Of course - their single chip M1 Max GPU performance got ~30-40% of high end desktop GPU, so it sounded impossible that a dual-GPU solution could do 100% of high end desktop GPU. When Apple omits actual direct comparison it sounds and looks so bad for them. Especially on these fake graphs! If they are confident in their numbers, use numbers. So silly.
It's a test of raw GPU performance. Testing bandwidth would be difficult to synthesize, and also irrelevant for the vast majority of "real world" workloads besides maybe GPGPU computations which are almost always always written with CUDA anyways.
There really isn't all that much software that does the same exact thing on both NVIDIA and Apple hardware, so benchmarks will largely favor the hardware that has had more integration.
Unlike AMD & Intel, which are easily directly compared, I think there'll be many big differences between Apple and NVIDIA. They do many things in completely different ways.
It'll be interesting to see more comparisons as more software is adapted and written to fully take advantage of Apple hardware, the way much of it is currently written to fully take advantage of NVIDIA.
It’s fair if “better” means “better for end-users”, which is all that matters in practice. It’s especially fair given that Apple is generally credited for having a capabilities focus rather than spec focus, i.e., they are credited for the same end-user focus.
I think the article is poorer for not explicitly making this argument, but it’s right to focus on end-user capabilities, IMO.
Better for end users depends entirely on their workloads, and is a moving target. Presumably as Apple’s new chip gets more adoption more games will be designed and optimized for it. There are already games (partly) optimized for it, so why not run those? Why is Tomb Raider more representative of an end-user workload than World of Warcraft?
I agree it's a moving target, but the point is that the current capabilities matter most, potential future capabilities matter less, and spec sheet or artificially constrained benchmark comparisons matter least. I'm bullish on Apple Silicon, but it is far more likely for your average gamer or GPU compute coder to think that the M1 Ultra is not better than an RTX 3090. It's not being "unfair" if you don't highlight a game that has been specifically optimized, if such games are few and far between, and it is far from assured that that will change. A really good comparison would highlight those cases where AS-specific optimization does improve things significantly, but even there, it should really emphasize that the comparison of M1 Ultra to an RTX 3090 is only favorable given an artificial power constraint on the latter (Apple mention it rather than emphasize it, but a good journalist will emphasize it).
So is AS more technically impressive, and does it have more growth potential in the hardware and software (as adoption hopefully grows)? Absolutely. Is it currently materially "better"? For a desktop, by almost any measure, no.
This article is responding to a very specific slide that Apple showed claiming they were faster than the fastest GPU. This article demonstrates that along a wide range of expected functions for a GPU, Apple's slide was misleading (in multiple ways). While it is definitely true that Apple's new system is a hog for multimedia problems, it's not a "3d-GPU-rendering" system, it's a "2d-GPU-rendering" system, and that's a secondary market for GPUs that nvidia didn't optimize for.
Further, I'm fairly certain if nvidia and adobe did optimize more for GPUs, that their products would be significantly faster. However, at that point we start to get into some very interesting issues about whether the GPU should be a bus-attached, or core-attached device.
>a very specific slide that Apple showed claiming they were faster than the fastest GPU
That is not what the slide showed. The slide shows Apple's chip reaching the same performance numbers given the same amount of power delivery on each GPU. The Verge, this PC Gamer article, and you, are conveniently missing that clearly presented metric.
Yeah, Apple's slide seems carefully cropped to show discrete GPU limited going to 320W, so the M1 comes on top on a performance vs. power ratio. But the 3090 Ti has TBP of 450W [1], which likely explains why its raw performance easily beats the M1.
Note also "GPU performance" is a metric that includes many things. What did they test, specifically, here? ML training? Gaming? Or 2D ops used commonly in photoshop?
I found this review useful: https://www.theverge.com/22981815/apple-mac-studio-m1-ultra-...
TL;DR: many professionals who work in 2D or psuedo3D could buy this and immediately be more productive than they were on their previous Apple device. But that chart is still extremely misleading.
that's kind of my beef w/ the 3000 series gpu's in general.
