Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
I'd like a faster GPU as well, but let's be realistic in your comparisons. One is a CPU with an integrated GPU, the other is just a GPU. Of course a dedicated GPU is going to outperform the CPU.
I can't say what apple will do but I will say my M4 Max Studio is one of the best desktops I've ever owned. I think there's something to build upon. The GPU is the weakest link in this setup, whether we're talking AI processing or gaming but as a complete system, it's been great.

I'm not sold that Apple is going all in on gaming and one could make the argument that there just isn't a ROI on them developing GPU technology to compete with Nividia just to satisfy gamers.
 
  • Like
Reactions: splifingate
I'd like a faster GPU as well, but let's be realistic in your comparisons. One is a CPU with an integrated GPU, the other is just a GPU. Of course a dedicated GPU is going to outperform the CPU. If you're going try and do a fair comparison, then compare apples to apples. Compare the M4 to any AMD or Intel CPU with integrated graphics.

While I am a big proponent of meaningful comparisons, I’d say it is fair to compare within the same product category. Apple positions the Studio as a high-performance compact workstation, so it probably makes sense to ask for high-end performance. Apple GPUs are not bad at all, and they are state of the art in some aspects, but the desktop Mac performance still remains underwhelming.

Take again the Nvidia’s Spark that is being discussed here. It’s a very compact platform, 140W, same price as M4 Max Studio. Yet the GPU is two times faster (and 20x faster for ML). Now, Spark is vapor ware currently, and it always will be a low-volume, likely subsidized product, but is it is unrealistic to expect better performance from an M4 studio than a mid-range gaming laptop?
 
While I am a big proponent of meaningful comparisons, I’d say it is fair to compare within the same product category. Apple positions the Studio as a high-performance compact workstation, so it probably makes sense to ask for high-end performance. Apple GPUs are not bad at all, and they are state of the art in some aspects, but the desktop Mac performance still remains underwhelming.

Take again the Nvidia’s Spark that is being discussed here. It’s a very compact platform, 140W, same price as M4 Max Studio. Yet the GPU is two times faster (and 20x faster for ML). Now, Spark is vapor ware currently, and it always will be a low-volume, likely subsidized product, but is it is unrealistic to expect better performance from an M4 studio than a mid-range gaming laptop?
No disagreement with the statements made. I’m curious where you got the specs for the Spark’s performance?
 
Last edited:
I am really confused what people think these machines are for.
More importantly how if at all, impacts Apple's innovation and improvements with their GPUs? I don't see the Studio, or Mac Pro being direct competitors of this, but I could be wrong.
 
  • Like
Reactions: splifingate
More importantly how if at all, impacts Apple's innovation and improvements with their GPUs? I don't see the Studio, or Mac Pro being direct competitors of this, but I could be wrong.

They could be, if Apple delivers a strong GPU matmul implementations. I wouldn’t expect them to ship a FP4 or similar reduced-precision hardware this soon, but they could achieve competitive performance on FP16 and similar formats.
 
  • Like
Reactions: splifingate
They could be, if Apple delivers a strong GPU matmul implementations. I wouldn’t expect them to ship a FP4 or similar reduced-precision hardware this soon, but they could achieve competitive performance on FP16 and similar formats.
Possibly a coincidence but the MLX people just updated it to support FP4.
1756492581087.png
 
Last edited:
Possibly a coincidence but the MLX people just updated it to support FP4.
View attachment 2541962

That’s support for the storage data format - it is converted to some other, natively supported data for processing on the GPU. Metal does not currently offer FP4 as a native format for tensor operations. But who knows, maybe unreleased hardware will introduce a language extension.
 
  • Like
Reactions: OptimusGrime
I don't see the Studio, or Mac Pro being direct competitors of this, but I could be wrong.

Time will tell, no?

I am no "Pro", and I really want to get myself a setup which can help allow me to dive into the river of understanding, and knowledge wrt: 'AI'

It would be nice if I could select a segment of AAPL's Offerings that could help me steer my boat into these previously-uncharted waters . . . those of us that find ourselves in the space where the tire-meets-the-road are feeling left-behind with the latest offerings....

The math works-out to the favor of the Digx, vs the RTX Pro 6000 (and, def. the M3 Ultra).

My M2MS is totally-fantastic. Until something else +++ comes-along, I'm content to float in this temporal cul-de-sac ;)
 
I am no "Pro", and I really want to get myself a setup which can help allow me to dive into the river of understanding, and knowledge wrt: 'AI'
I'm not a pro but what little I do know of AI is you need a lot of certain compute cores, such as Nvidia's cuda cores, and lots of fast ram. I thought I'd be able to dabble with some LLMs on my studio with 36GB of ram - boy was I surprised at how limiting that 36GB of ram is.

It seems with the DGX Spark with 128GB is selling at 4,000 dollars (that's what google is telling me) - A similarly configured Studio will run you about 3,700 dollars (128GB of ram/1TB of storage).
 
  • Like
Reactions: splifingate
you need a lot of certain compute cores, such as Nvidia's cuda cores, and lots of fast ram

My last nVidia purchase was a 980Ti for my MP 5,1 (which was moved to my Dell T5500 (both of which are stored in the third bedroom)).

The EPYC AM5 system I recently built is absolutely (totally) fantastic performance-wise, but it seems I'm now having to revert to an abacus for the math with regards to investment v. return.

