Apple M1 Ultra - 2nd highest multicore score, lost to 64-core AMD Threadripper.

[gjsman]

3 minutes ago, AluminiumTech said:

If you take Apple's word for it, the M1 Ultra graphics is 80% faster than a W6900X which is a 6900XT and thus the 3090.

That would be for the 64-GPU-core version though, which is an extra $1000. Expensive, but considering that you need two perfectly-binned M1 Max chips... understandable from Apple's perspective even if hard to stomach. 

 

It will be interesting to see how the 48-GPU-core $4000 Mac Studio performs. I am personally expecting the 32-GPU-core M1 Max to perform higher than we saw it in the MacBook Pros. 

[HRD]

1 minute ago, gjsman said:

I actually think it may be at least believable. Consider when AnandTech did their tests:

 

 

 

Take the M1 Max score, increase it a bit because the M1 Max at the time was tested inside of a laptop chassis without desktop-level cooling, and then double it. 3090 territory? 

 

And this is not even running natively. 

The Pathless (Apple Arcade) was running under rosetta and got recently updated to be a native ARM game and better metal support 

the performance gains are huge

at 1080p the fps went from 45-60 to locked 120fps after the update on the m1 max.

 

 

[gjsman]

IKR? It's much more powerful than the games would let on. Most of the games for testing purposes are actually running under two layers of transaction. The first being Rosetta 2 (x86 -> ARM), but the second being MoltenVK (Vulkan -> Metal).

 

A game that was natively ARM and Metal would, potentially, run screaming fast. Considering Apple has been significantly less affected by scalpers and delays and price increases than everybody else (even if it may take them several weeks to ship), it could be a great gaming solution if game developers gave in. But they probably won't unless Apple entices them with some deals.

[HRD]

5 minutes ago, AluminiumTech said:

If you take Apple's word for it, the M1 Ultra graphics is 80% faster than a W6900X which is a 6900XT and thus the 3090.

Apple said up to 80% faster and I think it is true for some tasks for example "Apple's M1 Max clearly has an excellent media playback engine that outperforms not only standalone mobile GPUs, but even Nvidia's top-of-the-range GeForce RTX 3090." m1 ultra has double the media engine. 

 

source:

https://www.tomshardware.com/news/apple-m1-max-benchmarked-in-adobe-premier-pro

[leadeater]

64 Zen 3 cores clocked lower than TR, EPYC 7763, does around ~40,000 for those interested. ST EPYC3 is not great, ~1300.

[gjsman]

1 minute ago, leadeater said:

64 Zen 3 cores clocked lower than TR, EPYC 7763, does around ~40,000 for those interested. ST EPYC3 is not great, ~1300.

It also costs more than $5,000 for the chip alone if you go the eBay route, with an MSRP of over $10,000. No onboard graphics included. Is it really a great option?

[Paul Thexton]

15 minutes ago, gjsman said:

It also costs more than $5,000 for the chip alone if you go the eBay route, with an MSRP of over $10,000. No onboard graphics included. Is it really a great option?

If you *need* the raw compute power for what you’re creating, to then make money as a business, $10k probably isn’t an issue in all honesty.

 

I don’t personally have any experience of needing that kind of core/thread count other than to use it for OS virtualisation so I’m not sure what you’d be doing that would need it, I do have a friend who’s an Indy game dev who complains of long compile times using Unreal engine with his thread ripper (no idea what gen / sku he has) so I guess maybe that would be an example work load?

 

 

[Kisai]

17 minutes ago, gjsman said:

It also costs more than $5,000 for the chip alone if you go the eBay route, with an MSRP of over $10,000. No onboard graphics included. Is it really a great option?

No, not really.

 

Basically if RTX 3080's were still available for $700, and TR's were somehow priced affordably, that's the only way it would make this Mac look like a bad value.

 

Now personally, I feel that the Mac Studio is a great value all things considered. I hate the lack of onboard storage, which means you have to buy more than than you would need just so there's enough to last the 8-year life of the machine, where as a more typical desktop you could replace a smaller SSD after 2 years if it's worn out. All I'm asking from Apple is make the drive replaceable without having to destroy the machine. I could live with not replacing the RAM if it comes with enough.

