Just wanted to briefly comment on this. From what I understand, there are de-facto only two dies. The Pro and Max are the same floor plan, it’s just that Pro is a partial tape-out. I can imagine Apple saves a lot of money like this v
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
M3 Chip Generation - Discussion Megathread
- Thread starter scottrichardson
- Start date
- Sort by reaction score
Just wanted to briefly comment on this. From what I understand, there are de-facto only two dies. The Pro and Max are the same floor plan, it’s just that Pro is a partial tape-out. I can imagine Apple saves a lot of money like this v
That approach saves lots of money right up until it doesn't scale. At some point, you have to actually put in the work to scale past two 'halves'. But yes, these dies are design with subsystems and to get to a bigger die you primarily add more subsystems ( and perhaps a better network between the subsystems as it gets much bigger).
So a Max is mostly more copies of 'stuff' had in the Pro 'half' of the tapeout. [ plus some relatively narrow changes. ] Scale the subsystems number down and adjust the floorplan a bit and get a 'plain' Mn . Basically same 'building blocks'.
From what I understand the Max simply has more GPU cores (and additional memory controllers/cache to support them)+ what's needed for UltraFusion. I think this method will work well for them for the foreseeable future, unless they want to start differentiating Pro and Max more.
For the laptop products that works great. The upper half of the desktop line up not as much.
However, not doing any desktop optimized design saves money too.
The question gets to whether want to be competitive across the whole line up or maximize the Scrooge McDuck factor.
First Apple patent I am aware of that explicitly mentions four-die Extreme-style configuration: https://patentscope.wipo.int/search/en/detail.jsf?docId=US400934247&_cid=P11-LJTAW4-84142-1
Looks like a different approach than UltraFusion too — an intermediate routing network rather than direct fabric connection.
Looks like a different approach than UltraFusion too — an intermediate routing network rather than direct fabric connection.
So a hypothetical eight SoC system with each SoC containing:
- 8-core CPU (6P/2E)
- 48-core GPU (w/hardware ray-tracing)
- 16-core Neural Engine
- 128GB ECC LPDDR5X RAM
- 64-core CPU (48P/16E)
- 384-core GPU (w/hardware ray-tracing)
- 128-core Neural Engine
- 1024GB ECC LPDDR5X RAM
I can see ways to do that, but it’d either be in NUMA territory, or they’d have to get really creative with packaging and thermals. I wouldn’t hold my breath (but I’d be really interested in their solution).
Why eight? The patent depicts four. Of course, the language is rather generic, but until now the diagrams were fairly close to the actual products. For example, if you look at their “scalable system on a chip” family of patents, that only shows up to two dies, and corresponds almost 1-to-1 to the M1/M2 Ultra. Not to say we’ll never see an eight die system, but let’s not jump before the horses.
There were some rumors that the M3 Pro uses 8+8 CPU config, and the GPU core count of the Max will likely remain close to 40 (depends on whether they use the 3D packaging though). So a hypothetical M3 Extreme would have 32+32 CPU cores and around 20000 GPU compute units.
macOS currently scales to 64 cores/threads, eight SoCs to spread the maximum CPU core count across means more GPU cores...?
Ray tracing is not a memory intensive algorithm. In fact, I don't think you need anything like 8 GB of RAM to render a 4K image. It can be done (slowly) one pixel at a time.
I find it curious that you won’t entertain the idea of Apple patching the kernel to scale beyond 64 cores, but readily jump at the entirely hypothetical notion of eight SoCs
Anyway, the SoC configurations you describe don’t make any sense to me as they would make weak laptop chips. The big advantage of Apple Silicon is economy of scale. Gimping your CPU doesn’t make much sence. Btw, 4x 8+8 CPU clusters give you 64 cores, so you are still within the limit.
But less GPU cores, I am just trying to shoehorn more GPU cores into each SoC...! ;^p
There’s no stopping Apple designing a GPU + memory controller only die, so that’s another avenue that Apple can take.
How can this be? Current algorithms use BVH to reduce intersection time.
Any algorithm that renders game scenes at less than 60 fps is useless.
Well, are we talking about your fantasy or are we talking about what would make business sense for Apple? I don't have any problem with either, it's just that I personally find the first one a bit less interesting
If you are working with triangles, raytracing will take strictly more memory than rasterization (as you have to store the triangle data + the spatial index). If you are working with other type of spacial data (pixels, curves, etc.), raytracing might use less memory, but then again you can always generate geometry on the GPU using mesh shaders... When it comes to high-fidelity retraced scenes, they can be huge, which makes the GPU memory a real problem. Disney's Moana scene for example is over 90GB uncompressed.
And sure, on the basic level you can do a brute-force raytracing at no extra memory cost, but that is trivially uninteresting.
If Apple goes with an 8+8 config on each core, they would only need four SoCs. Depending on the overall die size (which is a major unknown given we have yet to see any 3nm parts announced), it could be possible to fit that configuration into any Mac currently sold on the market, although cooling considerations could limit it to models such as the Mac Studio and Mac Pro.
