I’m factoring in the complete SOC here. The M2 ultra performs well relative to AMD and Nvidia while using dramatically less power.
MS leak illustrates new console development cycle
- Thread starter Shifty Geezer
- Start date
In terms of what? AMD has only recently with the 7000 series apu's have moved to 5nm while Apple has been on 5nm since the m1 processor. Apple also says they are on a second generation of 5nm process which would be the n5p process by tsmc. Who knows if AMD or Nvidia even have access to it.
When you look at the amount of chips apple buy's from tsmc both nvidia and amd combined is still behind it. So its obvous they would have a micron process advantage vs everyone else.
The 7800xt is 5nm and uses a good bit more power than the entire Apple SOC while being slower.
On the CPU side it isn’t even close. Apple is just so far ahead of anything Intel and AMD are doing WRT efficiency.
Slower in what? vague statments don't help. People went gaga over the m1 but it was largely forgotten once other companies had access to the same fab process and the same will happen with the m2. However I am sure once again Apple will be on a new generation of micron processes and the same discussions will happen again.
Slower in the currently available benchmarks. M2 ultra ranges between 4070ti and 4080.
Flappy Pannus
Veteran
Do you have game benchmarks which indicate this? The issue with the M2 Ultra in particular is that the multi-die GPU's are very difficult to scale well in games, applications are far less problematic. What are you basing the 7800XT is slower than the M1 Ultra on? Links?
No actual game benchmarks yet, but Apple says the GPU is recognized as a single entity and not multi GPU. Im basing the 7800xt being slower than M2 ultra on the 4070ti being faster than 7800xt.
Last edited:
Flappy Pannus
Veteran
Yes, but that's not my contention. I'm talking about how game performance scales regardless of how transparent the architecture is.
M2 game benchmarks may be rare, but there are certainly game benchmarks for the M1 Ultra. Best case the M2's GPU is 30% faster. I can certainly see for certain applications, especially ones that involve image processing and can really take advantage of the unified architecture, the M1/M2 Ultra can more than hold their own, especially in watts/performance. But as mentioned, those can scale far more readily to multi-die GPU's than games.
I get that comparing game performance between two different platforms is tricky and not necessarily indicative as an overall assessment of the architecture, especially with one of the platforms having a relatively unique architecture/API. Also, being in a potential console, the multi-GPU design could be explicitly targeted by devs to avoid the pitfalls they incur when they have to be abstracted on PC's.
But if we're going to be comparing it in the here and now, and especially to the 7800XT, native Metal and M1 games are the best we can do to make that assessment, and looking at titles like Village and No Mans Sky on the M1 Ultra, I don't see anything to justify the statement that the M2 Ultra is really a 7800xt in gaming performance, let alone superior - at least not now.
Maybe with older titles like World of Warcraft where it seems to fare quite well (going by LTT's look at the M1 Studio), but benchmarks I've seen of modern games that utilize modern rendering features (sans ray tracing (obv not RDNA's strong suit either)) it's not really a 7800XT competitor atm.
Edit: Found a video with games tested on the M2 Ultra. It's hard to compare perfectly as the scenes may differ, but in this scene the settings are at 'high', but no RT - and using Metal FX Quality, which actually renders at 1080p native (according to DF). It's getting ~100fps in this scene. While 7800XT benches in this title are thin on the ground as it's a little old to be included in reviews, a 6800XT averages 100fps, maxxed no RT - at 4K native.
Last edited:
DavidGraham
Veteran
That's not even near the realm of true. When
DF tested RE8 on the desktop M1 Ultra and M1 Max, with max setrings, 4K, no RT, the M1 Ultra is providing performance similar to a 3080m (140w) laptop, and the M1 Max provided half the fps of the M1 Ultra.
An M2 Ultra is maybe 30% faster than M1 Ultra. So at best you are looking at a desktop 4060 performance, and that's being generous.
Flappy Pannus
Veteran
Yes, I addressed that, being a relative immature platform for AA/AAA games doesn't help to accurately evaluate the architectural chops of their GPU. But the fact that the pickings are slim is irrelevant, these are the games we have to go by - you're the one who brought in the performance comparisons, so then give some justification for that. Like, cite...something?
There's also other fully natives games in that video - Layers of Fear (2023) and Baldurs Gate 3 for example. Layers of Fear, 4K high, no RT, in this scene gets 23fps on the M2 Ultra:
On my 3060, same scene, same settings: Exact same performance.