 

[Forbidden Wafer]

1 hour ago, AluminiumTech said:

If you take Apple's word for it, the M1 Ultra graphics is 80% faster than a W6900X which is a 6900XT and thus the 3090.

That is their problem. Too much snake oil even when they hold a nice advantage or are super competitive.

[gjsman]

Right - Apple's answer to the EPYC is going to be in their actually-Pro lineup. The EPYC is not competitive value-wise with the Studio. It's like asking who needs or should buy a MacBook Air when you can have a MacBook Pro 16" fully-decked-out.

 

5 minutes ago, Kisai said:

Now personally, I feel that the Mac Studio is a great value all things considered. I hate the lack of onboard storage, which means you have to buy more than than you would need just so there's enough to last the 8-year life of the machine, where as a more typical desktop you could replace a smaller SSD after 2 years if it's worn out. All I'm asking from Apple is make the drive replaceable without having to destroy the machine. I could live with not replacing the RAM if it comes with enough.

It's weird because Apple highlighted in the announcement that the (up to 6) Thunderbolt ports is fantastic for external storage. I got a strong "you should use external storage" vibe from both the pricing and the announcement.

 

I wonder if this is one of Apple's clever "faux Pro" tricks. For example, the $1000 monitor stand. All actual pro users have VESA mounts and multiple screens, so they aren't offended by it because it never interested them. The only people buying it are "faux Pros" and to those that can afford it, the horror of the price tag is a perk. Similarly for storage, I speculate all actual Pros are using network storage or external storage (thus the storage prices do not affect them), while "faux Pros" load up on the storage.

[Arika]

Until Apple start selling their chips for non-apple OEMs to use, the numbers are worthless to me.

It's impressive, don't get me wrong, but also pointless since i'm never going to buy an Apple computer.

 

edit: well maybe not never, depends on if MS can get windows running on it through offical means (not like a hackentosh) and actually gets all the performance out of it.

[Paul Thexton]

1 minute ago, Kisai said:

Now personally, I feel that the Mac Studio is a great value all things considered.

If I were considering buying one then I think I’d be tempted by the Max rather than the Ultra personally. £2000 plus maybe another £500 or so for upgrade options seems reasonable to me, but really that’s simply so I could use it for large dev projects. 
 

I think any general consumers would be better off with a M1 Mac mini, sure it’s not going to play any Mac ported games as well, but in all honesty if playing games is a core requirement of any PC purchase then I would recommend just buying a Windows machine within budget instead.

8 minutes ago, Kisai said:

I hate the lack of onboard storage, which means you have to buy more than than you would need just so there's enough to last the 8-year life of the machine

Yep. Soldered ram or not, an unpopulated m.2 slot for additional storage is top of my wish list for Macs.

 

[gjsman]

16 minutes ago, Arika S said:

edit: well maybe not never, depends on if MS can get windows running on it through offical means (not like a hackentosh) and actually gets all the performance out of it.

An Apple Executive said when Apple Silicon was announced that they were A-OK with Microsoft bringing Windows to Apple Silicon, but Microsoft didn't want to. 

 

Over a year later (last November), it was confirmed what I had (privately) long-suspected, that Microsoft had a Windows on ARM exclusivity deal with Qualcomm. As of November it was "expiring soon" but we don't know exactly when. I suspect that it was for five years (considering we got first Windows on ARM in what, 2016?) and expired at the beginning of 2022 but I have no proof of either speculations. 

[LAwLz]

4 hours ago, jaslion said:

They test a limited amount of cpu benchmarks. Things even basic cpu's can excel at. Which is not representative of real world performance unless those simple calculations are the only thing ever happening. Then yeah this thing smashes but that is such an extreme outlier of a usecase it's not representative.

 

I mean basically every benchmark is a best case scenario. Load up a real world task and boom usually entirely different story. i mean the m1 is a perfect example. Absolutly dominates in geekbench but load up video editing, gaming, rendering,... and there you go a entirely different story. Cpu's it obliterated now do the opposite.

