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.
Lets talk iGPU
- Thread starter Danny82
- Start date
- Sort by reaction score
Why do you say this? I've always assumed that Apple were pissed at Nvidia because of the endless GPU failures that cost them a fortune in repairs to generations of MacBooks. CUDA is mostly for heavy duty number crunching isn't it? As far as know Apple doesn't have an equivalent to bind together GPUs into computing clusters.
Someone mentioned the Afterburner card in the Mac Pro. I have speculated elsewhere that this card might be used as a specialist GPU replacement of sorts. Since its essentially an array of FPGAs, it can be reprogrammed by Apple for other tasks. I imagine there is a variety of graphical jobs it could do if Apple want it to.
The other thing people don't seem to be expecting with Apple Silicon, is any significant changes to the Mac lineup as a whole. I haven't seen anyone but me point out that last time they changed CPUs they changed all their Mac model names except the iMac but even if they keep names, the variety of extra types of cores and modular features they mentioned in the keynote means Apple can build Macs for any specialist market they feel is worthwhile. They could offer a Mac Pro for video with a different chip to the Mac Pro for audio. They could do one with a massive stack of ML cores or modules, or the neural engines they use. They can do anything they want now.
A Mac Mini designed for high throughput and ultra low power in data centres. If it can even get close to a 1U Dell or HP rack server, those are going to sell in serious volumes. Because you can fit 8 of them in a 1U slot and 24 in 3U and either shrink your data centre or x10 its capacity without needing to upgrade the power supply.
It would be interesting to know all these details, but I am afraid it is locked behind seven legal seals
CUDA is Nvidia's "native" GPU programming framework. Nvidia did some very clever (if underhanded) marketing to put themselves in a position where they pretty much control all the GPU-driven scientific and machine learning workflows. Apple is trying to establish themselves as a new powerhouse in this segment. They want you to use Metal and CoreML that are optimized for their hardware. But given how much tools are readily available that utilize CUDA, if Apple gave Nvidia access to the Mac, nobody would bother using Apple's frameworks (and hence Apple's hardware).
The way I see it Apple's goal is to create an environment where developers and researches are aggressively encouraged to use Apple's APIs. The strategy to achieve this is a) release some really good hardware that people will want to use and b) eliminate the options so that people are forced to use Apple's API and Apple's API only. Basically, Apple is asking everyone to spend some additional effort but is promising a worthy payoff if they do so.
And it is not clear that Apple will succeed. The performance of Apple's GPUs and ML cores is still a bit unknown, as Apple does not publish clearly interpretable metrics. What I was able to find is that latest iPhone Neural Engine is supposed to deliver around 5 TFLOPS (I assume this is half-precision operations). Nvidia's RTX 2060 tensor cores offer over 50 TFLOPS (not sure if this is 32bit or 16bit). But I am not sure what this means in practice.
You can use Metal for this. Not for the distributed case though, that is a logic layer you will have to implement yourself. I am not too familiar with the capabilities of modern CUDA in this regard, it was a while that I had a look.
exactly.
As a software engineer specialized in machine learning, nobody uses Apple CoreML. Anyway, you don’t train a model on Apple’s hardware. You train on Linux servers.
In 16-bit MATMUL OPs on Tensor Cores.
The framework now includes more and more features for distributed computing, multi-GPU computing. PyTorch and Tensorflow also implements layers for this which ease the task a lot (also because it’s in Python and not in C++).
Agree that the navi will and maybe support the intel macs.. but cant imagine Apple suddenly releasing a dGPU as great as the one from Nvidia and AMD.. and definitely not an iGPU that competes with dGPU.. would probably take them a few more years.. of course that being said, it would be nice if they can come out their own dGPU so soon
Anyhow, if the iGPU is 2x more powerful than the one on crappy intel will be most welcome
The “crappy” Intel iGPU is faster than anything Apple currently has available.
And next year, Intel is launching Tiger Lake which has an iGPU that is supposed to be 3 times faster than the Ice Lake iGPU (that currently already beats everything Apple has).
Might want to check your information. The iPad Pro GPU (2 years old chip at this point) is about twice as fast that current Iris Plus. Leaked benchmarks of Tiger Lake G7 would place it around 20-30% faster than A12Z GPU. I very much doubt it will outperform the final Mac chip running at higher power.
That is not true. The iGPU in the 2020 13" MBP scores over 10.000 in the metal benchmark, which is higher than the iPad Pro.
Curious if anyone got information on the TDP for latest intel iris and a12z..? Just want to imagine how much further apple gpu can push themselves below it takes a big hit on battery life.. like for the mbp16 5600m running at TDP 50w..
A12Z is rated at 15W TDP. So I guess 3 times of the current the performance if they increase the TDP to 50W? (assuming performance scales linear with TDP)
A12Z scores 12000 in geekbench 5 compute. And it’s about 60% in graphics benchmarks such as 3Dmark. So in the sense you have set me right - I misspoke about being twice as fast. It’s only like 50% faster.
I'm sorry, but this doesn't seem to agree with the data I have seen. Looking at Geekbench 5 results for OpenCL and Metal, the A12Z exceeds any Intel iGPU (Iris Plus seems to be the best one) by a healthy margin.
The Anandtech reviews of the 2018 iPad Pro also show the A12X doing well well against laptops, even beating those with an MX150 dGPU, and not that far behind those with a GTX1060: https://www.anandtech.com/show/13661/the-2018-apple-ipad-pro-11-inch-review/6
Which benchmarks lead to you conclude otherwise? Please provide some evidence of your claim - I'm open to learning something new!