Yeah, they're fast and all. But at significantly more TDP and with default coolers larger (> 300 mm) that can fit into my (previously perfectly fine "full size GPU") case? Are they really that much better than my 1080?
I'm aware of that (it was covered already in the press), and it's clear what the slide intended to show. This is standard apple reality distortion field. I work in the field of high performance computing and we know how to interpret a graph that is intended to deceive the naive.
I'm of mixed minds...my PC is useful for certain things. But my Apple laptops w/ iGPU's did not require as much hardware debugging, sketchy firmware downloads/updates, less fans to clean, large PSUs that blew and had to be replaced etc...
I agree, my patience also wears thin getting 'standard' parts to work together when they don't want to.
On the other hand, how far can you trust proprietary all-in-one solutions? At least with the modular stuff you have some control over the system, and a choice in selecting a preferred configuration - the all-in-one is a black box.
presumably, a lot of technical info has been coming out of Apple (and hopefully like minded companies) re vetting of chips, mobo, OS-built security, T2 chip encryption etc.
Vs. the modular stuff - I'd be more confident if security was part of the norm, but I don't get the feeling it's a priority on the like of a lot of 3rd party vendors like Asus or Gigabyte etc.
Look at the chart. The question isn't about "max power" performance as much as it is power consumption: Apple claims its chip is reaching similar performance numbers to its competitor given the same amount of power. In fact, the Verge's article[0] that this PC Gamer article is shamelessly and effortlessly reposting, does not dispute that, but is moving the goalpost to "well what if the competitor had more power?" Yawn
The M1-era is a very difficult time for the crowd who just wants to criticize Apple for the sake of it. I use both an M1 and an M1 Max every day alongside Threadrippers and Xeons and while Apple misses a lot, they fucking killed it this time.
The M1s are my desktop and laptop respectively, the Xeons are for deployment scenarios, and the Threadripper box is used mostly for big builds.
I don't have my current project building on `darwin-aarch64` yet so I haven't been able to do a comparison, but I suspect that once I can get my hands on an M1 Ultra that it will build a big C++/Haskell application faster than the 64-core Threadripper does.
800GB/s of memory bandwidth is insane. You write programs differently for a computer where your main memory is effectively an on-die L3 cache at 128GB.
Oh I don't think I'd go so far as to say building big C++ (or Rust, or Haskell, or ...) projects requires a ton of compute: it's just one way to tune the inner loop of your development process, the so-called "edit-compile-test" cycle.
I've worked on projects that take hours to build from scratch and tens of minutes to do incremental builds on. When making a non-trivial modification and running the tests is 14 minutes, it's just human nature to end up with a 30-minute coffee-trip/HN-browse.
When it's 30 minutes, you only get 16 of those in a standard work day. Getting anything meaningful done takes days or weeks when you only get to change your code 16 times a day. You can change Python code 16 times an hour if the changes are individually modest.
It’s definitely the crowd showing why apple will never target the larger gamer market. Apple would have to target a market audience that is pretty opposite what apple tech is about. As one YouTuber said it “gamers are high touch, etc”.
Though I do have to fault apple for even bothering to compare this SoC to the 3090. They were always going to get very strong animus from people who wouldn’t bother to remember the TDP qualifier “up to 200w”.
If they really want to compare themselves to the 3090 then they need to make an SoC whose TDP approaches the TDP of a comparable CPU+GPU at low and high ends. Ie make the chip much bigger.
On Apple's graph there is a clear line showing that their performance is better, irrespective of power.
If Apple wanted to play in the "perf for watts" space, then they should have made sure to keep their claims within that space. Everyone is playing to the goalpost that Apple established.
Yeah, the arguments sounds good until you go and look at the actual chart from Apple.
The 3090 can apparently draw around 380 Watts. [1] Apple shows its performance leveling off at 320W, whereas the M1 Ultra is performing better at 100W, and not leveling off. But assume for a second that the M1 Ultra suddenly flatlines right there. Fair enough, they don't want to show that. But in what world does that 3090 curve ever end up substantially out-performing the M1 Ultra with just 60 more Watts?
