Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.

crazy dave

macrumors 65816
Sep 9, 2010
1,453
1,229
Is there an official explanation for why the 64 core variant is only 3,000 points higher than the 32 core variant? Something is not right here.
These tests may have been done in Windows. There's a problem with Windows scheduling (at least in 10 and below, I don't know about 11) and high high core CPUs. I can't remember where the problems start, but it's high. There are workarounds but Geekbench scores in particular seems to take a hit because of it. I think someone mentioned in the thread above, the scores are higher and more reasonable in Linux.
 

crazy dave

macrumors 65816
Sep 9, 2010
1,453
1,229
GFXBench barely uses the CPU, I don't see why it shouldn't saturate the GPU in all offscreen tests.
And there is at least one subtest (Manhattan 1440 offscreen) where the M1 ultra is more than 1.5 times better than the best M1 Max, which suggests that this is the 64-core variant... and that performance scaling is therefore not great (about 1.5x instead of 2x).
We now have two tests of the M1 ultra showing almost identical results (the average being very close to the top result). So this appears to the representative of the what the M1 ultra GPU can do in this benchmark app. It's clearly not faster than the 3090 except in "Manhattan".

A test doesn't have to be limited by the CPU in order for it to not saturate the GPU - especially for compute workloads. I'll admit I'm less familiar with graphics workloads but I understand that it's similar. The issue is that not enough meaningful work is being sent to the GPU. This is especially a problem on older benches that were designed for less powerful graphics cards: you eventually get a topping out as not enough of the GPU's resources can be brought to bear before the problem is over or the frame is outputted. You can even see this in the relatively recent GB5 compute test: scores both for the M1 Max (apparently doesn't spin its clocks fast enough before the subtests are over) and in the higher Nvidia GPUs (spins its clocks up fast, but simply too many cores that subtests don't saturate with work) don't scale properly. I've done this myself on my own work where the beefier GPU on the identical CPU doesn't look as good as is promised until the problem set gets bigger.

Looking at the scores (including but not limited to for Manhattan 1440p offscreen and Aztec High), the M1 Pro can be 2.1x faster than the M1, where it should be only 2x. I've noticed that there don't appear to be different categories for different GPU core counts of the different M1s (7 vs 8, 14 vs 16, 24 vs 32, 48 vs 64). So hmmm ... that makes things extra interesting potentially ... especially since these are also user submitted (i.e. averages and especially top scores can be increased from overclocked components) and there are a few odd ones if you try to get into the details (like an M1 - plain M1 - that scores better than any GPU ever). Hopefully the Anandtech review will be out soon.

Still if this is the 64-core variant, then, yes, the scaling is not so good on Aztec High where it had been excellent before and it would be ... nice to know where it is more powerful ;).
 
Last edited:

jeanlain

macrumors 68020
Mar 14, 2009
2,461
955
. The issue is that not enough meaningful work is being sent to the GPU. This is especially a problem on older benches that were designed for less powerful graphics cards:
But GFXBench appears to scale well on other powerful cards, like the 6900XT.
Maybe the Metal M1 Ultra driver doesn't send tasks to all cores when it determines that power can be saved?
 

leman

macrumors Core
Oct 14, 2008
19,521
19,678
These tests may have been done in Windows. There's a problem with Windows scheduling (at least in 10 and below, I don't know about 11) and high high core CPUs. I can't remember where the problems start, but it's high. There are workarounds but Geekbench scores in particular seems to take a hit because of it. I think someone mentioned in the thread above, the scores are higher and more reasonable in Linux.

If I understand it correctly, the issue is that Windows kernel does not scale beyond 64 threads (32 cores with SMT). If you want to use more CPU cores in the same process you have to use specialised APIs or something like that. Geekbench probably does not do that.

But GFXBench appears to scale well on other powerful cards, like the 6900XT.
Maybe the Metal M1 Ultra driver doesn't send tasks to all cores when it determines that power can be saved?