He later turns on Metal FX Quality, which - if it works the same as it does in Village - is the same base res as DLSS performance. Metal FX Quality nets him 39fps, while DLSS performance gives me 55fps in that scene. He only gets into the 55 fps range by turning on Metal FX Performance mode, which again - if it works the same as it does in Village - is just a spatial upscaler at 1080p, which well, sucks. So best-case scenario for this game is 3060 level performance, but likely well below considering reconstruction.
Last edited:
It really doesn't matter because this goes against what you are saying. We are also still comparing different micron nodes which plays a large role in power usage and performance
Frenetic Pony
Veteran
Why on earth is anyone here even questioning Apple's GPUs? They're running late last gen titles on a phone, not just as benchmarks but with long term throttling. The upcoming Nvidia based Switch 2 is going to do the same thing using at least twice and more likely three times the power draw. They've done an amazing job in perf per watt and are easily top of the heap there.
Flappy Pannus
Veteran
We're talking about the M2 Ultra my man, specially how it was supposedly more performant than the 7800XT. The M1/M2 line is far more relevant in hypothesizing about future APU's in consoles than phone soc's. Maybe read the posts before asking why people are talking about something?
Because this is a tech forum wanting to understand the relative performance of different GPUs and SOCs. Why wouldn't we question Apple alongside everything else??
Apple M2 Ultra GPU Outpaces RTX 4070 Ti in Early Compute Benchmarks
M2 Ultra with Metal API versus RTX 4070 Ti and RTX 4080 using OpenCL
Some numbers from apps almost certainly much better tuned to take advantage of the hardware. With directed developer support this chip would almost certainly be performing much better than the game comparisons. Not to mention that in a console, developers could find ways to offload work to all the specialized function blocks present on the SoC.
Flappy Pannus
Veteran
Apple M2 Ultra GPU Outpaces RTX 4070 Ti in Early Compute Benchmarks
M2 Ultra with Metal API versus RTX 4070 Ti and RTX 4080 using OpenCLwww.tomshardware.com
Some numbers from apps almost certainly much better tuned to take advantage of the hardware.
Huzzah, finally some form of a benchmark.
Thing is, I addressed that previously as well. Apps can far more readily take advantage of a multi-chip GPU architecture, if games scaled as well as apps both Nvidia and AMD would have brought MCM to the consumer market by now. There's no contention that the M2 performs great in GPU apps.
The very article you're citing:
Toms Hardware said:
Maybe, maybe not, due to the above, or other architectural quirks that are not so amenable to modern game engines atm. I just think if you're going to tout the gaming prowess of a GPU, you would use evidence of its gaming prowess. Death Stranding is the next big title up, I guess we'll see how it fares? Maybe it will buck the trend so far of being closer to 3060ti performance than 4070ti in games.
Regardless, on 'directed developer support' - do you seriously think Apple hasn't given engineering resources and assistance to Capcom, when they've been on stage touting their games coming to Mac/iOS in several keynotes by now? Sure, Capcom could have fucked it up, lord knows their PC ports can be a little spotty at points (from the DF review, it had massive shader stuttering on Mac OS at launch, but I don't know if recent patches have cleared this up). But Metal as an API is nearly 10 years old now, the M1 debuted 3 years ago. The platform is certainly not weathered, but it's not that virgin.
I'm expecting performance to improve as both developer experience and OS/driver updates iterate, that much is obvious. And it's also obvious on many performance/watt metrics, the M1/M2's are in a class of their own - no one is saying they're not impressive APU's - I like them enough that I have 2, an M1 Mini and a M2 air. But game loads are highly variable in what they stress on a GPU, especially when compared to apps or benchmarking suites. We just don't have the evidence at this point that Apple has demonstrated the expertise to create a high-wattage (relative to what a console would target) APU that is performant on modern gaming workloads. There's a lot of work that goes into designing a chip with a certain power envelope, you can't just hit the shrink/expand button and multiply performance or power savings by the die area.
If they're relevant to games, sure - the PS5 and SX already have many specialized blocks, heck the PS4 had a seperate ARM soc that handled stuff like drm and background downloading.
Perhaps even the multi-GPU architecture could be taken advantage of more readily if it wasn't abstracted like I hypothesized, albeit you probably don't want to make life for developers new to the platform more complicated than it already would be.
BTW, while all of this is interesting, we also have to bear in mind relative to inclusion in consoles, just what the M2 Ultra consists of - it's 132 billion transistors. The PS5 APU by comparison, is 11. Yes, that M2 has 64GB of on-package DDR5 ram with a wide bus that really bumps up those numbers, but that's a necessary part of its architecture to get that excellent bandwidth. The point is though, it's an immensely complex chip that is in no way economically feasible for a $500 console, now or likely several years from now - there's a reason the cheapest platform you can get an M2 Ultra in starts at $4k. We're all just spitballing here of course with the next gen ~4 years out, but really at this point if we're going to bringing the M2 Ultra into the discussion, we might as well be hypothesizing about what a 6090 would look like inside a PS6 too. Neither is really applicable.
