely you made a programming mistake which renders multithreading (or multi-processing) ineffective. However, Super User isn’t the place to ask about that; Stack Overflow is. If you want to go that way, remember to include the gist of the code and whatnot, according to the rules and customs over there. | |
| Sep 1, 2023 at 11:27 | history | edited | albert | CC BY-SA 4.0 |
added 429 characters in body
|
| Sep 1, 2023 at 11:22 | history | edited | albert | CC BY-SA 4.0 |
added 429 characters in body
|
| Sep 1, 2023 at 11:16 | history | asked | albert | CC BY-SA 4.0 |