News Report: Apple Testing M3 Pro Chips With 12 CPU and 18 GPU Cores

[Admin]

Apple developers are claimed to be testing M3 Pro chips with up to 12 CPU and 18 GPU cores and 36 GB of onboard RAM.

Report: Apple Testing M3 Pro Chips With 12 CPU and 18 GPU Cores : Read more

 

[Dr3ams]

Not an Apple user. But I think Apple's OS would do well if they sold a compatible version for PCs. Same goes for any CPUs and GPUs they develop.

 

[Ogotai]

But I think Apple's OS would do well if they sold a compatible version for PCs. Same goes for any CPUs and GPUs they develop

not gonna happen, if they did, who would apple control its products and such ?

 

[Elusive Ruse]

M2 is basically M1 with more cores, hopefully M3 will offer a proper leap.

 

[lmcnabney]

36GB?
Are we counting cache and GDDR now?

 

[Sippincider]

hopefully M3 will offer a proper leap.

Hopefully it'll be available in a proper replacement for the 27" iMac!

 

[Jimbojan]

I would bet Intel's Meteor Lake CPU for laptop will perform better than Apple's M3, with power efficiency improvement significant from last time (from Intel 7 to Intel4). In fact, Intel's follow on chip will be competitive with any ARM chip for server in power efficiency, and better performance, thus the threat from ARM will be disappeared. Intel was spending $7B/quarter to improve the fab, the cost to make chip will be lower than TSMC ( Apple's). Intel's future is far brighter than and chip design companies like AMD or NVDA, and Apple, with US government chip-act help, no other foreign designer can be more competitive than Intel.

 

[Kamen Rider Blade]

I would bet Intel's Meteor Lake CPU for laptop will perform better than Apple's M3, with power efficiency improvement significant from last time (from Intel 7 to Intel4). In fact, Intel's follow on chip will be competitive with any ARM chip for server in power efficiency, and better performance, thus the threat from ARM will be disappeared. Intel was spending $7B/quarter to improve the fab, the cost to make chip will be lower than TSMC ( Apple's). Intel's future is far brighter than and chip design companies like AMD or NVDA, and Apple, with US government chip-act help, no other foreign designer can be more competitive than Intel.

Proof is in the final product, show don't tell.

Talk is cheap, deliver on what you claim / promise, otherwise it's hype.

 

[peachpuff]

Proof is in the final product, show don't tell.

Talk is cheap, deliver on what you claim / promise, otherwise it's hype.

I heard the m3 gpu will be faster than a 4090ti in Quad sli...

 

[Amdlova]

I heard the m3 gpu will be faster than a 4090ti in Quad sli...

Thanks God. Something needs to stop Nvidia. M1 breaks the 3090 ti now the m3 will show no mercy...

 

[thestryker]

I'm somewhat curious if they've moved to a new core design yet, but I'm guessing not. One place that Apple will likely forever be ahead of X86 is on the integrated graphics side. While AMD/Intel could put out a chip with a much bigger IGP the costs would end up being significant and likely untenable for the OEM market. Though I am hoping as packaging gets better there are more options available so we see more customized chips.

 

[newtechldtech]

Apple really needs to enter the handheld gaming consoles with their chip , it will blow ANY handheld gaming device in the market out of the water.

 

[brandonjclark]

My daily work machine is a M1 Macbook Pro. It's an outstanding piece of hardware. Dare it say it, but if it ran Windows it'd be better.

 

[Ogotai]

M1 breaks the 3090 ti now the m3 will show no mercy...

um no it doesnt for the most part, in performance per watt, yes, in raw performance, not really and i doubt the m3 will and until it is released, its nothing but speculation and rumor

M1 Ultra Doesn't Beat Out Nvidia's RTX 3090 GPU Despite Apple's Charts

Despite Apple's claims and charts, the new M1 Ultra chip is not able to outperform Nvidia's RTX 3090 in terms of raw GPU performance,...

www.macrumors.com

M1 Ultra vs Nvidia RTX 3090

In Apple’s March “Peek Performance” Event, Apple revealed to the world their latest and greatest Apple Silicon to date: the M1 Ultra in, for the first time in over a decade, a brand new Mac, the Mac Studio. During the presentations, Apple made many bold claims such as astounding 800 MB/s memory...

techjourneyman.com

M1 Ultra GPU comparison with Nvidia was misleading - 9to5Mac

When Apple compared the M1 Ultra GPU performance to what was then Nvidia's most powerful graphics card, the company's chart and quote ...

