eGPUs

NickK

NickK

NickK

Soldato
Joined
13 Jan 2003
Posts
24,820
My old 2011 MBP has had a replacement logic board, now one of the board's memory channels (or the socket) has died so I'm down to a slow 8GB. The Mrs suggested I get a new mac with the bonus.

I don't game but I have and would use a high power GPU for image processing and programming for astrophotography. For programming the data is stored inside the GPU and displayed from the GPU typically, with the computer just uploading and orchestrating the execution of the GPU kernels (programs). Although numerical accuracy with GPUs has never been their strong point (slow IEEE FP support) they're still faster than a CPU.

I've seen the black magic's eGPU however the GPU is soldered and non-upgradable. I see there are other designs.

Therefore I'm considering a eGPU enclosure with a AMD RVII based GPU card with 16GB graphics memory. From my GPU coding (going back to 2006) I'm clear that it would be beneficial to have 32GB in the laptop to ensure shadowed/mapped textures don't result in swap space use - although for the last 3 years I've not done much due to the dying MBP. In terms of CPU power a i7 or i9 wold be good.

I'm aware that having an enclosure really needs a beefy PSU for both the MBP and the GPU.

So any recommendations?

edit: RVII not R9.
 
Last edited:
I'm currently using a Sonnet eGPU breakaway box and would recommend it. I have the 550w version so it can charge the MBP as well.

They also have a 350W version but you won't be able to charge the MBP during use. You could always switch the PSU in them as they use SFX power supplies.

Be aware that if you plan to use it in bootcamp it's not a simple plug and play.

https://www.imore.com/how-set-your-egpu-preferred-rendering-device-macos-mojave
 
momo56 said:
I'm currently using a Sonnet eGPU breakaway box and would recommend it. I have the 550w version so it can charge the MBP as well.

They also have a 350W version but you won't be able to charge the MBP during use. You could always switch the PSU in them as they use SFX power supplies.

Be aware that if you plan to use it in bootcamp it's not a simple plug and play.

https://www.imore.com/how-set-your-egpu-preferred-rendering-device-macos-mojave
Click to expand...

Yup - was thinking of a Sonnet with a Radeon VII. So I think it would be the 650W version I'd need.

Are you using a small 13" MBP?

edit: seems this guy is using one: https://egpu.io/forums/builds/mid-2...2gbps-tb3-razer-core-macos-10-14-4-theitsage/
 
Last edited:
Interesting topic this.

I currently have a base spec 13" Pro but i'm liking the idea of replacing both it and my home machine with a 4 or 6 core Macbook Pro and using my Vega 64 in a eGPU.

Trouble is it's a fair amount to put down on something that might not work well for the odd bit of gaming (World of Warships, Apex Legends, Project Cars 2 etc) and should I go ahead, I can't decide whether the the 6 core is worth the extra over the quad and if so, whether the fact it has a dGPU onboard might make using an eGPU more difficult.
 
Interesting article here: http://timdettmers.com/2018/12/16/deep-learning-hardware-guide/

The issue used to be that OpenCL and the original GPGPU use of textures required the data be loaded into the texture pinned buffer in system ram then copied into GPU ram before the kernel was executed. With large data sets the memory of the host machine disappeared rapidly.

For a GPU the PCIe bus was designed for fast upload from system to GPU memory, however downloading from textures back to the host buffer took time so most of the time the less fetching of data back, the better. In the system I wrote - I loaded image frames into the GPU continuously, processed and then rendered the output directly into the GPU output in a window.

For the purpose I want, a 32GB GPU would be hampered by 16GB host memory (only using for unpinned, ie interim results etc perhaps). It's very annoying that the i7 15" is limited to 16GB and the i9 has thermal issues (32GB) given it ends up 3K of machine then add eGPU. I'm thinking that a 13" MBP with a eGPU would probably address the issues for me (and not fry the mac). However I know 16GB disappears quickly and would hamper the GPU. I've had QC i7, 16GB of system memory and 1GB of VRAM since 2011.. I may end up waiting for the next refresh to see. Come on Apple!
 
What is tasty is the Mac mini can be upgraded to 64GB and is a decide price. Mini+VII eGPU would be hardcore and costs less than a Mac laptop.
 
Plenty of user eGPU setups can be found here : https://egpu.io

I have been tempted by eGPU numerous times, I guess I'm just waiting for an AMD card worth buying :) (to ensure macOS compatibility)
 
shiver said:
Plenty of user eGPU setups can be found here : https://egpu.io

I have been tempted by eGPU numerous times, I guess I'm just waiting for an AMD card worth buying :) (to ensure macOS compatibility)
Click to expand...

Noted a couple of references to egpu.io but not really looked more than google hits. However I just noted this piece: https://egpu.io/setup-guide-external-graphics-card-mac/ Look down at the difference between the bridge vs CPU attachment of the T3 bus. The bridge strangles with SATA and other attached bus traffic.

Will check the chipset of the mini.. it has T3 attached via the bridge, not the CPU like the lcelake chipset.
 
Last edited:
Just found out the 2018 Mac mini i7 is miles faster than my 2011 MBP i7. Multicore bench mark of 9K vs 23K.

Not even going into the VII!

In the past I’ve used clFFT on GPU in real-time image alignment and FFTW on cpu. No metal FFT yet. clFFT is a bit slow in fixing bugs etc so i’m not sure if Apple will make their own library tools.

FFTs are typically code generators that take the FFT and input/outputs required and generate an execution plan. Then each time the fft is needed the pregenerated code is used.

Seriously tempted by a mini + eGPU as I can get a better machine for what I want for cheaper.
 
You must log in or register to reply here.
Back
Top Bottom