Wednesday, March 9th 2022
Apple's Brand New Mac Studio With the M1 Ultra CPU Gets First Benchmark Figures
Less than 24 hours after Apple's launch event, the first Geekbench numbers for the new Apple M1 Ultra CPU are out and the numbers are interesting to say the least. For starters, the system the Geekbench numbers are from, is the top of the range 20 Core SKU with 128 GB of RAM. This helps us get some additional insight into Apple's new CPUs. As Apple didn't provide much in technical terms yesterday, nor on its website, we now know that the clock speed of the M1 Ultra is the same 3.2 GHz as the regular M1. It also appears that the CPU cache remains the same, even though Geekbench is only listing the cache of the efficiency cores for some reason.
Although Geekbench isn't a reliable cross-platform benchmark, we do at least get an idea of how the new SoC from Apple performs. The single core performance is more or less on par with the Apple M1 Max, but loses out quite easily to Intel's Alder Lake processors. However, once we move to the multi-threaded test, the M1 Ultra really shows what it's capable of. Surprisingly the performance scaling is almost linear with the double of performance CPU cores compared to the M1 Max, which suggests that Apple's multi-chip module design is extremely capable. The interesting thing will be to see how well this design scales for GPU intensive applications. Stepping outside of the Apple ecosystem, the M1 Ultra ends up somewhere around an AMD Ryzen Threadripper 3970X in terms of multi-core performance. Scaling over some of the detailed tests aren't somewhere between 80-90 percent depending on the particular test compared to the M1 Max, if we compare to the faster results on Geekbench, which is still quite impressive considering we're looking at two M1 Max CPUs that are technically glued together.
Sources:
Geekbench, via Benchleaks
Although Geekbench isn't a reliable cross-platform benchmark, we do at least get an idea of how the new SoC from Apple performs. The single core performance is more or less on par with the Apple M1 Max, but loses out quite easily to Intel's Alder Lake processors. However, once we move to the multi-threaded test, the M1 Ultra really shows what it's capable of. Surprisingly the performance scaling is almost linear with the double of performance CPU cores compared to the M1 Max, which suggests that Apple's multi-chip module design is extremely capable. The interesting thing will be to see how well this design scales for GPU intensive applications. Stepping outside of the Apple ecosystem, the M1 Ultra ends up somewhere around an AMD Ryzen Threadripper 3970X in terms of multi-core performance. Scaling over some of the detailed tests aren't somewhere between 80-90 percent depending on the particular test compared to the M1 Max, if we compare to the faster results on Geekbench, which is still quite impressive considering we're looking at two M1 Max CPUs that are technically glued together.
38 Comments on Apple's Brand New Mac Studio With the M1 Ultra CPU Gets First Benchmark Figures
Amazing performance.
Also, Geekbench, AKA the benchmark for smartphones. Entirely appropriate for this Apple garbage.
Performance of a 32 Core/64 Thread processor that runs at 3.7 GHz and has a default power draw of 280 watts without any IGP ( we must add around 50-60 watts for an entry level IGP just to get any video output from this processor, making it to 330 watts )
is matched by
a 20 core, 20 thread processor, that runs at 3.2 GHz and has a maximum power draw of 100 watts while matching the GPU render performance of Core i9-12900K + RTX 3090 ( based on the footnote 2 and 3 and graphs on this Apple link www.apple.com/tr/newsroom/2022/03/apple-unveils-m1-ultra-the-worlds-most-powerful-chip-for-a-personal-computer/ )
The difference is 37.5% less cores, %13.5 less max frequency, 69% less power draw in favor of M1 Ultra.
That is really amazing performance.
I, personally, didn't like Apple branded computers with Intel CPU's at all; and to get more performance I had built Hackintosh'es for much less cost. My last Hackintosh build was based on Ryzen 3700X, 32 GB RAM, RX 460 and lots and lots of NVMe storage, powered by a 850W CoolerMaster Gold PSU, and I could match the best of the best Mac Pro desktop performance while compiling code; granted it was working as loud as a tractor.
Then the Mac mini running on M1 chip came out. I didn't have enough money, so I bought an 8 Core M1, 16 GB RAM and 512 GB NVMe disk as that was what I could afford. The moment I got the box in my hand, I thought that was a joke, that damn thing was weightless, smaller than 1/20 of Ryzen Hackintosh and the weight was 1/10 of it. The moment I fired it up the same code compile, there was absolutely no noise at all, 0 decibel, and the same code compile finished faster.