I have mentioned asymmetrical SoC configurations in the past; one regular SoC paired with one GPU-specific SoC...
Well, are we talking about your fantasy or are we talking about what would make business sense for Apple? I don't have any problem with either, it's just that I personally find the first one a bit less interesting
It's a rumor forum, it's all fantasy until Apple releases a product...?
What I would really like to see would be the aforementioned asymmetrical SoC configuration, or maybe a chiplet configuration...
Just something that gives the high-end Apple silicon more GPU cores...!
Ideally, Apple finds a way to integrated ASi GPUs into a system and folks can just add an ASi card or two for more GPU horsepower...
8+8 configuration and four SoCs means less GPU cores, I am trying to maximize GPU core count, as that is where high-end Apple silicon is lacking...?
Well, technically, "GPU-specific SoC" is contradictory. If it is not complete, it is not a "System on a Chip". It is not unimaginable that they might build an adjunct graphics chip, except the demand for such a configuration within the scope of Apple's target market makes it not really economically viable. They could simply divide the package content to support whichever sizes of GPUs they think will sell, though that leads to other complications: the L3 blocks surround the GPU blocks, so they would probably have to put L3 on its own chip within te package.
They are faced with all manner of design decisions to make the ideal products. In the end, bragging rights comes in behind economics.
So a little over thinking on my end but...a few things to contribute/speculate/question on this thread.
If the Rumor of "base model M3 Pro" to be believed at 12core (6P+6E) 19GPU and 36gb of Ram...
If We to scale that based on the current M2 MAX memory options that apple applied..
M3 pro should have 36GB, 72GB options.
M3 MAX should have 72GB, 144GB, 216gb(?) options
M3 ultra Should have 144GB, 288GB, 576GB options
M3 ULTRA instinct (lol) 288GB, 576GB, 1152GB options
Pretty much what everyone wants in their Mac pros.
My biggest Concern though is that this won't matter much with the Current I/O ports that we have.
Essentially all this power only for the macbooks/imacs/mac studios to still come with Thunderbolt 4...could be a bottle neck here.
So really the only exciting product that will benefit the most from the M3 upgrades is the Mac pro with the now available PCI 5 lanes, everything else is kind of capped until Thunderbolt 5 is out on the M4 or so.
What is everyones thoughts here on this.
If the Rumor of "base model M3 Pro" to be believed at 12core (6P+6E) 19GPU and 36gb of Ram...
If We to scale that based on the current M2 MAX memory options that apple applied..
M3 pro should have 36GB, 72GB options.
M3 MAX should have 72GB, 144GB, 216gb(?) options
M3 ultra Should have 144GB, 288GB, 576GB options
M3 ULTRA instinct (lol) 288GB, 576GB, 1152GB options
Pretty much what everyone wants in their Mac pros.
My biggest Concern though is that this won't matter much with the Current I/O ports that we have.
Essentially all this power only for the macbooks/imacs/mac studios to still come with Thunderbolt 4...could be a bottle neck here.
So really the only exciting product that will benefit the most from the M3 upgrades is the Mac pro with the now available PCI 5 lanes, everything else is kind of capped until Thunderbolt 5 is out on the M4 or so.
What is everyones thoughts here on this.
Oh, I have no prejudice at all agains folks describing their dream configurations, just personally I am more interested in trying to guess what Apple will actually ship. I don't think that what you propose fits well with their product strategy.
I doubt that we will see this kind of asymmetric configurations from Apple. They will likely continue leveraging the economies of scale by offering a very small number of dies (right now they de-facto have two die designs per family!) that can fit multiple roles, with focus on mobile performance first.
Asymmetrical designs will only benefit desktops, and that hasn't really been Apple's focus since the begin of the transition. Patents they have published suggest that they are exploring 3D packaging (likely to make better use of expensive next-gen nodes), but there is absolutely nothing related to asymmetric MCM technology.
Well, yeah, that's not going to happen, simply because they are not going to invest millions and millions of USD just so that a handful of Mac Pro users have an upgrade path. Replaceable SoC boards are more likely, but who knows.
The GPU core count is fine. The M2 Ultra has the same amount of cores as RTX 4080. The problem is the low operating frequency.
Apple usually test development machines with full RAM configuration.
Which could mean that 36 GB RAM could be max capacity for BTO configuration on MacBook Pro with M3 series Chip.
So
12 is the new 8?
24 the new 16?
32 the new 24?
I think Apple would be applauded for that kind of change in the base model... But in true Apple fashion, they'd probably throw all the goodwill away by increasing the prices for memory upgrades
Yo, smart dudes. What can Apple do to make the Mac Studio/Mac Pro competitive with the i9 and 4090 in 3D applications that demand a beefy GPU?
Optimize the software.
Already you have hardware that is plenty capable, but because of the software that is bound to CUDA - you get very high performance on Nvidia GPUs. Start demanding optimization of software, and applications for other platforms, before you people demand "faster horses".
The horses are as fast as possible, to competitive levels. Software isn't.