The graph is clearly wrong.
Unless you change the argument to saying that they are measuring something silly that doesn't matter in the real world.
So, if I made a GPU that has 1GFLOP (but can never go above that) and uses 0.000001W, would you jerk me off because "tEcHniCaLlY It'S aBoUt pOwEr CoNsuMpTiOn", even though it's wildly inferior, just because you extrapolate that to GFLOPS if it was using 300W ?
Can you answer the simple question about why the dotted line stretching from the max apple performance is higher than the GPU?Because that's what the chart says (in addition to the power details).
> For the most graphics-intensive needs, like 3D rendering and complex image processing, M1 Ultra has a 64-core GPU — 8x the size of M1 — delivering faster performance than even the highest-end PC GPU available while using 200 fewer watts of power.
This is the exact statement Apple made. Do note that the words "similar performance numbers at the same power" are not in there.
Did anyone say m1 ultra is faster than 3090? If they did, surely they didn't mean games. Productivity is not gaming. Fast overall is not the same as delivering 144+ frames each second.
First, Why post an article that just refers to testing done at The Verge instead of linking to The Verge's own article?
Second, the article at the Verge is based on running an x86 game under Rosetta's x86 emulation instead of testing the performance of a native ARM and Metal game.
Strangely, The Verge's Mac Studio review covers the fact that the Adobe suite of software had to be updated to support ARM and Metal before you could see the full performance of the hardware without emulation, but they completely ignore this for games.
>With many recent professional Mac reviews, the story has been: Apple makes great hardware, but the software — specifically Adobe’s Creative Cloud — isn’t optimized for Apple’s ideas.
But at this point, Adobe has had quite a bit of time to catch up. Apple has sold a whole bunch of M1 devices — the demand is here, and Adobe has responded. Premiere, Photoshop, Lightroom, Audition, Media Encoder, and more are running natively on Apple Silicon (and After Effects is available in beta). And the story has basically flipped: the Mac Studio now appears to be one of the few computers that can unlock the full power of Adobe’s software.
> Strangely, their review covers the fact that the Adobe suite of software had to be updated to support ARM and Metal before you could see the full performance of the hardware without emulation, but they completely ignore this for games.
When do you reckon that Tomb Raider is getting updated for ARM?
Yes because consumers want to know how the hardware performs for their existing catalog of games, many if not all of which won't be ported? Obviously depends on what you play etc
There's nothing wrong with reporting the level of performance you'll get for emulated games too, but the claim being made is that the hardware can't do better than that.
To make the headline accurate, you have to add a disclaimer, going from:
>Apple's new M1 Ultra isn't faster than an RTX 3090
to
>Apple's new M1 Ultra isn't faster than an RTX 3090 when running x86 games under emulation
They also test the native performance of Photoshop's notoriously computationally intensive brushes.
>Ultra also enabled Alex to use features of Photoshop that he’s never been able to use before. The process that blew all of our professionals’ minds was the brushes. Alex was able to paint a white layer with a media brush, and it was instant — it really looked like he was painting with a physical brush. That tool requires so much computing power that many current machines can’t handle it. It was a breeze for the Studio.
“Having a machine like this in my daily process would change everything,” Alex told me after his testing period.
That would have been a much better take to include than the graphics apple shared in the announcement. Probably would have sold more units faster, not that it was needed.
Tomb Raider is only relevant in this discussion because it has a very detailed and fine built-in benchmark system. Not every game does have that (like neither WoW or Baldur's Gate 3)
Tomb Raider loads a pre-rendered scene that is the same every single time you start the benchmark thus you can compare it across systems. That's the point of it and hence became a staple alongside with other games (some games are even notorious about it because that's their only use case [0])
WoW doesn't have anything like that. And everything can affect the frame rate from the mobs in the zone, the weather, the unique players with unique transmog, the raids are different every single time etc.