The way I remember the explanation given by more knowledgeable people, it is about power management. The task is too short to trigger the high-performance GPU mode. The power management of dGPUs is probably different, they don't have to care about energy conservation as much.
 
  • Like
Reactions: bobcomer

Lihp8270

macrumors 65816
Dec 31, 2016
1,143
1,608
The Mac Studio will be used soon but the AMD Ryzen Threadripper 5990X will be also used soon.
I think it is better to compare these two and not using the old 3990X to show how good Apple is compared to a very old CPU xD.
The 3990X is also half the price at 4K than the OP claimed.
 

Lihp8270

macrumors 65816
Dec 31, 2016
1,143
1,608
These tests may have been done in Windows. There's a problem with Windows scheduling (at least in 10 and below, I don't know about 11) and high high core CPUs. I can't remember where the problems start, but it's high. There are workarounds but Geekbench scores in particular seems to take a hit because of it. I think someone mentioned in the thread above, the scores are higher and more reasonable in Linux.
The updates to the Windows scheduler was to work with the new Intel P and E core chips.

Other (x86) CPUs shouldn’t be affected by the changes.
 

dugbug

macrumors 68000
Original poster
Aug 23, 2008
1,929
2,147
Somewhere in Florida
The 3990X is also half the price at 4K than the OP claimed.


newegg is 8K

and amazon has all vendors but one at 8k. That one vendor is 4K. I dunno, they were all 8K when I posted.
 

Lihp8270

macrumors 65816
Dec 31, 2016
1,143
1,608
newegg is 8K

and amazon has all vendors but one at 8k. That one vendor is 4K. I dunno, they were all 8K when I posted.
Looks like those are scalped end of line prices as that certainly isn’t typical retail pricing.

In the UK I believe 3990X retail pricing was at £3900 ish
 
  • Like
Reactions: mi7chy

ader42

macrumors 6502
Jun 30, 2012
436
390
Is it reasonable to compare pricing of a single computer part (PC) to an entire computer (Mac)?

It would make more sense to compare a complete system price (disreagrd KVM).

With reasonably equivalent gfx card, ssd, ram and thunderbolt ports you will be talking £8k imho even with a 3080 Ti

Going from a 32 core 3975WX to a 64 core 3995WX costs an extra £2,000.

The Mac Studio looks cheap to me.
 

jeanlain

macrumors 68020
Mar 14, 2009
2,461
955
The way I remember the explanation given by more knowledgeable people, it is about power management. The task is too short to trigger the high-performance GPU mode. The power management of dGPUs is probably different, they don't have to care about energy conservation as much.
I don't think GFXBench subtests are too short. These tests run for about 1 min or more.
 

mi7chy

macrumors G4
Oct 24, 2014
10,623
11,296
Last edited:
  • Like
Reactions: Ulfric

mi7chy

macrumors G4
Oct 24, 2014
10,623
11,296

deconstruct60

macrumors G5
Mar 10, 2009
12,493
4,053
System Mac13,2 Apple M1 Ultra 3220 MHz (20 cores)
Uploaded March 9th, 2022
iro
Platform macOS
Single-Core Score 1793
Multi-Core Score 24055

System Mac13,2 Apple M1 Ultra 3220 MHz (20 cores)
Uploaded March 8th, 2022
Platform macOS
Single-Core Score 1793
Multi-Core Score 24055
Is
that head-banging multicore score seems to hold. To get an idea of just how big 24K is on multicore, here are the top CPU scores. The only one that beat the ultra? The $8000 3990X cpu.
AMD Ryzen Threadripper 3990X
2.9 GHz (64 cores)
25157

AMD Ryzen Threadripper 3970X
3.7 GHz (32 cores)
22334

Intel Xeon W-3175X
3.1 GHz (28 cores)
20910

the official AMD price for 3995WX is $5,489 .
https://www.tomshardware.com/news/amd-threadripper-pro-retail-pricing-6000-64-cores-3000-32-cores

street prices for loose retail is somewhat like the aftermarket GPU market; way off MSRP. Should get better for 3000 series as the 5000 series rolls out ( demand for 3000 should drop significantly )

Geekbench on 3995WX where filter out Windows :

46,993 on MT . So not even close .