Put all 9K in one basket, or put a 9K product in another 8K product . . . it appears that there are not enough beads in my lattice ;)

I reserved the DGX Duo earlier this year, and it looks like it will soon be pay-up, or pass-on.

I'm approaching this idea with the assumption that most of the x-edge current toolkits are NV-centric, so here I am.

I kinda like the DGX approach to the idea that a layman such as I can dip their toes into the water without committing to a Threadripper+1-4 RTX Pro Blackwell Max-Q's (or flagrant intimacy with some cloud-based solution). My personal business should remain personal, and financially viable.

I'm excited at the prospect of being provided with a platform to combine all of this into a single package; a package I can actually power with a 200Amp household circuit.

I hope the next Ultra Studio offers something similar.
 
I think the point is to have unified memory. But this thread is all over the place with people jumping back and forth about setups, so it's hard to tell.

Wanted to validate your reply, and state that I appreciate.

I'm teaching teenagers HS Science on the daily, and one of the highest hurdles my students regularly have to leap is keeping everyone on the same page 🤷‍♂️

To whit:

I'm really excited by the prospect of the next generation of Ultra Studio, and looking forwards to the intro. to the M5, and even more unified memory within the SoC.
 
I reserved the DGX Duo earlier this year, and it looks like it will soon be pay-up, or pass-on.

I'm approaching this idea with the assumption that most of the x-edge current toolkits are NV-centric, so here I am.

I kinda like the DGX approach to the idea that a layman such as I can dip their toes into the water without committing to a Threadripper+1-4 RTX Pro Blackwell Max-Q's (or flagrant intimacy with some cloud-based solution). My personal business should remain personal, and financially viable.
If you are a layman with no goals beyond aimless experimentation, you shouldn't be buying a DGX or any other form of dedicated AI hardware. It won't be useful for anything else and you don't personally need it.

The reason you don't need it is that you don't need to be on the cutting edge to learn. Back when I was an electrical engineering undergrad student, one of the most important project courses I did was based on dinky little 8-bit microcontrollers, at a time when cheap home computers were probably at least a thousand times faster. Another important one was a lab partner and I implementing a CPU - but once again that CPU was nowhere close to high performance. Speed wasn't necessary to learn the principles of how things worked. If you're operating at the equivalent of undergrad level, or less, which is what it sounds like, there's no need to invest in expensive tools yet. You can dip your toes in the water using hardware that's much more accessible - hardware that you probably already have - on much smaller models.

Finally, LLMs and all other forms of "generative AI" are a hype bubble that is going to pop, not the foundational technology of the next 25 years. If you're curious about them, cool, but don't fall for the FOMO being spread by those inflating the bubble.
 
If you are a layman with no goals beyond aimless experimentation, you shouldn't be buying a DGX or any other form of dedicated AI hardware. It won't be useful for anything else and you don't personally need it.

The reason you don't need it is that you don't need to be on the cutting edge to learn. Back when I was an electrical engineering undergrad student, one of the most important project courses I did was based on dinky little 8-bit microcontrollers, at a time when cheap home computers were probably at least a thousand times faster. Another important one was a lab partner and I implementing a CPU - but once again that CPU was nowhere close to high performance. Speed wasn't necessary to learn the principles of how things worked. If you're operating at the equivalent of undergrad level, or less, which is what it sounds like, there's no need to invest in expensive tools yet. You can dip your toes in the water using hardware that's much more accessible - hardware that you probably already have - on much smaller models.

Finally, LLMs and all other forms of "generative AI" are a hype bubble that is going to pop, not the foundational technology of the next 25 years. If you're curious about them, cool, but don't fall for the FOMO being spread by those inflating the bubble.

Appreciate the thoughts, mr_roboto! Serious considerations.

I'm six decades into life, teaching all the HS Sciences (in the same class (4X on-the-daily)), and started my 2y-accelerated MAT this Summer.

That being said, I just want to live a little further towards the edge. Currency is a double edged sword, and the balance goes a long way. I don't need the best, but I do need something that allows me to carry a decent conversation at the Inference/Learning Party ;)

Having been bit in the asterix hard by multiple failed purchases (substantial $ lost) from Teh Bay, I'm a little reluctant to gravitate further towards the Budget light on the porch...
 
What is an "inference machine"?
A lot of people want to run local LLMs and some of them are looking for any hardware that could do it including Strix Halo, Macs, and this Sparks machine because they can all go up to 128GB for relatively cheap.

However, this Nvidia Sparks machine is not designed for cost effective local inference. It's designed fro CUDA devs to mimic big DGX racks. It just so happens to be ok for inferencing.
 
  • Like
Reactions: splifingate
A lot of people want to run local LLMs and some of them are looking for any hardware that could do it including Strix Halo, Macs, and this Sparks machine because they can all go up to 128GB for relatively cheap.

However, this Nvidia Sparks machine is not designed for cost effective local inference. It's designed fro CUDA devs to mimic big DGX racks. It just so happens to be ok for inferencing.

Thank you.

Seeing as I now have a extremely-performant AM5 Epyc 128GB system in the house, how might I best use it to help meet such an objective?

Blackwell RTX PRO's are ~9K, and the M3U/DGX Systems are about the same...there is "Having Dosh", and "Having Unlimited Dosh" (the prior to which I can lay-claim to) 🤷‍♂️
 
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.