It's not just loading the game and there you have X framerate and that's the score, that's not a benchmark.
Perhaps, but this discussion would be moot if Apple did the same, “M1 GPU faster that RTX [under specific scenario]”. But they didn’t do that and here we are.
Are you expecting Intel or AMD to announce a new chip and only tell you the level of performance you can expect while running ARM software under emulation?
Is there any evidence that running under Rosetta impacts GPU performance measurably? Shaders are compiled on the fly for a given GPU so draw calls coming in from native vs emulated shouldn't have a big impact unless the CPU is the bottleneck.
GPUs are usually the performance bottleneck at 4k so Rosetta may not even have a measurable impact on framerate for some of the tests. Beyond that, what games are out there that have otherwise identical implementations that run on Windows/x86 and MacOS/Arm?
Why compare games at all? It is a really common use of GPUs so Apple claiming just "faster" rather than "faster in X use case" or "more FP computations per second" opens them up to gaming comparisons.
It's not just Adobe or games, takes 3D rendering for example, both Redshift and Cycle are now using Metal (the Metal backend for Cycle was even submitted by Apple themselves), Nvidia's GPUs absolutely destroy their (supposed) M1 equivalents, with or without using the ray tracing hardware.
In Blenders case, there are still several patches left to merge in and even the Blender page strictly mentions that it's early integration without much performance work done yet.
The technically interesting bit about the Ultra's GPU is that it makes multiple GPU dies look like one physical GPU to software:
>if you could somehow link up multiple GPUs with a ridiculous amount die-to-die bandwidth – enough to replicate their internal bandwidth – then you might just be able to use them together in a single task. This has made combining multiple GPUs in a transparent fashion something of a holy grail of multi-GPU design. It’s a problem that multiple companies have been working on for over a decade, and it would seem that Apple is charting new ground by being the first company to pull it off.
Biggest discussion was on the TomsHardware submission[0] (170pts, 215comments), the verge was also submitted[1], but no real traction (7pts, 2comments).
It's not about PC gaming though, more like the consoles. Apple can absolutely disrupt the market of Sony and MS (not so much of Nintendo). The MSRP of PS5 and Xbox are $399/$499, if Apple can target this segment they would have a field day. The Apple TV already exist and maybe not now but 1-2-3 years down the line they can fit the next generation M1 Ultra there.
BUT you need a very strong game back catalog (see the PS exclusive games and the GamePass on Xbox which is a killer deal). I'm not sure Apple is willing to go that far, just by judging their Apple TV exclusive line up for example compared to Netflix and Amazon. To me it seems they rather have the iOS games on the Mac than have actual AAA games on the Mac. But that might change in the future, who knows.
In the end it's all wishful thinking (or not) but I'm pretty sure competition is only good for the consumer market.
68 comments
[ 3.7 ms ] story [ 131 ms ] threadThere really isn't all that much software that does the same exact thing on both NVIDIA and Apple hardware, so benchmarks will largely favor the hardware that has had more integration.
Unlike AMD & Intel, which are easily directly compared, I think there'll be many big differences between Apple and NVIDIA. They do many things in completely different ways.
It'll be interesting to see more comparisons as more software is adapted and written to fully take advantage of Apple hardware, the way much of it is currently written to fully take advantage of NVIDIA.
I think the article is poorer for not explicitly making this argument, but it’s right to focus on end-user capabilities, IMO.
So is AS more technically impressive, and does it have more growth potential in the hardware and software (as adoption hopefully grows)? Absolutely. Is it currently materially "better"? For a desktop, by almost any measure, no.
Further, I'm fairly certain if nvidia and adobe did optimize more for GPUs, that their products would be significantly faster. However, at that point we start to get into some very interesting issues about whether the GPU should be a bus-attached, or core-attached device.
That is not what the slide showed. The slide shows Apple's chip reaching the same performance numbers given the same amount of power delivery on each GPU. The Verge, this PC Gamer article, and you, are conveniently missing that clearly presented metric.
