MP All Models Mac Pro X,Y…WHAT IF…

[Amethyst]

Hello guys, i just ask my friend about mac pro, below are his info.
1 This January, he tested 32 core m3 mac studio, and it works perfectly fine.
2 BTW there are no mac studio/ mac pro, he get after that.
3 He has an info about M4 chip but he really can't tell me. T_T
4 There are various prototype of Mac pro pci-e compute card (with unknown Apple AX chip) has been test by him.

 

[ZombiePhysicist]

Hello guys, i just ask my friend about mac pro, below are his info.
1 This January, he tested 32 core m3 mac studio, and it works perfectly fine.
2 BTW there are no mac studio/ mac pro, he get after that.
3 He has an info about M4 chip but he really can't tell me. T_T
4 There are various prototype of Mac pro pci-e compute card (with unknown Apple AX chip) has been test by him.

Does he have any hints if the next Mac Pro will have PCIe 5 and Thunderbolt 5? Looks like the new Mac Pro might come out around June of 2025 with an M4Extreme (2 Ultra's with a fusion connector between them).

Thank you and your friend for the great hints!

 

[Harry Haller]

M4 Extreme! And I kind of called it about a week ago.

Good call.
The Gurman and Max haters are gonna have massive head explosions. 🤣

 

[Boil]

Everyone sleeping on the M4 Extreme Mac Pro Cube...!

And if y'all think the M4 Ultra & M4 Extreme are pretty snappy, wait until the N3X variants drop...!

;^p

 

[kiiso]

We must summon the most noble, most mighty, one and only @Amethyst, to give us some hints! 🙏🔮

Hello guys, i just ask my friend about mac pro, below are his info.
1 This January, he tested 32 core m3 mac studio, and it works perfectly fine.
2 BTW there are no mac studio/ mac pro, he get after that.
3 He has an info about M4 chip but he really can't tell me. T_T
4 There are various prototype of Mac pro pci-e compute card (with unknown Apple AX chip) has been test by him.

It worked! 😳😍🔮

 

[avkills]

Now wouldn't it be icing on the cake if that "compute card" worked in 2019 Mac Pros.

 

[Harry Haller]

Now wouldn't it be icing on the cake if that "compute card" worked in 2019 Mac Pros.

That would be innovative.
So It’s not gonna happen.

 

[Boil]

M4 Extreme Mac Pro will give us more GPU & Neural Engine horsepower for local work, either 3D graphics or AI stuff...

M4 Extreme should also give more available bandwidth to the Mac Pro PCIe slots, allowing the hypothetical "compute cards" (GPU cores, and possibly more Neural Engine cores) to be used for "remote" render jobs or "remote" AI processing tasks...?

So one keeps busy working on the SoC GPU cores / Neural Engine cores, while "remote" jobs are batched over to the "compute cards"...?

 

[Regulus67]

Now wouldn't it be icing on the cake if that "compute card" worked in 2019 Mac Pros.

That would prevent Apple to claim how much better the new mac is in all areas.

I suspect that is part of the reason Apple didn't continue to offer new graphic cards for the Intel version, as well.
God forbid if the Intel Mac Pro 7.1 could outperform the new M2 Mac Pro in the graphic capability
That is ONLY allowed in the PC markets 😆

 

[H. Flower]

So not only is the Mac Pro not dead, but its now going to be able to compete with NVDIA PC workstation?

Is this for real?

 

[Harry Haller]

So not only is the Mac Pro not dead, but its now going to be able to compete with NVDIA PC workstation?

Is this for real?

Maybe, maybe not.
From what I’ve seen I’d say the odds are 50/50 that Apple has a comparable workstation to what’s available elsewhere. Recent rumors say M4 Mac Pro 2nd half 2025. Hoping Apple says something at WWDC. We’ll see.

 

[deconstruct60]

Now wouldn't it be icing on the cake if that "compute card" worked in 2019 Mac Pros.

A “compute card “ with a A18 on it ? What is that really going to substantively compute on ? A-series has even less PCIe aggregate bandwidth than the M1 . There is compute ‘ horsepower’ in the chip bbut how do you get raw data in/out ?


Even an AnnX ( A14X , A18X ) going to get you something like a plain M3/M4 . That will be more like a T2 supplement than a “compute card“ . It would keep it very simple as a pure 75W standard bus powered card . PCIin that case would likely be around x4 PCI-e v4. Which would be a substantive mismatch to the MP 2013 PCI-e ve limitations. ( capped at x4 v3 ). They can kludge with a x4 v4 to x8 v3 switch , but the desupport is coming for MP 2019 .


They would get way more sales volumes if targeted baslleine PCI v4 systems ( and made it mostly host system independent. Mac on a card more so than entirely dependent upon MP userbase. ).

 

[deconstruct60]

M4 Extreme! And I kind of called it about a week ago.

Max Tech ? Really? Just 8-9 days ago he had a video that M4 ultra was coming in June or so . Now is he is not on weak footing foundation?