What are you on about? The M1 does just as well in things like CPU-based video encoding, rendering and compiling as it does in geekbench. 

 

4 hours ago, Eigenvektor said:

Geekbech is just a single benchmark, so it doesn't necessarily represent all possible workloads. I've also read several times that Geekbench seems to favor macOS and the performance difference you see isn't always mirrored by other benchmarks.

 

Not trying to dis M1 though. I have an M1 Pro in my MBP at work and it is a nice little machine

Geekbench is not a single benchmark... 

 

 

 

What is it with people who doesn't understand geekbench trying to discredit it? It's infuriating. Geekbench is a fine benchmark. It's more than fine. I'd say it's very good for testing CPUs. It's basically SPEC but with slightly different tests and a much smaller dataset. The smaller dataset means it doesn't stress memory as much, but in the case of the M1 that is actually a drawback since memory is one of the things the M1 excels at. 

[Paul Thexton]

I don’t even know why this suddenly came to mind, but if you’ll forgive the analogy, here’s canned generic cpu benchmarks explained in the context of political opinion surveys from a (very) old BBC comedy

 

 

ask different questions, get different answers

[leadeater]

1 hour ago, gjsman said:

It also costs more than $5,000 for the chip alone if you go the eBay route, with an MSRP of over $10,000. No onboard graphics included. Is it really a great option?

That wasn't the point, the point was when you can buy a Zen3 TR it'll be that performance or a slight bit higher. The current comparison TR is not using current Zen3 arch and nobody should be buying a new TR right now with that in mind. A 5950X would quite often do just as well or better for a lot of things, not all things of course but it's rather nonsensical to be buying those right now, or anything for that matter lol.

 

I stated the reference point, the 7763, so people can know where the 64 Zen3 cores currently exist. Because 64 Zen3 is not a theoretical, it's actually a thing so worth looking at.

 

Edit:

At least you can buy a Mac for the sticker price, hell at least you can buy a Mac at all.

[PeachGr]

5 hours ago, Spindel said:

Please explain how GB is not a indicatoon of real world performance. 
 

Enlighten me please.

A CPU is working differently in different scenarios, and this is why on CPU benchmarks you see different scenarios benefit one, or the other CPU. If it was one to rule them all, them 100% of people would but it (as it happened before ryzen era). Examples are code compilation, gaming, CPU rendering, updates per second, simulations etc.

[Alex Atkin UK]

3 hours ago, gjsman said:

I actually think it may be at least believable. Consider when AnandTech did their tests:

 

 

 

Take the M1 Max score, increase it a bit because the M1 Max at the time was tested inside of a laptop chassis without desktop-level cooling, and then double it. 3090 territory? 

 

Then Zen Pinball Party must be REALLY poorly written as it hammers the M1 Pro.  In comparison Pinball FX3, which looks a whole lot better, ran like a breeze at 4K 60Hz on an RTX 2080.  I mean its possible in rewriting the game for a new engine (I believe its Unreal, like the new Pinball FX will be) sapped performance, but it just goes to show how complicated comparisons are.

Just like a CPU, a GPU has specific functions.  It absolutely might match or beat a 3090 in some operations, but might absolutely suck in others.  You can't simply say its better or worse, it depends entirely on the workload.

 

I mean look at Switch, you can port a lot of games there but they aren't running at remotely comparable graphics settings.  Are we sure games like Tomb Raider have all the same effects engaged on Mac?  I was under the impression DX12 translation was impossible as Metal doesn't have a comparable API, so no idea how Vulkan is doing it.

We just don't have a big enough supply of compatible titles to make a good comparison, at least in gaming.  We actually have a better chance in Linux if they get the GPU fully working, of course we'd need the x86 compatibility layer too which probably wont happen.  Rosetta 2 doesn't actually support the entire x86_64 instruction set even on MacOS.

[hishnash]

1 hour ago, Alex Atkin UK said:

  I was under the impression DX12 translation was impossible as Metal doesn't have a comparable API, so no idea how Vulkan is doing it.

