Marvin

cgWerks said:

Marvin said:
By comparison, the M2 Ultra is faster than the 28-core Intel chip and faster (27TFLOPs, assuming both GPUs fully used) than the higher-end Radeon GPUs.

Is that really the case, though? Maybe in terms of TFLOPs, but w/o hardware for RT and such, won't certain 3D tasks will perform very poorly compared to a Radeon? I've watched several videos using Studios or MBPs in different 3D/CAD apps and comparing them to PCs (even laptops) where they don't do so well. I suspect some of this is software optimization, but I have to wonder how much is hardware related.

My understanding (maybe even based on a link you gave me some time ago?) is that the GPU approaches are quite different, with advantages and disadvantages (not just TFLOPS numbers). As far as this applies, the problem is that most of the market is developing for, and accustomed to, AMD/Nvidia, so where Apple falls short, it will be quite noticeable in workflows.

Marvin said:
The only market worth targeting is the enthusiast market that would today buy an i9-13900K + Nvidia 4090 for around $4k. The M2 Ultra is within 30% of the CPU and probably 1/3-1/4 the GPU (lower when using hardware raytracing) for $7k.

Yes, this is really the issue, and not just for 3D people, but people wanting to game on the Mac. But, i think it goes even further than that. What about someone who'd buy a mid-level gaming PC with a 3070 or 3080? Then we're in the under $2k range. I still don't think the M2 Ultra will compete, but even the M2 Max becomes expensive. If the GPU is on the table, Apple has no mid-to-high end anymore, though better lower-end than most of their past lineup. It becomes more about justifying any spending beyond a mini or MBA.

Hardware RT definitely makes a difference for some things. Some apps only use it properly on Nvidia (Optix), AMD doesn't do so well in Blender:

Blender opendata

The M2 Ultra is on page 2 and in the same performance range as AMD's 7900XTX as well as all the mid-range Nvidia GPUs.

Nvidia's RT means that anything above a desktop 3070/4060 is faster and the $1600 4090 is 3-4x faster. Nvidia is doing a great job getting optimal software and hardware working together:

https://www.pcgamer.com/nvidia-rtx-40-series-ray-tracing-performance/
https://www.nvidia.com/en-us/geforce/news/geforce-gtx-dxr-ray-tracing-available-now/
https://developer.nvidia.com/blog/accelerating-inference-up-to-6x-faster-in-pytorch-with-torch-tensorrt/

For standard rendering, the GFXBench 4K Aztec 3D test has the 4090 highest again at 507FPS, M2 Ultra 299FPS, about 60% of the performance. M2 Ultra is pretty much even with AMD's highest end GPUs.

gfx bench

Given that Apple wouldn't support Nvidia, a Mac Pro would have offered those high-end AMD GPUs and allowed two of them. The M2 Ultra is equivalent to an Intel Mac Pro with 24-core Xeon W-3345 (250W) plus an AMD W7900 (295W), excluding RT cores.

Hardware RT takes time to implement properly and should allow for motion blur and shader execution order. Apple is doing the software side just now. If it can make it into M3, the M3 Ultra will have the ability to beat all the other GPUs in the Studio form factor but the M2 Ultra looks like it's performing well.

agreer4 said:

I am looking at a Mac Mini - I dont think I need an M-Pro chip but I cant wrap my head around the idea that 8gb RAM is enough for pretty much any computer at this point assuming you want to have it last more than a year or two. Is my vast experiance in Windows configurations clouding my judgment here? do I really need 16GB on the M-class side of the world or has Apple figured out a way to make things not get bottlenecked by swapping, even to NVME/PCIE based flash, its still a bottle neck. I dont want to end up with a brick 2 years from now because I was too cheap to pay the (very over priced) $200 for 8gb extra ram, but on principle i dont wanna spend that much extra on ram. 

My planed useage is for all the basic computer stuff and maybe some media creation - When i had my Mac Mini / Macbooks (2004-2019) i loved Garage Band and am considering stepping up to Logic Pro. Not sure if that would need the ram.

16GB RAM is still best for media work because media gets uncompressed into memory. 8GB can hold up ok for performance due to the SSD but it will be swapping a lot. The system will use 2-3GB to boot the OS and this is also used for the graphics memory, which would be 1GB. App memory left after this is 4GB. A single browser tab can use 0.5GB RAM.

Most media apps will say 4-8GB minimum requirement and 16GB recommended like Photoshop:

https://helpx.adobe.com/photoshop/system-requirements.html

If you open two apps like Photoshop + Final Cut Pro, that doesn't leave enough to meet the minimum for both. 8GB RAM is for the most basic use case of browsing, document-based workflows, emails etc. Any kind of audio/video/image work should get 16GB.

tht said:

flori said:

Looks like Ultra M2 beats all graphics cards by a wide margin with a score of 281,948.  Radeon RX 6950XT is at 244,820.

https://browser.geekbench.com/metal-benchmarks?mibextid=Zxz2cZ

Looks like a mistake by Primate Labs. That score is more than 2x the score of the M2 Max. Unless there are a lot of Compute sub benches that are memory bottlenecked which the M2 Ultra overcomes, hard to believe that score. 

Geekbench compute scores vary a lot so can be quite unreliable but there are also two different models each of Max and Ultra and they don't say which they are testing, they describe them by CPU cores.

https://www.apple.com/shop/buy-mac/mac-studio/12-core-cpu-30-core-gpu-16-core-neural-engine-32gb-memory-512gb#

There's M2 Max with 30 and 38-core GPUs. M2 Ultra with 60 and 76-core GPUs.

A 76-core Ultra would be expected to be more than 2x a 30-core M2 Max.

