Afaik the Intel Integrated GPUs either have a hard limit on the addressable RAM or are capped to half the available RAM.
I don’t know of any AMD G series equipped laptops with that much memory , but admittedly I haven’t looked closely at the options.
Then there’s the issue of memory sharing of actual data resources. Albeit, this is down to Software, but more software for Mac can assume shared memory to take advantage of it, versus other brands because iGPUs have historically been very limited.
The AMD laptops don't need to be in the G series to have an iGPU - they all do. The G series only exists for desktops.
> Albeit, this is down to Software, but more software for Mac can assume shared memory to take advantage of it, versus other brands because iGPUs have historically been very limited.
AFAIK there is no such software that only supports Macs. You can query these features at runtime, and it's easier to do so than to rewrite your renderer for Metal.
Ah yeah I always get tripped up by AMDs offerings.
I didn’t say the software needed to only support Mac, but if they have a metal backend (as many things have multiple backends) they know they can spend the time to get a bigger ROI due to the number of shared memory macs as a percentage of all macs.
...Not exactly and this is a silly comparison anyways
Apple dynamically scales the GPU memory while Windows GPU memory must be reserved. If you have 64GB of Ram and want 32GB of VRAM you now...always.... have 32GB of system memory.
But furthermore...that iGPU is like 1/10th the performance of an M1 Max lol, absolutely nobody should EVER do this on an iGPU. Their argument is that on a dedicated card you just can't do this.
It is not true that iGPU memory always has to be reserved anymore. Some programs and APIs don't support it, but you can address main system memory from many iGPUs and vice-versa. In Vulkan for example you can do this by setting the HOST_COHERENT, HOST_VISIBLE and DEVICE_VISIBLE for a buffer. Of course, this is subject to driver bugs, and you have to be careful with caches that may or may not be shared depending on the specific iGPU.
Beyond this, newer iGPUs like the RX680M are comparable in performance to an M1 Pro. Certainly sufficient for any model visualization task, since that's what we're talking about.
In those applications you're likely to be limited by raster performance, and I can't find any metrics for that. TFlop for TFlop, the 780M which is the main competitor should be within 10-30%, but that is not relevant for this application.
The only other application where this could be relevant would be 3D rendering, where the 780M would win by a large margin due to RT acceleration.