9to5mac.com

Despite Apple's claims, M1 Ultra GPU is not as powerful as RTX 3090 - VideoCardz.com

No, Apple M1 GPU is not as powerful as RTX 3090 Apple may need to reconsider its M1 silicon marketing. The company was claiming that their newest chip called M1 Ultra equipped with 64-core GPU is as good as high-end desktop GPU while consuming 200W less power. As proven today, this claim...

videocardz.com

Apple M1 Ultra Benchmarked, Gets Destroyed by the RTX 3090

RTX 3090 owners can rest easy. Apple's new chip is no threat to them.

www.extremetech.com

it also seems, apple limited the 3090s performance some what to boot. and keep in mind this is for the 3090, not the 3090 ti if it doenst really beat the 3090, its not beating the 3090 ti

i quote :

Proof is in the final product, show don't tell.

Talk is cheap, deliver on what you claim / promise, otherwise it's hype.

 

[bit_user]

Not an Apple user. But I think Apple's OS would do well if they sold a compatible version for PCs.

Aside from all the other reasons not to do this, a big one is that Apple just migrated almost completely off of x86! Why on earth would they backtrack on that?

Also, maybe you're too young to remember this, but Apple actually tried the model of selling their OS to run on 3rd party hardware, for a couple years, back in the late 90's. I even knew a guy who had a licensed Mac clone (this was back when they were PowerPC-based, so the clones were actually built expressly to run MacOS). IIRC, one of the first things Steve Jobs did, upon his return to Apple, was to put a stop to this practice.

Apple makes huge margins on their hardware. If they would sell the OS separately, either it would be too expensive to be very interesting to most, or Apple would be giving up a significant chunk of revenue. Also, it would be more costly for Apple, because they would now have to support a much wider range of hardware than they do today, and that translates into higher development and testing costs.

Same goes for any CPUs and GPUs they develop.

Not sure quite what you mean by this part.

 

[bit_user]

36GB?
Are we counting cache and GDDR now?

No, there's no way it has that much cache. Also, it wouldn't make sense to count that as memory capacity, because cache is non-exclusive and developers would want to know how much actual RAM they can use for their apps. While you might expect a marketing department to try tricks like including cache, it wouldn't make sense for a developer to say that.

However, I was wondering how they arrived at that number. Maybe 3 stacks of 12 GB? Or, maybe there's a 4 GB block of GDDR6X?

 

[bit_user]

I would bet Intel's Meteor Lake CPU for laptop will perform better than Apple's M3, with power efficiency improvement significant from last time (from Intel 7 to Intel4).

Not in the same power budget. Don't forget that the M3 will be using TSMC N3. Furthermore, Apple is willing to burn more die area and other chip costs on maximizing power-efficiency than Intel. For instance, Apple gets an efficiency benefit from in-package LPDDR5X, which Meteor Lake will not have.

In fact, Intel's follow on chip will be competitive with any ARM chip for server in power efficiency, and better performance, thus the threat from ARM will be disappeared.

Even if that's true, you're forgetting about performance per mm^2, where there's no chance Intel can match ARM.

 

[bit_user]

One place that Apple will likely forever be ahead of X86 is on the integrated graphics side. While AMD/Intel could put out a chip with a much bigger IGP the costs would end up being significant and likely untenable for the OEM market.

Meteor Lake appears to support tGPU chiplets of 2 different sizes. Arrow Lake is rumored to have a large tGPU option, but it will need in-package DRAM (like Apple) to avoid being memory-bottlenecked.

I read some analysis of Meteor Lake which suggests the CPU and GPU tiles can actually grow large enough to overhang the contact array.

Meteor Lake Die Shot and Architecture Analysis – Why Is Intel 4 Only A 40% Area Reduction Versus Intel 7?

Die shot analysis shows 40% area reduction on identical structures from Intel 7 to Intel 4. Analysis on Redwood Cove and Crestmont architecture, Foveros Omni, Intel's use of TSMC N3 in the GPU, and SOC/IO chiplets

www.semianalysis.com

 

[thestryker]

Arrow Lake is rumored to have a large tGPU option, but it will need in-package DRAM (like Apple) to avoid being memory-bottlenecked.

​

There are 3 main possibilities I could see for improving IGP memory bandwidth (unless the L4 cache leak turns out to be related to this): on package HBM just for feeding GPU, moving all DRAM on package with a larger memory bus, or moving to a 256 bit memory bus.