BTW#2 - Just remembered Oliver from DF did actually do a dive on the M1 Max/Ultra comparing it to a 3090. Problem is though this was a year ago, so many of the games are not fully native. However IME, Rosetta games suffer very little performance loss when we're in GPU-limited scenarios, the far bigger barrier to performance in older titles are when games are using OpenGL vs Metal. Shadow of the Tomb Raider is one such game, it's Rosetta but uses Metal. Going by these numbers, the M1 Ultra certainly seems to fare quite a bit better relative to its PC counterparts than what we're seeing in Village:
That's putting the M1 more in line with 3070 territory. However, when I ran that benchmark on my own system, same settings:
...Oliver's 3090 results really don't jibe with this, he should be getting in the 80's* with those settings if my 3060 is pulling 47. Maybe a Nvidia driver update in the past year significantly upped performance? I do have an overlock but we're talking maybe ~2fps improvement with it. So another game where the M2 Ultra, assuming perfect scaling, would be performing maybe like a 4060? Better than Village at least I guess, but again, understandably not ideal as not fully native.
(Also of note though is the non-linear scaling you get from doubling GPU resources with the M1 Ultra over the M1 Max, further indicating you can't really take compute benchmarking suite results and translate that to hypothetical game performance.)
*ok I looked around the web - 3090's are getting 90-100+fps at 4K non-RT maxxed in the SOTT benchmark (and that includes HBAO+ which is quite a bit more demanding than BTAO). Those benchmarks were also around the 3090's launch, so it's not a recent driver update. No idea how Oliver got just 65fps here.
Last edited:
DavidGraham
Veteran
Geekbench is not a graphics benchmark, it tests different compute and video/image processing workloads, which are offloaded well into the specific blocks in the M2 silicon. Aztec mobile benchmark doesn't scale well on high end GPUs.
A more contemporary compute heavy workload is Cinebench 2024, here the 4090 is 3x times faster than M2 Ultra, the 3070 is 20% faster.
Here the 3060Ti is considerably faster than M1 Ultra.
Kinda weird with MAC going ARM, that Xbox would embrace the architecture windows doesnt run well on (yet). That would make 2 of 3 consoles at least including Switch. Mobile of course. X86 is getting squeezed.
I'm for it, just for something different. Majority of R&D is probably in mobile space going forward anyway.
As to all this ML, NPU this and that, I'm still not sold. Google made a big deal out of the Tensor chip NPU and Tensore sucks compared to other state of the art Qualcomm chips. I'm just not sold on the alleged AI revolution. MS has a bad habit of fixating on something stupid that ends up deriding the hardware (I believe arguably the 360 EDRAM, definitely the One ESRAM, fit this category). Is this the next example?? Now it's hardly cut and dried, as a gresat upscaling solution is basically equivalent to a more powerful GPU. Still, is there really going to be a need vs whatever incremental improvements AMD has to FSR2 etc by that point? I'd be a bit skeptical. But really, kinda too early to speculate.
One thing people arent noting here is it means MS is doing a tenth gen console for sure, at least. Since a lot of people seem to doubt MS intentions to stay in the console market every gen. Since it wouldn't even come out til 2028 according to the images, that's something like 5 (remaining current gen)+7 (presumed "tenth" gen length at least)=12+ more years...
I'm for it, just for something different. Majority of R&D is probably in mobile space going forward anyway.
As to all this ML, NPU this and that, I'm still not sold. Google made a big deal out of the Tensor chip NPU and Tensore sucks compared to other state of the art Qualcomm chips. I'm just not sold on the alleged AI revolution. MS has a bad habit of fixating on something stupid that ends up deriding the hardware (I believe arguably the 360 EDRAM, definitely the One ESRAM, fit this category). Is this the next example?? Now it's hardly cut and dried, as a gresat upscaling solution is basically equivalent to a more powerful GPU. Still, is there really going to be a need vs whatever incremental improvements AMD has to FSR2 etc by that point? I'd be a bit skeptical. But really, kinda too early to speculate.
One thing people arent noting here is it means MS is doing a tenth gen console for sure, at least. Since a lot of people seem to doubt MS intentions to stay in the console market every gen. Since it wouldn't even come out til 2028 according to the images, that's something like 5 (remaining current gen)+7 (presumed "tenth" gen length at least)=12+ more years...