[1] https://www.hardwaretimes.com/nvidia-made-the-rtx-3090-ti-as...
Note also "GPU performance" is a metric that includes many things. What did they test, specifically, here? ML training? Gaming? Or 2D ops used commonly in photoshop?
I found this review useful: https://www.theverge.com/22981815/apple-mac-studio-m1-ultra-... TL;DR: many professionals who work in 2D or psuedo3D could buy this and immediately be more productive than they were on their previous Apple device. But that chart is still extremely misleading.
Yeah, they're fast and all. But at significantly more TDP and with default coolers larger (> 300 mm) that can fit into my (previously perfectly fine "full size GPU") case? Are they really that much better than my 1080?
Putting CPU, RAM, and GPU all on a single chip, or a multi-chip integrated package clearly has massive advantages.
I think this is another signal that the the era of of the modular computer is over, unfortunately (I love modularity, standards, diy tinkering, etc)
On the other hand, how far can you trust proprietary all-in-one solutions? At least with the modular stuff you have some control over the system, and a choice in selecting a preferred configuration - the all-in-one is a black box.
Vs. the modular stuff - I'd be more confident if security was part of the norm, but I don't get the feeling it's a priority on the like of a lot of 3rd party vendors like Asus or Gigabyte etc.
[0] https://www.theverge.com/2022/3/17/22982915/apple-m1-ultra-r...
I don't have my current project building on `darwin-aarch64` yet so I haven't been able to do a comparison, but I suspect that once I can get my hands on an M1 Ultra that it will build a big C++/Haskell application faster than the 64-core Threadripper does.
800GB/s of memory bandwidth is insane. You write programs differently for a computer where your main memory is effectively an on-die L3 cache at 128GB.
I've worked on projects that take hours to build from scratch and tens of minutes to do incremental builds on. When making a non-trivial modification and running the tests is 14 minutes, it's just human nature to end up with a 30-minute coffee-trip/HN-browse.
When it's 30 minutes, you only get 16 of those in a standard work day. Getting anything meaningful done takes days or weeks when you only get to change your code 16 times a day. You can change Python code 16 times an hour if the changes are individually modest.
Though I do have to fault apple for even bothering to compare this SoC to the 3090. They were always going to get very strong animus from people who wouldn’t bother to remember the TDP qualifier “up to 200w”.
If they really want to compare themselves to the 3090 then they need to make an SoC whose TDP approaches the TDP of a comparable CPU+GPU at low and high ends. Ie make the chip much bigger.
If Apple wanted to play in the "perf for watts" space, then they should have made sure to keep their claims within that space. Everyone is playing to the goalpost that Apple established.
The 3090 can apparently draw around 380 Watts. [1] Apple shows its performance leveling off at 320W, whereas the M1 Ultra is performing better at 100W, and not leveling off. But assume for a second that the M1 Ultra suddenly flatlines right there. Fair enough, they don't want to show that. But in what world does that 3090 curve ever end up substantially out-performing the M1 Ultra with just 60 more Watts?
The graph is clearly wrong.
Unless you change the argument to saying that they are measuring something silly that doesn't matter in the real world.
[1]. https://www.youtube.com/watch?v=xMpNdP3h0Hs
Edit: And I'm an Apple fan, writing on my M1 Air, which I love.
This is the exact statement Apple made. Do note that the words "similar performance numbers at the same power" are not in there.
You are the one moving goalposts here.
Second, the article at the Verge is based on running an x86 game under Rosetta's x86 emulation instead of testing the performance of a native ARM and Metal game.
Strangely, The Verge's Mac Studio review covers the fact that the Adobe suite of software had to be updated to support ARM and Metal before you could see the full performance of the hardware without emulation, but they completely ignore this for games.
>With many recent professional Mac reviews, the story has been: Apple makes great hardware, but the software — specifically Adobe’s Creative Cloud — isn’t optimized for Apple’s ideas.