With the move to tiles I could actually see adding a single 8GB package of HBM potentially being viable as a SKU line. This seems like it would be the most space efficient way to improve things without a lot of extra engineering.

On package DRAM would probably be the best way to approach it, for mobile especially, as they could really juice the bandwidth and cover both CPU/GPU. This would take up a lot of space and would likely require moving to a larger socket size. On the other hand it would lead to lower power consumption and a more easily controlled experience.

Personally I'd like them to move up to a "quad" channel memory bus for all current "dual" channel chips. It would allow for extra memory bandwidth to the GPU, and also a switch to 1DPC for desktop computing without losing capacity. This would also allow memory to run at much higher clocks with 4 slots populated. While it wouldn't be enough for a run into mid-range dGPU performance it should be capable of eliminating the need for a ~$200 or lower card. This would probably lead to more soldered memory in mobile though.

 

[bit_user]

on package HBM just for feeding GPU, moving all DRAM on package with a larger memory bus, or moving to a 256 bit memory bus.

That's not happening for Meteor Lake, which was my point.

On package DRAM would probably be the best way to approach it, for mobile especially, as they could really juice the bandwidth and cover both CPU/GPU.

That's what Apple's M-series does. They use LPDDR memory, though. The extra bandwidth comes from scaling up to 512-bit (8-channel), at the high-end. But, the baseline model (in M1 and M2) is still 128-bit, so it only offers energy savings and not extra bandwidth.

For 8 GB to be viable, you need to use memory compression. That's what Apple does, anyhow.

Personally I'd like them to move up to a "quad" channel memory bus for all current "dual" channel chips.

Not all models, as it adds cost and burns more power. Plus, it's unnecessary, unless you have lots of CPU or GPU cores.

While it wouldn't be enough for a run into mid-range dGPU performance it should be capable of eliminating the need for a ~$200 or lower card. This would probably lead to more soldered memory in mobile though.

Based on the Geekbench scores, Apple's M1 Ultra is roughly comparable to a RTX 3070, but it's using 1024-bit (16-channel equivalent) of LPDDR5, for ~800 GB/s of bandwidth.

 

[magbarn]

Meanwhile, they'll put in a $25 stock 256gb SSD in the base model and then charge you $800 to upgrade to a $100 2TB SSD.

 

[evdjj3j]

I'd like to see Intel and AMD start producing similar x86 SOCs with strong GPUs to cut into nVidia sales. NVidia has gotten ridiculous with the pricing.

 

[thestryker]

Not all models, as it adds cost and burns more power. Plus, it's unnecessary, unless you have lots of CPU or GPU cores.

The memory controller itself being capable doesn't mean that you're always using all of the channels. By itself the memory controller being more capable should be no real cost difference to Intel nor use notably more power it's only when utilizing the extra channels that comes into play.

AFAIK Intel hasn't run different memory controllers across the desktop/mobile in the history of the core i series. Though with tiles it's entirely possible doing so wouldn't be cost prohibitive, but that would likely be determined by where the memory controller lies. If it's still with the CPU die then this likely wouldn't be viable.

 

[bit_user]

The memory controller itself being capable doesn't mean that you're always using all of the channels. By itself the memory controller being more capable should be no real cost difference to Intel nor use notably more power it's only when utilizing the extra channels that comes into play.

Since I've been reading die-shot analysis of Alder Lake and Meteor Lake, I would take issue that "extra memory channels do no harm". Memory controllers appear to occupy a nontrivial amount of die space, which costs money. Therefore, I'd expect they wouldn't overprovision the memory bandwidth of SKUs relative to their actual bandwidth needs, especially when it's on a separate tile. They can pair the right I/O tile for the CPU and GPU tiles of a particular CPU.

 

[thestryker]

Since I've been reading die-shot analysis of Alder Lake and Meteor Lake, I would take issue that "extra memory channels do no harm". Memory controllers appear to occupy a nontrivial amount of die space, which costs money. Therefore, I'd expect they wouldn't overprovision the memory bandwidth of SKUs relative to their actual bandwidth needs, especially when it's on a separate tile. They can pair the right I/O tile for the CPU and GPU tiles of a particular CPU.

Alder/Raptor Lake is DDR4/5 memory controller, so it isn't comparative to others in general. I haven't seen a Meteor Lake breakdown of the SoC tile, but I would assume there are already multiple versions of it so perhaps they could do multiple memory controller types.