m1, egpu, XDR x3

[kelvin.lau]

I'm currently running 3x XDR displays with a MacBook Pro. I know my setup is an outlier, but I'm pretty sad to hear the m1 chip doesn't support BlackMagic eGPUs. They're the reason why I could hook up my 3 XDRs in the first place.

I'm curious if there's any news about m1 support for eGPUs. Did they ever say there will be support later?

 

[Pressure]

Wait for the 16” replacement or equivalent next year. There is currently no eGPU support for M1 and no way to drive 3 XDRs.

 

[leman]

I'm curious if there's any news about m1 support for eGPUs. Did they ever say there will be support later?

As far as I know, Apple has not provided any information action about eGPU on Apple Silicon Macs. Since eGPUs require separate drivers (which are expensive to write and maintain for the new system) and a different programming model to Apple GPUs, I do not believe that we will ever have them with these new Macs.

 

[Hexley]

Wait for the 16” replacement or equivalent next year. There is currently no eGPU support for M1 and no way to drive 3 XDRs.

I agree. 16" will come out between January and WWDC 2021 in June.

 

[torncanvas]

As far as I know, Apple has not provided any information action about eGPU on Apple Silicon Macs. Since eGPUs require separate drivers (which are expensive to write and maintain for the new system) and a different programming model to Apple GPUs, I do not believe that we will ever have them with these new Macs.

With that said, there’s no information either way, and delivering GPU performance that would rival AMD’s highest end Big Navi GPU (which matches the new Nvidia 3090) while also finding a way to scale Apple Silicon up to Mac Pro levels, would be a much harder milestone than any existing Apple Silicon achievement.

It’s one thing to come up with a mobile GPU at 50W TDP that rivals competitors; it’s a whole other level of effort to deliver one at 250W. And in order to get people to buy an AS Mac Pro, the GPU needs to be at least as powerful as the next AMD card released whenever Mac Pros come out.

Therefore, for these reasons, it’s more likely than not that there will be AMD cards for AS Macs, which means an eGPU will be supported even in a Hackintosh type way in the future, since lack of drivers is the only thing holding support back.

 

[Takuro]

eGPU support may or may not come. Apple Insider has a story that people managed to get an eGPU detected in System Information, but that's a far cry from actual support.

 

[leman]

With that said, there’s no information either way, and delivering GPU performance that would rival AMD’s highest end Big Navi GPU (which matches the new Nvidia 3090) while also finding a way to scale Apple Silicon up to Mac Pro levels, would be a much harder milestone than any existing Apple Silicon achievement.

It’s one thing to come up with a mobile GPU at 50W TDP that rivals competitors; it’s a whole other level of effort to deliver one at 250W. And in order to get people to buy an AS Mac Pro, the GPU needs to be at least as powerful as the next AMD card released whenever Mac Pros come out.

I think it is much simpler to upscale a smaller GPU than to downscale a larger GPU. Apple GPU IP is scalable, it's competitive with anything that AMD or Nvidia have to offer when you look at the building blocks, and it has undeniable advantages in performance-per-watt over the competitors (about 1.5-2x in compute and 2-3x in graphics). I see no technical reason why Apple would be unable to build a large GPU cluster containing 32 or more cores, that would give you 4096 ALUs and over 10 TFLOPS while still being a sub 50W system. The only puzzle bit missing is a fast interconnect capable of wiring all those cores, but it is certain they already working on it (they were aggressively hiring interconnect engineers a while ago). Memory is pretty much a solved problem (one can always use HBM or some other wide-interface RAM) and Apple has been using an Infinity Cache-like structure long before AMD had it.

Given what Appel already did in chip business, building a large GPU is probably not difficult for them at all. It's just expensive. And they have the cash.

Therefore, for these reasons, it’s more likely than not that there will be AMD cards for AS Macs, which means an eGPU will be supported even in a Hackintosh type way in the future, since lack of drivers is the only thing holding support back.

We will see. From developer's standpoint, using AMD GPUs for AS Macs would be a big step back. It would break down the very convenient programming model that Apple has implemented so far and it would sabotage many pro applications. Unified memory is a big thing, and I don't see how they can have it in combination with a third-party GPU. Unless they have AMD develop one for them in close collaboration with Apple CPU/system's team. At this point it's easier and cheaper for Apple to build one themselves — as I mentioned above, they have the tech for that...

