M3 Ultra Chip is Only 10% Faster Than M4 Max in Benchmark Results

[rp2011]

I'm quite surprised to read many comments pointing out the AI usage. I get that AI is the thing of the moment, still: what about the regular stuff these machines were supposed to work with? You know: video editing, 2D/3D graphics and audio editing.

Are all professionals suddenly turned into AI farm builders or what? To do what, then? OK, you have a remarkable computing power on hand. And what about it? What data are your locally running AI models going to compute?

Considering the RAM/GPU features, is an M3 Ultra more suitable for video editing than M4 Max? Shouldn't this question be more appropriate for the regular user to evaluate the M4/M3 differences?

The M4 Max, even on MacBook, is more than good enough for *most professionals. I'm probably going to choose it over the Ultra this time around, and I use it for graphics work that requires plenty of GPU power and memory. Also, "professional work" changes. As computers become more powerful, they will be put to work by a new group of professionals for more and more demanding tasks, such as AI. This shouldn't be surprising.

Having said all that, I really wanted an Ultra based on the M4 as that would really have felt worth it for me vs only 8% performance increase. I'll save my money as I am not running LLMS.

 

[pksv]

The M3 Ultra’s large unified memory enables it to load massive models at a lower cost.

However, for running dense or hybrid models and most importantly for training (the most important thing for an AI developer) its performance falls short. In these cases, the lower memory refresh rate and bandwidth compared to dedicated GPUs make it less suitable.

You are also right. 1 to 1 Nvidia is still a king, if only money is no object. Apple cannot compete with them when it comes to performance. But that is not the point here, at a cost that is an order of magnitude lower Apple has found its niche and will be used in many places. Having the ability to run a non-distilled model privately for $30k-40k is a bit of a game changer. All the more so when you consider that these are complete devices, all you have to do is plug them into the power.

 

[rp2011]

I honestly think this was a bit of a late scramble to put something together for the LLM space, which has become popular, with Nvidia releasing Project Digits. Apple probably had this relatively underwhelming M3 Ultra on the back burner but knew it was something that could be configured with massive amounts of RAM.


This is going to be one of the only systems (if not the only) with effectively over 500GB of VRAM to run an LLM. Project Digits is $3000 with only 192GB.

Yep, especially as we are seeing PC makers jumping in to try and scoop the spotlight from the improvised Mac Mini clusters that folks are cobbling together for this.

 

[chucker23n1]

the odd thing is that it's nowhere near twice as fast as an m3 max. Any explanation for that?

Geekbench 6 penalizes high CPU core counts, because in practical use, few workloads take advantage of them.

 

[pksv]

I honestly think this was a bit of a late scramble to put something together for the LLM space, which has become popular, with Nvidia releasing Project Digits. Apple probably had this relatively underwhelming M3 Ultra on the back burner but knew it was something that could be configured with massive amounts of RAM.

Very valid points and very much possible that that was the case.

This is going to be one of the only systems (if not the only) with effectively over 500GB of VRAM to run an LLM. Project Digits is $3000 with only 192GB.

Correction: Project Digits will have 128GB VRAM not 192GB
And very important: you can't buy Digits yet, while you can purchase M3U now and have it on your desk in a week.

 

[mrFab]

I can confirm this result from our own testing (we have a review unit) - the odd thing is that it's nowhere near twice as fast as an m3 max. Any explanation for that?

Geekbench 6 penalizes high CPU core counts, because in practical use, few workloads take advantage of them.

It should be noted that chip fusion comes with an overhead, as the communication between the chips adds latency, which limits the theoretical gain.

 

[rp2011]

You are also right. 1 to 1 Nvidia is still a king, if only money is no object. Apple cannot compete with them when it comes to performance. But that is not the point here, at a cost that is an order of magnitude lower Apple has found its niche and will be used in many places. Having the ability to run a non-distilled model privately for $30k-40k is a bit of a game changer. All the more so when you consider that these are complete devices, all you have to do is plug them into the power.

It's looking like Apple IS capable of competing with Nvidia with consumer GPU's if they really wanted to. The M4 Max is proof of that. I don't know why they seem to have been so adamantly opposed to giving consumers decent GPU performance for so long. Hopefully, this is starting to change. I agree with those who think Apple has been caught off guard by the direction computers have taken in terms of GPU's. Graciously ceding all that money to Nvidia does not sound like it was done on purpose. It sounds like plain o'l bad decision-making.

 

[pksv]

It's looking like Apple IS capable of competing with Nvidia with consumer GPU's if they really wanted to.

Of course, this is a derivative. Both Nvidia and Apple chips are manufactured by TSMC, so in theory they can produce exactly the same thing. The difference is that Nvidia doesn't care about power consumption, size, and temperatures.

 

[chucker23n1]

in theory they can produce exactly the same thing

This is only true if core design doesn't matter, which it does.

 

[Veinticinco]

Max? Ultra? All so confusing now so think I’ll just wait for the Uber chipsets.