Sadly I watched a chunk of this . At about 9:07 he flashes a more full full Gurman quote:

“The Hidra chip is designed for the ‌Mac Pro‌, which suggests it is an "Ultra" or "Extreme" tier chip. As for the ‌Mac Studio‌, Apple is testing versions with an unreleased M3-era chip and a variation of the M4 Brava processor that would presumably be higher tier than the M4 Pro and M4 Max "Brava" chips.”

It can either be an Ultra or Extreme doesn’t make it an Extreme solution .

The other old thing is that apparently M3 Pro and M3 Max got different code names .( Lobos and Palma ) . Gurman seems to be trying to toss 3 dies into the Bravapile to spin the super isolation for Hidra .


Hidra can be designed for Mac Pro and still get stuffed into a Mac Studio. The M2 Ultra has gobs of PCIe bandwidth that goes completely unused in the MS deployment . Also doesn’t make any sense at all that the MP woypuld end up with only an extreme chip . The entry price would jump even higher .That is just going to result in fewer MP sold .

To use the long edge of a bigger Ultra chip for UltraFusion doesn’t make any sense like his one mock-up suggests ( going to need memory channels ) .

Decent chance Hidra is just a “big as rectified” Ultra chip . Instead of doing a more chiplet design , Apple just goes monolithic at the top also . Only mimic Nvidia in building as large a die as possible.
Hidra ends up more squared off than the “max” . A big enough floor plan devaintion to get a different codeword.

On the Mac Pro , Gurman has been ‘off’ . Half sized MP … never came . Gurman mentions GPUs cards for M-series MP … never came . Etc. also has thrown shade at Mac Studio disappearing ( not being a replacement to large screen iMac) . That has a flakey track record also .

 

[Boil]

New Mn Ultra that is a monolithic die as big as the reticle allows...

New GPU/NPU-specific die, again as large as the reticle will allow...

This GPU/NPU die is paired with the new Ultra using the method(s) from this post to make the Mn Extreme...

This GPU/NPU is also used (one, two, four dies...?) for the ASi "compute card"...

LPDDR5X would be the utilized RAM; up to 1TB for the Mn Extreme & up to 2TB per four-die ASi "compute card"...?

Maybe even higher clocks if using the N3X process...! ;^p

 

[singhs.apps]

Hello guys, i just ask my friend about mac pro, below are his info.
1 This January, he tested 32 core m3 mac studio, and it works perfectly fine.
2 BTW there are no mac studio/ mac pro, he get after that.
3 He has an info about M4 chip but he really can't tell me. T_T
4 There are various prototype of Mac pro pci-e compute card (with unknown Apple AX chip) has been test by him.

The summon worked.
All hail the Sandman, weaver of dreams.

 

[Regulus67]

Max Tech ? Really? Just 8-9 days ago he had a video that M4 ultra was coming in June or so . Now is he is not on weak footing foundation?

I only watch his videos if he is testing a real product. He has been talking about Mac Pro with M-series extreme chips for two years now. I have never heard Apple officially talk about any extreme chip

 

[impulse462]

So not only is the Mac Pro not dead, but its now going to be able to compete with NVDIA PC workstation?

Is this for real?

I think this is a little premature given how easy it is to program in CUDA and apple hasn't really shown an API to do that yet. As in, knowing the overall architecture of the gpu (e.g. global, shared memory.etc) write a .cu file and have it run on a distributed cluster of GPUs.

However when it comes to replicating pytorch/tensorflow implementations of neural networks, given their new MLX release it does bode hopeful on that end. I am working on a project using MLX right now actually, but I have found that it does NOT still support the ANE that is on my 16" m1 max mbp which is so ridiculous. I don't understand why apple does not fully open up ANE to researchers/developers and my hunch is that the existing ANEs in M1-M3 (which are designed for inference only) are not that great.

This is maybe a reason as to why the M4 chips are being hyped up so much -- however I personally would temper my expectations until anything is released. There's a reason its taken nvidia over a decade to get to this point.

 

[Boil]

I have never heard Apple officially talk about any extreme chip

Maybe we get a preview at WWDC 2024, the M4 Extreme in the all-new Mac Pro Cube...! ;^p

 

[Harry Haller]

Maybe we get a preview at WWDC 2024, the M4 Extreme in the all-new Mac Pro Cube...! ;^p

Why the Cube as opposed to Cheese Grater 2?
By Cube do you mean Studio form factor?

 

[deconstruct60]

Why the Cube as opposed to Cheese Grater 2?
By Cube do you mean Studio form factor?

The ' ;^p ' ascii emoji is a critical piece there.

Because the Cube is a both more unlikely than an Extreme and also a allusion back to a old glory days of Jobs and somewhat the NeXT Cube ( which the Mac Cube somewhat mimiced ) . Combine two unlikely things and get something even more 'rare' and also manage to rustle up all of the ruckus the Mac Pro 2013 dredged up.

