News First Apple M1 Ultra Benchmark Posted, Nearly Matches Threadripper 3990X

[SkyBill40]

Reasonable people having ordinary conversations don't instantly bark about YOUR RESPONSIBILITY TO PROVE blah blah blah when they doubt something, especially if it's not an outlandish idea. They just politely ask for more info, and only get your level of rude if the person in my position is evasive and refuses to provide any support. You went nuclear from the word go. It's almost as if you're really angry about something. I wonder what that could be?

I simply stated you should provide your citation to refute the claim. That's it. A reasonable person would do so, yet here we are. Why? Moving on, I didn't "go nuclear" whatsoever. That's your perception of it and nothing more. I'm not even remotely agitated let alone angry and while I suppose a hot take on it might seem as such, you're reading way too far into it. Believe me there.

Good lord you actually wrote "Apple mad", and now you mad it got turned back on you, eh?

It was a light and jovial jab, nothing else. Not mad. Amused, perhaps... but not mad.

One can make a distinction between the Geekbench results database, which is not great, and Geekbench the benchmarking tool, which is decent.

It wasn't always that way. Someone in this thread linked to the thing where Linus Torvalds was bashing it. Without even clicking through the link I'm guessing it was one of his rants about its scores treating AES cryptography as a CPU test. He didn't like this because most CPU tests in GB are genuinely testing the CPU core itself, but the AES test will use an accelerator on most platforms.

Modern GB still does AES the same way, but weights it so low in the composite scores that it doesn't influence them much. Not perfect, but that objection was mostly dealt with.

Fair enough. Thank you for that take.

I think you're a little confused here. The Tom's article was about an anonymous Geekbench score showing up in the database. Nobody knows who put it there. It probably wasn't Apple.

One possibility is that a journalist got their review unit and ran the free version of GB5, which auto-uploads an anonymized report after every run. (One feature you unlock by paying for Geekbench is the ability to control these uploads, including not doing them at all.) Another possibility is that some troll decided to have some fun.

The benchmarks Apple uses to promote the Mac Studio are mostly professional creative apps - Final Cut Pro, Affinity Photo, Photoshop, and so on. They don't use Geekbench.

Again, fair enough.

My point was that this isn't any different. If you were as familiar with the merits and demerits of Geekbench as you pretended to be, you wouldn't find it a controversial idea that its database has a lot of fake and low quality scores.

I made no claim or assertion to my knowledge or lack thereof with regard to Geekbench; to be honest, I'm not as familiar with it as I am with several of the more well known and used benchmarking tools which this site runs to test gear. If there's a lot of bogus stuff in GB and it's seen as largely unreliable, well, it shouldn't be taken seriously or referenced but that would somewhat deflate the article to degree. Would you not agree with that? As you seem to have more insight on the merits or demerits of GB and its use, I'll defer to you on it.

 

[user7007]

Also the 3990x doesn't scale in many workloads. Compared to the 3970x (32 cores) it's barely faster in some things (like geekbench). So while they it's not lying to say the 3990x on windows scores poorly in geekbench and the M1 Ultra gets a similar result in MacOS - it's also likely comparing a best case M1 result to worst case Zen2 result and the article makes no attempt to qualify it.

 

[techconc]

Impressive result especially for the power usage.
Not completely unexpected when considering the M1 ultra uses 114 billion 5nm transistors while the treadripper uses 39 billion 7nm transistors. The M1 ultra chip is huge.

I’m not sure what point you’re trying to make here regarding transistor count. The AMD thread ripper is just a CPU. The M1 ultra is an SoC… meaning a CPU, GPU, ISP, NPU, Media encoders/decoders, Secure Enclave, etc. The CPU represents a relatively small part of the M1 Ultra transistor budget.

 

[ungo]

He ran a site called "Geek Patrol" for a number of years that was focused on Macs, though I don't think it's been online for while. You can still find it at archive.org, though newer entries appear to just be from squatting sites after they either sold the domain or let it expire.