So the main issue is the GPU phsycily does not support some of the features DX expects, there are other things these GPUs support that DX (and its GPU) do not support.  It would be just as hard (if not harder) to create a compatibility layer to run metal shaders on DX/VK, I would in fact say a lot harder due to some of the memory and other support (being able to issue draw calls from compute shaders without calling back through the cpu). But the biggest issue would be stuff like tile compute shaders and having the output of them be useable in the following fragment shaders, doing this on a non TBDR gpu will be almost impossible without massive perf impact of putting every draw call separated by a compute shader into its own render pass. 

[Dracarris]

6 hours ago, MageTank said:

It's not just CPU frequencies either. Memory performance plays a huge part in this and I can adjust both my operating frequency and timings (primary, secondary, tertiary) and dramatically inflate my scores despite it looking like JEDEC speeds because GB decided to query the WMIC MemoryChip Get ConfiguredClockSpeed, instead of WMIC MemoryChip Get Speed. Even then, they don't know your timings either, and we all know at this point that latency is king.

well okay that means you cannot trust speeds and timing that GB reports - and how exactly affects this the score? After all the score tells us in which time machine x gets workload y done. At which clock speeds and memory timings is pretty secondary, especially with RISC machines where e.g., CPU clock is anyways not comparable to CISC. Here we are interested in how Apple Silicon compares to x86 machines, not in scores on two identical CPU models.

[Kisai]

1 hour ago, Dracarris said:

well okay that means you cannot trust speeds and timing that GB reports - and how exactly affects this the score? 

I don't think it does. It's not like GB is using those values to calibrate the test. The calibration value is literately Intel Core i3-8100 = 1000, which every single Apple Silicon in the A12 or later exceeds.

In single threading, All the M1's get the same 17xx value, and the last Intel models are 12xx. Not a single Android device hits 1000 (no device in the database beats a i3-8100), the only thing with a score higher than 1000 is an ARM Windows laptop. The Intel CPU's alone, the i9-12900k is 1997, or almost exactly double the i3-8100.

 

You never look at the multi-core scores for comparative benchmark purposes. Because a single-core tells you what the "maximum" performance of a cpu core is, the multicore may clock down instead, and usually the highest multi-core scores are also the worst single-core performers. For example, the highest performing multi-core scores on the mac's are also a lot closer to 1000 on the single core.

 

Overall, GB is a good way to get a like-for-like score, because the GB software is available on mobile devices, where as there are few benchmarks that are available for all platforms.

[Spindel]

14 hours ago, jaslion said:

They test a limited amount of cpu benchmarks. Things even basic cpu's can excel at. Which is not representative of real world performance unless those simple calculations are the only thing ever happening. Then yeah this thing smashes but that is such an extreme outlier of a usecase it's not representative.

 

I mean basically every benchmark is a best case scenario. Load up a real world task and boom usually entirely different story. i mean the m1 is a perfect example. Absolutly dominates in geekbench but load up video editing, gaming, rendering,... and there you go a entirely different story. Cpu's it obliterated now do the opposite.

 

14 hours ago, Eigenvektor said:

Geekbech is just a single benchmark, so it doesn't necessarily represent all possible workloads. I've also read several times that Geekbench seems to favor macOS and the performance difference you see isn't always mirrored by other benchmarks.

 

Not trying to dis M1 though. I have an M1 Pro in my MBP at work and it is a nice little machine

 

14 hours ago, MageTank said:

Worry not, Past MageTank happens to be an expert on this subject:

 

He even cites sources, runs the tests and compares with others as well. I like that guy.

 

For those uninterested in clicking through the ramblings of a madman, the short of it is this: Geekbench grabs DMI/WMI information and reports it instead of probing actual clock speeds. Any clock speed listed in test results are impossible to use for comparison sake because there is simply no way to verify if they are in fact running at that clock speed. It would be like querying WMIC MemoryChip and confusing Configured Clock Speed with "Speed". Both report different values and grabbing the wrong one results in inaccuracy.