But at this point, Adobe has had quite a bit of time to catch up. Apple has sold a whole bunch of M1 devices — the demand is here, and Adobe has responded. Premiere, Photoshop, Lightroom, Audition, Media Encoder, and more are running natively on Apple Silicon (and After Effects is available in beta). And the story has basically flipped: the Mac Studio now appears to be one of the few computers that can unlock the full power of Adobe’s software.
https://www.theverge.com/22981815/apple-mac-studio-m1-ultra-...
When do you reckon that Tomb Raider is getting updated for ARM?
To make the headline accurate, you have to add a disclaimer, going from:
>Apple's new M1 Ultra isn't faster than an RTX 3090
to
>Apple's new M1 Ultra isn't faster than an RTX 3090 when running x86 games under emulation
>Apple's new M1 Ultra isn't faster than an RTX 3090 when running x86 games under emulation or when directly comparing Geekbench scores
Actually, I think I see why they went with the shorter title.
They are making it sound like the performance of an emulated x86 game on OpenGL is the only level of performance you can ever expect.
>Ultra also enabled Alex to use features of Photoshop that he’s never been able to use before. The process that blew all of our professionals’ minds was the brushes. Alex was able to paint a white layer with a media brush, and it was instant — it really looked like he was painting with a physical brush. That tool requires so much computing power that many current machines can’t handle it. It was a breeze for the Studio.
“Having a machine like this in my daily process would change everything,” Alex told me after his testing period.
https://www.theverge.com/22981815/apple-mac-studio-m1-ultra-...
Tomb Raider loads a pre-rendered scene that is the same every single time you start the benchmark thus you can compare it across systems. That's the point of it and hence became a staple alongside with other games (some games are even notorious about it because that's their only use case [0])
WoW doesn't have anything like that. And everything can affect the frame rate from the mobs in the zone, the weather, the unique players with unique transmog, the raids are different every single time etc.
It's not just loading the game and there you have X framerate and that's the score, that's not a benchmark.
0, https://www.kotaku.com.au/2016/01/the-game-thats-more-known-...
>Apple's new M1 Ultra isn't faster than an RTX 3090 when running x86 games under emulation
But it's not.
GPUs are usually the performance bottleneck at 4k so Rosetta may not even have a measurable impact on framerate for some of the tests. Beyond that, what games are out there that have otherwise identical implementations that run on Windows/x86 and MacOS/Arm?
Why compare games at all? It is a really common use of GPUs so Apple claiming just "faster" rather than "faster in X use case" or "more FP computations per second" opens them up to gaming comparisons.
https://youtu.be/HF_rMjjPoPs?t=118 https://www.youtube.com/watch?v=iS_uPukPuo4 https://youtu.be/i4CuZqXA4BM?t=359
Of course the 3090 is faster at gaming, which it was designed for.
It's faster at _literally everything else_. So is a 3080. So is a 3070, most likely.
>if you could somehow link up multiple GPUs with a ridiculous amount die-to-die bandwidth – enough to replicate their internal bandwidth – then you might just be able to use them together in a single task. This has made combining multiple GPUs in a transparent fashion something of a holy grail of multi-GPU design. It’s a problem that multiple companies have been working on for over a decade, and it would seem that Apple is charting new ground by being the first company to pull it off.
https://www.anandtech.com/show/17306/apple-announces-m1-ultr...
[0]: https://news.ycombinator.com/item?id=30719654 [1]: https://news.ycombinator.com/item?id=30715759
BUT you need a very strong game back catalog (see the PS exclusive games and the GamePass on Xbox which is a killer deal). I'm not sure Apple is willing to go that far, just by judging their Apple TV exclusive line up for example compared to Netflix and Amazon. To me it seems they rather have the iOS games on the Mac than have actual AAA games on the Mac. But that might change in the future, who knows.
In the end it's all wishful thinking (or not) but I'm pretty sure competition is only good for the consumer market.
Target it with what? A controller from Apple would run $399-$499 and it would be sold separately.
I’m no GPU expert. What does this mean long term?