Interesting, I looked around a little. I don't see an exclusively Mac site, though - the first random place in the timeline I clicked had posts about Sun hardware and Linux. The first post does say it's launching as a Mac focused site, but it seems like that focus got blurry quick.

Right off the bat, I can see they are compiling the benchmark software using the native Xcode compiler for MacOS and iOS, developed by Apple, while for the other platforms they use the Clang compiler... which was also originally developed by Apple, even if it has been maintained by the open source community for a while.

The Xcode compiler is an Apple distribution of Clang. They never ship exactly the standard open source release, but AFAIK the differences have nothing to do with performance optimization, it's more like Apple wanting to move slower on random things (e.g. one complaint I saw when searching is that a recent Apple Clang couldn't compile something coded with C++ Concepts, a fairly new C++20 feature, but mainstream Clang could).

If your objective is to measure CPU performance with as few confounding influences as possible, using approximately the same compiler for all platforms is usually a good thing.

And I would describe these tests as "semi-synthetic". They do utilize some open-source routines to perform real-world tasks, but not necessarily in ways that will be entirely representative of how actual software will handle the same routines. For that, one would need to run tests on actual, complete software, rather than simply running a quick test that spits out a single number derived from an arbitrary collection of tests.

To me, "synthetic" means something like Dhrystone. Which I hope everyone can agree is an absolutely awful benchmark.

For what it's worth, I'm pretty sure Poole's overall inspiration is SPECint/SPECfp (which fits with his background as a Sun guy), with the biggest philosophical difference being that instead of summarizing with separate scores for int and FP, he generates a single number which does a weighted blend of the two.

Much like SPEC, the individual benchmarks are slices of real world application code, packaged up with reference data to run against, and everything's structured so that other factors like disk I/O shouldn't pollute the scores.

Unlike SPEC, Poole designed it to run very quickly, and (in the free version) to only evaluate burst performance. It's explicitly a measure of the performance potential of a core, not its thermally limited long term behavior. Individual tests take approximately 1 second of wall clock time to run, and he inserts cooldown idle time between each one.

That behavior is actually friendlier to Intel than it is to Apple Silicon: in all models other than the M1 Air, M1 chips don't have short term burst to any significant extent. Intel, on the other hand, is quite aggressive with short term burst - look up Intel Thermal Velocity Boost.

(Expanding on the M1 Air: it's different from all other M1 Macs because it has no fan. The only knob available to keep SoC temperature controlled is to reduce CPU/GPU clock speed. On the CPU side, an Air running a long duration all-cores load seems to settle down at about 2/3 of the short term performance.)

 

[SpeedPc13YearsOldBuilder]

Is this really that impactful? Or just a benchmark.

It kinda depend on the apps that they used to benchmark, Apple's first party app perform like a beast while for 3rd party app? I don't think it will be that good

 

[PiranhaTech]

Something about Geekbench 5 scoring is really strange. The Core i9 12900k soundly beats the M1 Ultra in single threaded performance, but it loses by a wide margin in the multithreaded benchmark. M1 Ultra only has a few more cores than the 12900K. It's strange that it's multicore performance is so strong.

I ran Geekbench on my PC, and the benchmarks seem to be running loads that could target the SoC parts of the chip, not just the main CPU. There's a bunch of custom image processing chips on the SoC. It's a quite clever design, but yeah, can make skewed benchmarks.

This would be similar to benchmarking a PC with a newer discreet GPU in it vs rendering graphics on the CPU that doesn't have an integrated GPU.

There's also a chance that things like the on-die RAM and shared RAM between all of the SoC parts (CPU and GPU included) could favor certain loads on the SoC as well.