 

I can configure two identical systems, run Geekbench on both, and you'll never know I modified one system over the other. The listed information will be identical. I've done this in the past, it's just as easy to do it now.

 

11 hours ago, TheSage79 said:

Long Story short, General Compute Chips really can't be benchmarked for every single common scenario with one benchmark. Things like Geek Bench or Cinebench do the best they can, but there is so many different scenarios out there that it is just not feasible for a benchmark to be accurate beyond a certain degree. Geekbench, in particular, is geared more for mobile devices than desktops. Thats why when LTT does videos on various processors, or GPUS, they don't just show Cinebench scores, but show its performance across a multitude of applications (usually games). The other thing to keep in mind is that the M1 series chips are SoC's which is unlike most of anything in the PC world; so typical PC benchmarking doesn't necessarily translate well. If you want to know real world performance, you need to go out there and well... see the real world performance.

 

Personally, as someone who owns a 13" M1 MacBook Pro and has seen how well the M1 performs in real life, I can say the the M1 SoC is more than powerful enough for the vast majority of people. The M1 Pro, Max, or Ultra, are far more powerful than most people need and are much better suited for truly heavy hitting applications. 

 

9 hours ago, PeachGr said:

A CPU is working differently in different scenarios, and this is why on CPU benchmarks you see different scenarios benefit one, or the other CPU. If it was one to rule them all, them 100% of people would but it (as it happened before ryzen era). Examples are code compilation, gaming, CPU rendering, updates per second, simulations etc.

None of these answers my original question. 

 

While I agree that a benchmark is a benchmark and it might not always exactly reflect on your specific workload the fact is still that with in example GB you are comparing two identical workload mixes between two systems and thus it's indicative of how the relative performance between the systems is. Thus GB has validity, but of course it's not the end all be all answer. 

 

GB also has the advantage that it actually test a bunch of different workloads and weights the result unlike, in example, Cinebench that is only one specific work load that in my case is not at all relevant to how I use a computer. 

 

[Video Beagle]

20 minutes ago, Spindel said:

None of these answers my original question. 

well the real answer is that it tests show the M1 beating whatever processor they fanboy for (for whatever reason) so it's a "bad test".

[Spindel]

5 minutes ago, Video Beagle said:

well the real answer is that it tests show the M1 beating whatever processor they fanboy for (for whatever reason) so it's a "bad test".

I could have been mean and asked for an example of a CPU that scores high in GB but performs poorly in real life (other than when having a too small cooler). 

[leadeater]

1 hour ago, Spindel said:

While I agree that a benchmark is a benchmark and it might not always exactly reflect on your specific workload the fact is still that with in example GB you are comparing two identical workload mixes between two systems and thus it's indicative of how the relative performance between the systems is. Thus GB has validity, but of course it's not the end all be all answer. 

Benchmarks really only lose their validity when the workload, or working data set, is too dissimilar to actual real world applications doing the same instructions and tasks. The test path configuration i.e. the instruction being test would be absolutely valid however if the working data is nothing like actuality then it's not a representative test. Unless of course this is being done on purpose to show the difference in workload sizes and cache effects etc, stuff like that. There are valid times to run things in a more unrealistic way.

 

I think a lot of the criticisms people have/say about GB, whether they realize it or are just parroting, have to do with with GB stemming from a mobile benchmark and the data workload being light in comparison to many desktop application workloads. The danger there is attaining much higher thread performance from utilizing a lower level much faster cache than otherwise would be the case. Whether this hold true in GB5 or even when it's specifically run on desktop/laptop class CPUs I don't know. GB5 is not GB4-3 etc.

 

I personally much prefer SPEC as that is certainly more representative of desktop/workstation workloads and datasets but it takes forever to fully run and not everything can even run it either.

 

Basically for example two XTS-AES crypto benchmarks can result in very different results with different relative CPU to CPU performance. Stuff like that.

 

Translating benchmarks to application performance is not an easy or simple task because actual applications don't just do one thing, at one time, in one way. This is not a failing of GB5 this is a failing of everything like it including SPEC. But that's not why you run them.