An Extreme Chip ( if around 4 * Max ) would be over 300W TDP. The Ultra barely fits inside the thermal coverage of the Mac Studio. If double the package size and thermal drama maybe the Studio gets twice as tall.

All blissfully ignoring the MP 2019 chassis was built with a ~300W TDP SoC in mind in the first place.

 

[Amethyst]

Personally i think that compute card is the way apple can be divided mac pro buyer from mac studio.

PS. oh guys, i have a typo in my last post in (1) my friend have tested 32c m3 in both mac studio and mac pro form factors! not only mac studio in my last post.

 

[deconstruct60]

Recent rumors say M4 Mac Pro 2nd half 2025. Hoping Apple says something at WWDC. We’ll see.

WWDC 2025? More likely coupled to when the Mac Studio M4 gets announced than to WWDC (even if the MP is going to trail the MS by 3-6 months. ). If Apple skipped the M3 generation for the Mac Studio are they really going to wait around until WWDC 2025 to update it? I wouldn't bet on that. Amethyst's friend tested a MacPro M1 configuration that never made it out the door to retail also. M3 Studio may or may not ship in June 2024.

If Apple is squatting on a then two year old MP 2023 product waiting as long as possible into the end of 2025 isn't going to help them. Especially, if they M3-gen bump the MS in 2024 and leave the MP comatose the whole year. The flurry of md-higher end workstation stuff that Intel/AMD/Nvidia is shipping late this year is substantive.

If Apple updates the MP in 2024 with no "sneak peak' then there isn't going to be high need to 'sneak peak' it in 2025 either.

WWDC 2024 no way. Apple might 'sneak peak' something 6 months before shipping. Over a year? Not a chance. It is against standard corporate policy that in been in place for decades. That isn't really 'hope'. It is more ignoring what Apple has repeatedly said they won't do.

 

[deconstruct60]

Personally i think that compute card is the way apple can be divided mac pro buyer from mac studio.

Depends upon what the bandwidth is coming off the "compute card" is and how independent ( latency , system resources) the card is. A x8-16 PCI-e v4 link versus something lighter that might match up with Thunderbolt 4.
The MP 2019 shovels some 'compute' off the T2 over a x4 PCI-e v3 connection. The A-series based compute engine rumors don't really match up with a x8 PCI-e v4 solution.

The Afterburner card offloaded a very highly specialized type of compute. It wasn't a dGPU for more generalized 'compute'. Something A-series sized is odd. Smacks of wanting to put more than several on a card just to get lots of them on there for some narrow embarrassingly parallel thing.

PS. oh guys, i have a typo in my last post in (1) my friend have tested 32c m3 in both mac studio and mac pro form factors! not only mac studio in my last post.

That makes more sense if they want to get the Mac Pro out of the 'sleepy product' reputation. They are likely looking to sell more Ultras over a relatively small amount of time than if just do one of the two products. It wouldn't be surprising if Apple shifted to a more staggered launch with the MS and MP 6 or more months apart so can sell Ultras for more than just 12 months at a time. When more the MS to M4-gen, they measure how much inventory they have for M3-gen and use MP to soak that up before shifting MP to M4-gen. Rinse and repeat.

 

[impulse462]

Depends upon what the bandwidth is coming off the "compute card" is and how independent ( latency , system resources) the card is. A x8-16 PCI-e v4 link versus something lighter that might match up with Thunderbolt 4.
The MP 2019 shovels some 'compute' off the T2 over a x4 PCI-e v3 connection. The A-series based compute engine rumors don't really match up with a x8 PCI-e v4 solution.

The Afterburner card offloaded a very highly specialized type of compute. It wasn't a dGPU for more generalized 'compute'. Something A-series sized is odd. Smacks of wanting to put more than several on a card just to get lots of them on there for some narrow embarrassingly parallel thing.

There are 3 approaches I hypothesize:
1. stick with the unified memory approach. whatever program uses the accelerator card will have compute done on the card with access to only that memory on package
2. take nvidia approach and use hbm memory which will necessarily be memory bandwidth limited (relatively speaking, itll still be /fast/) but it goes against their design philosophy of apple silicon
3. some other way i haven't thought of

 

[mattspace]

There are 3 approaches I hypothesize:
1. stick with the unified memory approach. whatever program uses the accelerator card will have compute done on the card with access to only that memory on package
2. take nvidia approach and use hbm memory which will necessarily be memory bandwidth limited (relatively speaking, itll still be /fast/) but it goes against their design philosophy of apple silicon
3. some other way i haven't thought of

4. Use overt and "thought leader" subvert marketing and propaganda to convince people to lower their expectations to within the capbilities of the current paradigm of beefed-up iPads funning a macOS UI skin, until they can be convinced to do even less, but "enjoy" it because it's "spatial" and like all Apple executives, they surely have a "person" who does the actual work for which they're credited.