 

[hotaru.hino]

If anyone's curious, I posted a reply on the other M1 Ultra thread regarding some benchmark scores reviewers have posted: https://forums.tomshardware.com/thr...-cpu-64-core-gpu.3753325/page-2#post-22640464

The tl;dr is the M1 Ultra does almost match the Threadripper 3990X in Geekbench, but it loses by a wide margin (about 3x) in Cinebench R23.

 

[spongiemaster]

If anyone's curious, I posted a reply on the other M1 Ultra thread regarding some benchmark scores reviewers have posted: https://forums.tomshardware.com/thr...-cpu-64-core-gpu.3753325/page-2#post-22640464

The tl;dr is the M1 Ultra does almost match the Threadripper 3990X in Geekbench, but it loses by a wide margin (about 3x) in Cinebench R23.

It also doesn't beat a 3090 either.

Despite Apple's claims, M1 Ultra GPU is not as powerful as RTX 3090 - VideoCardz.com

 

[mrsense]

Only one question is needed. Can you play popular games on M1 system? I don’t think so.
If you want to compare a system to x86_64 systems make sure it can do that first.
Nobody “plays” Geekbench, except the geeks at Geekbench. If you are knowledgeable with computing world, you know Geekbench scores do not represent the actual system performance. It was only created to look Apple products look good.

 

[hotaru.hino]

Only one question is needed. Can you play popular games on M1 system? I don’t think so.
If you want to compare a system to x86_64 systems make sure it can do that first.
Nobody “plays” Geekbench, except the geeks at Geekbench. If you are knowledgeable with computing world, you know Geekbench scores do not represent the actual system performance. It was only created to look Apple products look good.

While I agree there's something iffy about Geekbench since it seems to be the only benchmark we see used to compare Apple with other systems, whether or not any Apple system can play games on it shouldn't really be relevant. Especially in the case of the Mac Studio: it's not intended to be sold as a gaming machine, so gaming benchmarks shouldn't have as much importance.

In a similar vein I don't expect reviews of Epyc or Xeon processors to have gaming benchmarks because the people buy them don't tend to be gamers.

 

[cryoburner]

...whether or not any Apple system can play games on it shouldn't really be relevant. Especially in the case of the Mac Studio: it's not intended to be sold as a gaming machine, so gaming benchmarks shouldn't have as much importance.

The gaming comparisons are likely a result of that nonsensical marketing slide from Apple showing the M1 Ultra supposedly being faster than a 3090 while drawing a fraction of the power, but Apple being incredibly vague about what sort of workload they were performing on it. So, naturally all the clickbait headlines tried to suggest the chip was going to be a "3090-killer", when it doesn't appear to perform anywhere remotely close in games, and even tends to be well behind in GPU compute workloads.

 

[gg83]

Only one question is needed. Can you play popular games on M1 system? I don’t think so.
If you want to compare a system to x86_64 systems make sure it can do that first.
Nobody “plays” Geekbench, except the geeks at Geekbench. If you are knowledgeable with computing world, you know Geekbench scores do not represent the actual system performance. It was only created to look Apple products look good.

Those geeks at geekbench geeking on benches.

 

[Historical Fidelity]

What makes me always excited about the apple M1 series is they always give full performance even if we're not plugged in power, not like the windows gaming notebook, which has significant performance between plugged in the power and using battery only

False….

 

[hotaru.hino]

False….

A cursory glance of what benchmarks I could find with the Macbook Pro M1 on battery compared to plugged in showed it's only 90% as powerful (https://eclecticlight.co/2021/12/08/do-m1-pro-and-max-cpus-run-slower-on-battery/). As a comparison with an AMD laptop, performance on battery can drop to ~67% of full performance (

View: https://youtu.be/0ISM_TxliRI?t=613
) So not quite "full performance", but Apple's M1 definitely doesn't lose as much as an x86 laptop.

Although if you do enable Lower Power Mode on M1 Macbooks, there is a much greater drop in performance (https://medium.com/macoclock/low-power-mode-on-macs-6ca03e402bf6) but I think this was meant to be used if you're down to your last quarter of the battery.