 

[torncanvas]

Valid counter argument. Intel has the same logic based on scaling up the Iris Xe, which reportedly has mediocre results. But Intel isn’t Apple and certainly Apple has a much better track record with GPUs.

I’m confident they can make a great 50W GPU, it’s more the 250W-350W range that I’m less confident in. They need to hit over 30, probably around 40 TFLOPS of shading horsepower in order to be competitive with the high end. If they can’t hit it, then it’d be best to use AMD.

It just seems like a lot to take on in addition to the other challenges. But they’re a big company so maybe they’re up to that challenge.

 

[ArPe]

eGPU support may or may not come. Apple Insider has a story that people managed to get an eGPU detected in System Information, but that's a far cry from actual support.

Thats no surprise. Any device attached to the Thunderbolt will be recognised, just the driver isn’t there. I have an eGPU attached to my Pro M1. I’m currently left to using it for the 4 USB ports and charging.

The AMD GPU will need drivers. It won’t be simple port of the x86 driver as Apple has to manage the card’s VRAM differently because the M1’s Unified Memory likes to handle all graphics and video data by default.

 

[ArPe]

Valid counter argument. Intel has the same logic based on scaling up the Iris Xe, which reportedly has mediocre results. But Intel isn’t Apple and certainly Apple has a much better track record with GPUs.

I’m confident they can make a great 50W GPU, it’s more the 250W-350W range that I’m less confident in. They need to hit over 30, probably around 40 TFLOPS of shading horsepower in order to be competitive with the high end. If they can’t hit it, then it’d be best to use AMD.

It just seems like a lot to take on in addition to the other challenges. But they’re a big company so maybe they’re up to that challenge.

Apple guys said in interview last week that users shouldn’t think about specs like TFLOPs and MHz because the unified architecture and dedicated chips make the system very different to compare against the traditional way. They said to focus on app performance.

The same could be said of when SGI implemented Unified Memory Architecture in their workstations in the late 90s. The O2 had amazing graphics performance at the time. Their Visual Workstation line implemented UMA on Intel chips by heavily modifying firmware and Windows HAL. The performance of the SGI VW couldn’t be compared one to one against a standard PC with the same CPUs and GPUs.

 

[leman]

I’m confident they can make a great 50W GPU, it’s more the 250W-350W range that I’m less confident in. They need to hit over 30, probably around 40 TFLOPS of shading horsepower in order to be competitive with the high end. If they can’t hit it, then it’d be best to use AMD.

To match something like a MI100, a 60-core Apple GPU would do. Then again, the question is whether they would want or need to build that kind of a GPU. Apple is not in a server business and they are not in a supercomputer business. I think we need to look less at specs and more that what Apple products are supposed to do. For example, there is a lot of focus on machine learning with modern GPUs, with dedicated matrix multiplication units across different precision ranges, but Apple already has this with the Neural Engine (actually, this bit is weird since Apple currently has matrix multiplication acceleration on CPU, GPU and Neural Engine). And since they use unified memory, they can utilize all different processors simultaneously without any latency or performance loss.

For example, will Apple ever introduce double-precision support on their GPUs? I don't think so. This is not an area where their products are used, so why would they waste die space on it? For heavy-duty scientific computation, you probably want a supercomputer anyway. And for some occasional double-precision work, you can either use the CPU (still hoping for that SVE support) or emulate extended precision using FP32 on the GPU.

But for more traditional Mac tasks like video rendering etc., a small Apple GPU is likely to outperforms a large (and much faster on paper) Nvidia or AMD GPU. Unified memory and software integration trump pure performance in this domain.

 

[Frank Philips]

About eGPU, well, Apple seem to consider that their graphics co-processor is enough.
On that point, time will tell.
We'll have to wait for independent benchmarks with GPU-hungry pro apps (when they'll be available) to know for sure.

Am I the only one to think that Apple will rather equip its higher-end computers (hi-end iMac, Mac Pro...) with an accelerator a la AfterBurner?

 

[Lammers]

I kind of wonder whether Apple’s support for eGPU in macOS, and its partnership with BlackMagic to build one, was an experiment to see what the take-up would be in the market. And given that there haven’t been any follow-ups to the BlackMagic eGPUs (and the Pro model was discontinued earlier this year without being replaced) we could infer that it didn’t sell well. And so we could further infer that the eGPU experiment might not have been a success. And in that case, it doesn’t take a stretch of the imagination to think that Apple might just jettison the feature with the M1 Macs.