Here's an M2 Ultra at 223k:
https://browser.geekbench.com/v6/compute/538846

281 / 223 = 1.26x. 76-core / 60-core = 1.26x.

Here's M2 Max at 122k and 144k:
https://browser.geekbench.com/v6/compute/541704
https://browser.geekbench.com/v6/compute/538762

281 / 144 = 1.95x. (presumably 76-core Ultra vs 38-core Max)
223 / 122 = 1.82x. (60-core Ultra vs 30-core Max)

There are also different power levels in the OS, these two Max pages have different clock speeds.
They also have different cooling capability. A Mac Studio will cool an M2 Max better than a Macbook Pro 14".

It looks like they have improved the Ultra GPU scaling though. We'll see with the Blender tests:

Blender Opendata

M1 Max (32C) is on page 5 with 1036.
M1 Ultra (64C) is on page 4 with 1831.

M2 Max (38C) is already above M1 Ultra on page 3 with 1917.
If M2 Ultra (76C) can scale to near 2x, that will bring it to the top of page 2. This is 1/3 from the top overall (and those others have hardware RT).

An M2 Ultra 76-core showed up on page 2 with 3412, 1.77x the M2 Max.
A 4090 is about 3.8x, 4080 is 2.8x, 3090 is 1.8x.

To get near to the top, they'd need a 50% boost from M3 to make a 114-core Ultra and then add a second GPU chip that is equivalent to another 114-core Ultra. Or maybe 50% boost + hardware RT would be enough.

lowededwookie said:

Parallels is awesome for some games but not all. But it’s still not a Mac game is it? It’s a Windows game running on Windows on a Mac.

we NEED native games.

That's why the porting toolkit Apple released is better. Here's Cyberpunk 2077 running just like a native Mac game. At 2:08, they show the dock icon and Mac environment:



Diablo IV:



There's no Windows OS involved. It's a Windows game but the compatibility layer intercepts all the requests to the Windows system and redirects them to the Mac system equivalents.

When the game says it wants to draw an object using DirectX 12, it gets caught and changed to being drawn using Metal instead. As far as the user is concerned, it gives the same feel as a Mac game, it just has some overhead due to the translation.

So far more than 40 games have been ported in just a week. To do native ports of those, each game studio/publisher has to get a porting team together as well as QA and a budget for doing the port. If the port isn't using a cross-platform engine, a Metal port would take 6 months and cost maybe $400,000. Using a cross-platform engine, it could probably be done in 1-2 months, around $100,000.

That sounds reasonable for a single high profile game but it doesn't work commercially doing it for just one game, there needs to be lots of them for gamers to invest in the platform. For 40 studios to go through that process is near impossible let alone the top 100 games for the past 20 years - 2,000+ games that aren't on Mac. It would take years for that to happen. The Game Toolkit can make this happen in 1 month. But it needs to be an officially supported method and easy for people to install the games.

Once the gaming audience is big enough on Mac, there will be a commercial reason to do native ports but there will always be the Porting Kit as a fallback where it's not feasible to go native.

thadec said:

Marvin said:

The high performance computing pro market is a rounding error in the computer industry and has been for a long time.

Not even close to being true. The workstation, small server and HEDT markets are booming. There was actually a Threadripper shortage last year. Intel launched Sapphire Rapids (Xeon W-2400 and W-3400) this year to compete with Threadripper in the desktop workstation market. AMD is going to launch the Threadripper 7900X this year along with the usual Pro 7900WX in order to compete with Intel in HEDTs. I am still wondering what is keeping Nvidia from getting in on the act with a stripped down version of their Grace chips targeting this market.

Engineers, architects, researchers etc. have always needed workstations. Do you think that these entire professions stopped existing? Moved to the cloud? Or can get their work done on MacBooks now? And now they have company. Thanks to YouTube and all that, the number of people into serious video editing and computer animation has gone through the roof. Add to those the AI/ML boom the past few years and now the LLM types is going to mean still more. Yet rounding error you say. 

 Apple not being able to do the Extreme until TSMC's 2nd gen 3nm process is ready until 2024, which forced them to just stick an M2 Ultra in a cheese grater and call it a day because needing to move forward with Sonoma meant that they couldn't wait any longer, is no reason to just go and make up stuff OK?

As I mentioned, it's worth targeting the HEDT enthusiast market (around $4k). The non-server market above this is a rounding error.

High-end Threadripper, Epyc and Xeons are mainly for servers. The Ultra chips are competitive with the cheaper desktop Threadrippers and i9s.

Many engineers, architects, researchers, Youtubers, artists have been able to migrate to Macbook Pros. The Max chips perform comparably to higher-end desktops and can outperform the 2019 Mac Pros.

https://www.tomshardware.com/news/nvidia-maintains-lead-as-sales-of-graphics-cards-hit-all-time-low-in-2022-jpr





Notebooks continue to erode desktop relevance. Lowest overall sales in over 10 years and falling. Nvidia's growth is in the server space with AI.

According to this, Nvidia sold about 30 million desktop GPUs in 2022. Notebook GPUs would be at least as many and their GPU revenue was $12b.

$12b / 60m units = $200 ASP.

High-end GPUs like the 4090 ($1600+) sell around 1m units per year. The majority of GPUs sold are consumer gaming cards:

https://store.steampowered.com/hwsurvey/videocard/

The Ultra Mac Studio and Mac Pro are aimed at the market that buys single 4090 GPUs. The market above this ($5k+) is well below 1%. The HEDT/enthusiast market is around 5-10%.

The M2 Ultra falls short of a $4k i9-13900 + 4090 but it's at least competitively priced in the Mac Studio. Chasing after a minuscule portion of the high-end market isn't a high priority, it can easily wait for M3 for hardware RT and possibly some extra GPU chip.