Tiger Lake iGPUs will no doubt be better, but then so will whatever Apple produces in their new SoCs.
The argument will be academic soon in any case, because what is certain is that after 2 years no Macs will be using Intel iGPUs. I'm confident that Apple will make Apple Silicon competitive with Intel CPU / iGPU performance to ensure its market position.
The more interesting question is how long Apple will take to match current discrete mobile and desktop GPUs. Will they even try to go up against NVidia and AMD in the desktop market? It would shake things up if they did!
I hope my research has helped you!
Gizmodo seems to disagree with you
So Just How Powerful Are Apple's New Laptop Chips Gonna Be
Apple officially confirmed in its WWDC keynote yesterday that, yes, it would be using its own processors in its laptops and desktops. Not a terribly
You are looking at a review from 2018. In 2020, Intel has released new 10nm chips with a new Ice Lake iGPU. That one scores 10241 in Geekbench 5 (metal) (the high-end 13" 2020 MBP uses this 10nm chip from Intel). The iPad Pro is around 8900.
Now if test results in the link are correct, Tiger lake will be a beast of iGPU.
Intel's Tiger Lake Xe iGPU is almost three times faster than the Ice Lake Iris Plus G7 iGPU and up to 40% faster than the AMD Vega 8 Renoir iGPUs in Battlefield V
While still not a threat to any decent currently available laptop dGPU, the Xe iGPU that will be featured in the Tiger Lake processors launching later this year should still be quite a significant improvement over the Ice Lake iGPUs. Our tests show that the Xe iGPU could even beat the Vega 8...
Gizmodo seems to disagree with you
So Just How Powerful Are Apple's New Laptop Chips Gonna Be
Apple officially confirmed in its WWDC keynote yesterday that, yes, it would be using its own processors in its laptops and desktops. Not a terriblygizmodo.com
Nice article. Somehow they did not include the upcoming Tiger Lake next year. Considering that Apple ARM will be released next year, it makes sense to compare it to Intel and AMD products in 2021 at least.
Also, Gidzmodo is really full of it. They claim based on how it runs Shadow of the Tomb Raider, it is the most powerful iGPU on the market? There are benchmark indicating that Intel iGPU's are even more powerful.
Last edited:
Again, iPad Pro with a 2 year old GPU is 11k-12k in Geekbench 5 (metal) that Ice Lake G7. And 50-60% faster in graphical benchmarks.
Tiger Lake is undoubtedly a big jump for Intel. But 80% performance increase over Iris Plus is not enough.
[automerge]1593940294[/automerge]
Apple ARM laptops are out this year.
Seriously, please stop for a moment and check your data. Your last couple of posts was you just making random claims.
Thanks.. understand that this is just guessing till we see the product.. but well.. the process of guessing and hoping is fun on its own rights
3 time the performance of a12z.. hmm.. this should be awesome right I see. I found the 8900 Geekbench in some article reviews of the iPad Pro. But seems that there are people on Geekbench scoring higher numbers (maybe due to newer iOS update?).
I got the 10241 from my own 2020 13" MBP, but I guess you can find higher number if you browse the test results.
I’m sure they are aware of it yet they come to this conclusion. Wouldn’t that tell you something about your data?
In fact I have never found any publishers tell otherwise (that Intel iGPU is faster than A12Z).
I'd say, let's wait for Tiger lake then. It is supposed to be released this year (not sure when it will actually end up in laptops). Also will be interesting to see what AMD will do on the iGPU side.
Yep, I hope they keep pushing each other! Faster integrated graphics means raising the baseline, which means better software and experience for users all across the board.
Even if Tiger Lake iGPU is faster than an A12Z, that chip is nearly three generations old and its running in a thermal envelope that is super tight compared to anything a Ice or Tiger Lake chip could possibly run in. My bet is the first Apple Silicon Mac will absolutely smoke any other iGPU available at release.
It does not. Performance scales linearly with the frequency, but power dissipation grows roughly cubically with the frequency, because in order to increase the frequency you typically also have to increase the voltage (which by itself causes the power to grow quadratically). In other words, the performance scales logarithmically with the power.
It is also not at all clear how far Apple's CPUs will actually be able to scale (i.e. up to which frequencies and power levels they are able to work reliably), since so far Apple has only made low-power variants. It's not as simple as slapping a bigger cooler on the CPU and pumping in more power.
It does not. Performance scales linearly with the frequency, but power dissipation grows roughly cubically with the frequency, because in order to increase the frequency you typically also have to increase the voltage (which by itself causes the power to grow quadratically). In other words, the performance scales logarithmically with the power.
It is also not at all clear how far Apple's CPUs will actually be able to scale (i.e. up to which frequencies and power levels they are able to work reliably), since so far Apple has only made low-power variants. It's not as simple as slapping a bigger cooler on the CPU and pumping in more power.
That is why I see larger (think of a Threadripper-sized APU) Apple Silicon SoCs for the Mac Pro line-up:
32 P cores / 12 E cores / 48 GPU cores / 32GB HBM2e UMA
48 P cores / 18 E cores / 64 GPU cores / 48GB HBM2e UMA
64 P cores / 24 E cores / 80 GPU cores / 64GB HBM2e UMA
The 2019 Mac Pro CPU heatsink / fans should be able to cool such an APUzilla.
Additional RAM (DDR5) & NVMe storage on logic board.
Expand the Mac Pro line-up with a new Mac Pro Cube, with integrated A/V DSPs & I/O for a complete entry-level DDC workstation! Bundle it with Final Cut & Logic. Silicon Graphics / SGI meets Video Toaster meets Apple Silicon!