I put the Ryzen on sale that evening, before going to bed.
When you reach the age 53 like me and after spending your last 41 years using computers and getting paid for working on computers your last 31 years, as far as the performance is the same, you appreciate noiseless performance over noisy performance.
But damn it Apple... These Apple computers are very expensive. Granted, Windows computers with similar performance are much heavier, much noisy, heat much more... and cost the same; but regardless, there is no denying that Apple computers are very expensive.
On another note: Please don't make the mistake of confusing the numbers provided by Apple press releases with numbers provided by Intel press releases ( Intel's numbers are always some kind of cherry on some obscure benchmarks that run some obscure instruction sets, my own experience, on the tasks that I am doing, says that Apple's numbers are real ).
Currently Intel has returned back to the old Pentium 4 days, it can increase performance only by increasing clock frequency that draws insane amount of power, which is unstainable.
What are both Intel and AMD plans for future to squeeze out more performance ? Increase parallelism by combining more CPU cores on multi-chip-modules, which Apple has already done.
On my official Apple store web page, for the new Mac Studio, the difference between ( M1 Max CPU + 64 GB RAM + 1 TB NVMe SSD ) and ( M1 Ultra CPU + 64 GB RAM + 1 TB NVMe SSD ) is around US$ 1,400; that is insane. OK, twice the CPU power, twice the GPU power... but that is still insane. :kookoo:
I wish I had the money to be an Apple fanboy :(
But the workstation announced here ( www.techpowerup.com/292759/lenovo-launches-p620-workstation-featuring-amd-ryzen-threadripper-5000-pro ) has a 1000 W PSU; whereas Mac Studio has 370W PSU; and that damn thing is slighylt bigger than a 1000W PSU itself.
That said, this is an impressive enough product in many ways, but Apple is clearly charging for that performance as well.It is considering the power draw and form factor.
See above with regards to Geekbench and I also mentioned it in the news post.
Here is the actual listing for the above article;
Apple M1Ultra 20core
browser.geekbench.com/v5/cpu/13345054
Here is a the 16c/32t Ryzen TR 3955WX;
browser.geekbench.com/v5/cpu/13364853
And here is the 32c/64t Ryzen TR 3975WX
browser.geekbench.com/v5/cpu/13173781
And the 64c/128t Ryzen TR 3995WX
browser.geekbench.com/v5/cpu/13296365
M1Ultra competes with AMD's offerings upto the 16core model but once we jump to the next model Ryzen cleans Apples clock. As AMD doesn't offer a 20c/40t model it leaves us looking for the average between those offerings. However, Intel does offer 20core model Xeons.
Intel Xeon Silver 4316 Workstation CPU
browser.geekbench.com/v5/cpu/12666256
Intel Xeon Gold 5320 Server CPU
browser.geekbench.com/v5/cpu/12452334
In these cases, Apple has an edge in the single core performance, but they in no way have the edge in multi threaded performance against either AMD or Intel.
It's also possible to find slower and faster versions of all the AMD and Intel CPU's in the Geekbench database, hence why there are a wide range of conclusions drawn.
Just a few examples of the Threadripper 3970 numbers.
browser.geekbench.com/search?utf8=%E2%9C%93&q=Threadripper+3970X
browser.geekbench.com/v4/cpu/compare/15870393?baseline=15728414
browser.geekbench.com/v5/cpu/compare/4653248?baseline=4642934
Look at how some of the math heavy tests where the x86 CPUs were categorically faster were removed in Geekbench 5, like SEGMM and SFFT. Why did they do that ? Who knows, I might have an idea or two but those are really basic tests which are useful especially if we're talking about CPUs that are meant to be used in workstations.
The Max doesn't out preform the Pro in Geekbench 5, they are both essentially tied. The Pro, Max and Ultra having the same single core speed makes sense, but the multicore speed seems off.
They said that mac pro is yet to come, so maybe wait for that before comparing to systems of that ’category’.
As for why Apple chips "miraculously" get faster with each release, it's probably due to them taking an interest in software optimization across the industry. They're porting every major application they can to their chips, of course you're going to see the results get better as they further optimize.