 

[rp2011]

Of course, this is a derivative. Both Nvidia and Apple chips are manufactured by TSMC, so in theory they can produce exactly the same thing. The difference is that Nvidia doesn't care about power consumption, size, and temperatures.

If this was the case AMD and Intel as well as everyone else would have already taken a slice of the Nvidia cake. If it were so easy.

 

[chucker23n1]

It should be noted that chip fusion comes with an overhead, as the communication between the chips adds latency, which limits the theoretical gain.

Sure, but I would still expect gains to be far bigger than 8% with double the core count, if you have a workload that can take advantage of 17+ cores.

For example, we can see (with the caveat that this is literally one benchmark run against one other benchmark run; not a high sample size) that Clang, Asset Compression, and Ray Tracer are 63.3%, 71.1%, and 90.5% faster, respectively — even though, at single-core, each of them is slower.

So that does suggest that when the workload does scale very well, the M3 Ultra's scheduler will take advantage.

(Of course, Amdahl's law, yadda yadda.)

 

[pksv]

This is only true if core design doesn't matter, which it does.

Only to a some small degree. All the competent brands can produce something that will be in a similar ballpark.
A perfect example is the MI300X from AMD, which is not only better but also cheaper compared to Nvidia's H100. What is truly unique about Nvidia and what made it a frontier in GPU and AI is their software layer - CUDA.

 

[pksv]

If this was the case AMD and Intel as well as everyone else would have already taken a slice of the Nvidia cake. If it were so easy.

But the final product is the hardware AND maily the software layer and this is where the difficulty lies.

 

[chucker23n1]

Only to a some small degree.

Nah. Qualcomm, for example, was able to make huge gains by buying Nuvia and investing in the core design.

 

[Bonte]

What would be the best choice, a M3 ultra with 96 GB RAM or a M4 max with 128 GB RAM?

Price is about 1500 dollars difference but also more cores and more TB connections on the M3. It’s a strange lineup.

 

[pksv]

Nah. Qualcomm, for example, was able to make huge gains by buying Nuvia and investing in the core design.

Read what I wrote again. "All the the competent brands" meaning that all the top companies are playing in the same ballpark of chip design. It's not like Nvidia's chip design makes their chips hardware wise better by huge percentage.
I even gave you an example of MI300X vs H100 I don't understand why you decided to ignore this.

 

[rp2011]

But the final product is the hardware AND maily the software layer and this is where the difficulty lies.

Yep Nvidia's Cuda technology has become so ubiquitous that so much of AI is built around it. Companies that can build their own chips and do not need compatability are starting to forge ahead with their own or alternate parallel computng technologies.

 

[Aeronauts]

I wish Apple would stop using these 1.something comparisons and just use percentages instead.

 

[3xBoom]

1.5x means 150% faster, not 50%.

You should be asking an LLM for the right answer

 

[pksv]

Exactly. Additionally, some companies work on the translation layer

Yep Nvidia's Cuda technology has become so ubiquitous that so much of AI is built around it. Companies that can build their own chips and do not need compatability are starting to forge ahead with their own or alternate parallel computng technologies.

Exactly. Additionally, some companies are working on the translation layers, so Nvidia's uniqueness will eventually end. It will remain the king due to the slots in TSMC, but they will definitely have to lower their margins and move to lower nodes.

 

[chucker23n1]

Read what I wrote again. "All the the competent brands" meaning that all the top companies are playing in the same ballpark of chip design.

Which is a highly subjective and pointless list. It lets you win any argument by simply asserting "well, it must be because they are/aren't competent".

I even gave you an example of MI300X vs H100 I don't understand why you decided to ignore this.

Because it isn't relevant. Obviously, both Apple and NVIDIA are watching each other and trying to learn from each other's design improvements. It costs huge chunks of money, and isn't as simple as "in theory they can produce exactly the same thing".

 

[pksv]

Which is a highly subjective and pointless list. It lets you win any argument by simply asserting "well, it must be because they are/aren't competent".



Because it isn't relevant. Obviously, both Apple and NVIDIA are watching each other and trying to learn from each other's design improvements. It costs huge chunks of money, and isn't as simple as "in theory they can produce exactly the same thing".

Lol. Ok mate, live by your beliefs ignoring reality if that's what you truly prefer, it makes no difference to me.

 

[anthony13]

Im betting for my purposes a m4 max with lots of ram will be better than an m3 ultra. I work with large 3d spaces in Vectorworks. I also typically have a lot of apps open. I don’t need the extra tb5 ports and only currently have two 5k displays. It seems like a lot of the rendering modes in Vectorworks also rely on processor over gpu.

 

[8KYUP]

Yes, I would've loved to see them put out an M4 ultra, but this allows LLM's like full Deepseek R1 671B (likely 8 bit MLX) to run on a stock, home computer for under $10,000. Apple wanted to get that out there because there is hunger for it. That's just crazy and I'm in.