As for 5995WX ( no geekbench scores but some passmark scores can compare )




There should be a Lenovo 600 series with it in a few weeks . AMD is only selling these to OEMs and the vendors can lock the CPU to their firmware . But CPU package prices probably not changing too much more. from 3000 series . Some uplift though. Package Max RAM capacity is way up : 2TB .




so yeah a $10k workstation but can provision on order lager capacity in some of these too.
 

Lihp8270

macrumors 65816
Dec 31, 2016
1,143
1,608
Is it reasonable to compare pricing of a single computer part (PC) to an entire computer (Mac)?

It would make more sense to compare a complete system price (disreagrd KVM).

With reasonably equivalent gfx card, ssd, ram and thunderbolt ports you will be talking £8k imho even with a 3080 Ti

Going from a 32 core 3975WX to a 64 core 3995WX costs an extra £2,000.

The Mac Studio looks cheap to me.
Not really.

If you have the correct workflow for Apple Silicone, and the real world figures are representative. The mac studio is extremely cheap for the performance.

However, it’s equally misleading to inflate the price of competition in a comparison. Particularly when you don’t even need to do so to make the point.
 

Lihp8270

macrumors 65816
Dec 31, 2016
1,143
1,608
the official AMD price for 3995WX is $5,489 .
https://www.tomshardware.com/news/amd-threadripper-pro-retail-pricing-6000-64-cores-3000-32-cores

street prices for loose retail is somewhat like the aftermarket GPU market; way off MSRP. Should get better for 3000 series as the 5000 series rolls out ( demand for 3000 should drop significantly )

Geekbench on 3995WX where filter out Windows :

46,993 on MT . So not even close .

As for 5995WX ( no geekbench scores but some passmark scores can compare )




There should be a Lenovo 600 series with it in a few weeks . AMD is only selling these to OEMs and the vendors can lock the CPU to their firmware . But CPU package prices probably not changing too much more. from 3000 series . Some uplift though. Package Max RAM capacity is way up : 2TB .




so yeah a $10k workstation but can provision on order lager capacity in some of these too.
Iirc I saw an article that the upcoming threadripper pro or whatever they call them are going to be OEM only and won’t be available to buy for consumers.

You’re only going to get them through buying a whole workstation machine.

AMD like they did last time they had the advantage over Intel, and like Intel did recently for years. They’re resting on their success and taking advantage of it. Instead of pushing on with what got them ahead.
 

crazy dave

macrumors 65816
Sep 9, 2010
1,453
1,229
The updates to the Windows scheduler was to work with the new Intel P and E core chips.

Other (x86) CPUs shouldn’t be affected by the changes.
I'm simply referring to the fact that I know the problems with high core count CPUs and GB were there in Windows 10 and below. I was not aware of it being fixed nor of it not being fixed in Windows 11 separate from the P/E scheduling.
 

crazy dave

macrumors 65816
Sep 9, 2010
1,453
1,229
But GFXBench appears to scale well on other powerful cards, like the 6900XT.
Maybe the Metal M1 Ultra driver doesn't send tasks to all cores when it determines that power can be saved?
Short answer: I dunno. Long answer too.
 

Admiral

macrumors 6502
Mar 14, 2015
408
991
The 3990X is also half the price at 4K than the OP claimed.

Four thousand dollars happens to be the price of the M1 Ultra Mac Studio itself, including its graphics capabilities. All in. That 3990X has to have water cooling or a mongo heatsink and fan ($150-200), go into a motherboard ($800), be equipped with 64GB RAM ($300), you gotta buy a case ($100), power supply (a beefy one — $150), and storage ($100), and then you have to buy a graphics card ($1000) — assuming you can find one. Mac Studio is a screaming bargain.
 
Last edited:
  • Like
Reactions: ader42
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.