2016 15" MacBook Pro (RP450) [6th,4C,H] + RX Vega 56 @ 32Gbps-TB3 (Mantiz Venus) + macOS 10.13 & Win10 [itsage]
After multiple delays, I finally got an RX Vega 56 through Best Buy. I chose this GPU because the RX Vega 64 seems to be quite a handful to handle for most eGPU enclosures on the market. It's also a direct competitor to a GTX 1070 which I have to test side-by-side.
In macOS High Sierra, it was plug-and-play. The RX Vega 56 was identified as R9 XXX. I changed the controller file inside AMD1000Controller.kext so that About this Mac shows the right card name. This is purely cosmetics and not necessary. The initial performance impression is that it's slightly faster than the R9 Fury. OpenGL is not working well in macOS. When the iMac Pro launch in December, the drivers will be a lot more optimized.
In Windows 10 Bootcamp, I got this eGPU going the same way I did with previous implementations [Bootcamp eGPU setup guide]. I needed to disable both the x16 PCIe Controller to the dGPU and x8 PCIe Controller to the left side TB3 ports in order to resolve error 12. AMD XConnect works great for internal display acceleration. At idle, the RX Vega 56 is really quiet. When it's loaded up, it gets loud but not intolerable.
Wow. Can you confirm issue on macOS current beta? Is Anti aliasing the cause? Have you tried a metal game?
barefeats has published an article
Thank you for being one of the first to publish a much anticipated AMD Vega eGPU implementation on a Macbook.
Now that you are using AIDA64 GPGPU benchmarking, which incidentally shows your good memory write (H2D equivalent) result, I'd be curious as to a comparative result of those AIDA64 AMD Vega54 GPGPU (OpenCL) results against Nvidia GTX1070/80/80Ti card.
Will this show AMD cards are still computationally superior to Nvidia and hence preferred by professionals and crypto miners?
@Nando4 there's something not right with the GTX 1080 Ti OpenCL number. It should be at least 200,000. I also tried this RX Vega 56 in my Mac Pro tower. The OpenCL number is very close to that of the RX Vega 64 from @barefeats' results.
@itsage, given macOS NVidia Pascal drivers are at beta and AMD drivers are still being developed, a Windows run of Geekbench would give a better computational comparison of those cards. Windows drivers are now mature though years of work.
I will receive my RX Vega 56 in few days! So I can help if needed
I was looking for the code inside the box to register for the Radeon Pack (2x included games). When I looked over the packaging tonight, I noticed the RX Vega 56 labels seem to be stickers. I carefully peeled them back and saw 64 underneath. This is similar to the packaging of the RX 480 last fall. Due to limited supply, the RX 480 4GB reference cards were actually 8GB cards with only 4GB enabled.
I checked using GPU-Z to confirm this is a Vega 56 (3584 shaders/64=56 compute units). I proceeded to flash it with the Vega 64 vBIOS. While the clock and memory speed improved, the shaders stayed at 3584.
The latest Crimson drivers brought some nice additions too. Look at what I found in AMD Radeon Settings. I undervolted the GPU and got a nice performance boost. These were ran in XConnect internal display mode.
@cvfunstuff, drivers for RX Vega in High Sierra is no where near where they should be. The GM candidate came out this afternoon but not much has changed for RX Vega cards.
Crimson drivers are getting better in Windows. The key is to undervolt these cards to get higher performance out of them. I set the GPU Clock at 1602 MHz/1020 mV for Stage 6 and 7. Memory clock is set at 950 MHz/950 mV. This yields almost a 10% improvement over stock settings.