Thanks to findings of okrasit and Fl0r!an, I was intrigued to write a small script that unleashes the full power of R9 Nano and RX 480 for calculation tasks on macOS 10.12.2. Nearly doubled clpeak single-precision compute results (GFLOPS) and Indigo Renderer Benchmark score.
The script is written only for experimental and educational purposes, I don’t take any responsibility if something goes wrong, and since this is a binary hack, I don’t see any continuity.
A word of warning, please do not edit PP_DisablePowerContainment key. It doubled OpenCL float4-float16 OpenCL computing performance with the RX 480, but it can also fry your Akitio. The total power consumption of RX 480 eGPU exceeded 190W! (I use custom boards).
http://www.insanelymac.com/forum/topic/313977-r9-nano/?p=2332854
https://www.tonymacx86.com/threads/enable-all-cores-r9-fury-cards.209892/#post-1393445
RX 480 (clpeak):
Device: AMD Radeon HD Baffin Unknown Prototype Compute Engine
Driver version : 1.2 (Dec 9 2016 21:43:55) (Macintosh)
Compute units : 36
Clock frequency : 1266 MHz
Global memory bandwidth (GBPS)
float : 202.40
float2 : 212.20
float4 : 215.84
float8 : 129.97
float16 : 60.01
Single-precision compute (GFLOPS)
float : 5644.43
float2 : 5634.59
float4 : 5608.17
float8 : 5574.38
float16 : 5517.69
R9 Nano (clpeak):
Device: AMD Radeon HD Baffin Unknown Prototype Compute Engine
Driver version : 1.2 (Dec 9 2016 21:43:55) (Macintosh)
Compute units : 64
Clock frequency : 1000 MHz
Global memory bandwidth (GBPS)
float : 402.93
float2 : 439.91
float4 : 419.14
float8 : 196.41
float16 : 109.40
Single-precision compute (GFLOPS)
float : 7180.09
float2 : 6990.68
float4 : 6820.10
float8 : 6769.77
float16 : 6657.53
chmod +x enable-baffin-CUs.sh
For R9 Nano:
sudo ./enable-baffin-CUs.sh fiji 64
For RX 480:
sudo ./enable-baffin-CUs.sh ellesmere 36
#!/bin/sh
#
# Script (enable-baffin-CUs.sh) by Goalque (goalque@gmail.com)
# Credit to okrasit and Fl0r!an:
# http://www.insanelymac.com/forum/topic/313977-r9-nano/?p=2332854
# https://www.tonymacx86.com/threads/enable-all-cores-r9-fury-cards.209892/#post-1393445
first_argument="$1"
second_argument="$2"
init_function=""
CU_count=""
pattern1="s/(\x48\xB8)(\x02)(\x00\x00\x00\x01\x00\x00\x00\x48\x89\x43\x54\xC7\x43\x7C)(\x08)(\x00\x00\x00)"
pattern2="s/(\x0F\x42\xC8)(\x89\x8B\x80\x00\x00\x00\x44\x88\xB3\x99\x00\x00\x00\x44\x88\x73\x20)"
pattern3="s/(\xE8)(\x49\x85\xFE\xFF)(\xBE\x48\x01\x00\x00\x4C\x89\xF7)"
if [[ "$first_argument" == "ellesmere" ]]then
init_function="\x46\xE4\x00\x00"
elif [[ "$first_argument" == "fiji" ]]then
init_function="\x73\x02\x01\x00"
elif [[ "$first_argument" == "baffin" ]]then
init_function="\x49\x85\xFE\xFF"
fi
if [[ "$second_argument" == 36 ]]then
CU_count="\x04\3\x12\5"
elif [[ "$second_argument" == 64 ]]then
CU_count="\x04\3\x20\5"
fi
if [[ "$init_function" != "" ]] && [[ "$CU_count" != "" ]]then
rsync -a /System/Library/Extensions/AMDRadeonX4100.kext/Contents/MacOS/AMDRadeonX4100 /tmp/AMDRadeonX4100
cat /tmp/AMDRadeonX4100 | perl -pe "$pattern1"$"/\1"$CU_count"/g" | perl -pe "$pattern2"$"/\x90\x90\x90\2/g" | perl -pe $pattern3$"/\1"$init_function"\3/g" > /tmp/AMDRadeonX4100_modified
rsync -a --delete /tmp/AMDRadeonX4100_modified /System/Library/Extensions/AMDRadeonX4100.kext/Contents/MacOS/AMDRadeonX4100
chown -R root:wheel /System/Library/Extensions/AMDRadeonX4100.kext/Contents/MacOS/AMDRadeonX4100
chmod -R 755 /System/Library/Extensions/AMDRadeonX4100.kext/Contents/MacOS/AMDRadeonX4100
rm /Volumes/Macintosh\ HD/System/Library/PrelinkedKernels/prelinkedkernel 2>/dev/null
rm /Volumes/Macintosh\ HD/System/Library/Caches/com.apple.kext.caches/Startup/kernelcache 2>/dev/null
touch /System/Library/Extensions
echo "Rebuilding caches..."
kextcache -q -update-volume /Volumes/Macintosh\ HD
echo "Ready."
else
echo "Invalid parameters."
fi
Yep, no problem. Remember that this script does the CU count patch only, nothing else. Who knows, maybe we’ll have a main course after an appetizer 😉
What is up with the cores limit exactly? Is it a Mac OS generic lock, or only locked because the Kexts are intended for different GPUs?
I take it this does not improve Metal/OpenGL performance?
Pending: Add my system information and expected eGPU configuration to my signature to give context to my posts
That’s a good question. There are a couple of values in hardware initialization functions. I guess the 16 CU limit is there because officially announced Polaris-based GPUs cannot utilize 64. Apple has quietly improved AMD drivers and added new device ids. Fiji is added purposely. This is a signal that they want to support a large variety of AMD GPUs.
I don’t know if this has an effect on shaders in OpenGL or Metal, probably not. Valley benchmark score did not seem to be affected.
As AmandTech pointed out, an AMD Compute Unit is made out of 4 SIMDs, where many CU's make the base of the GPU. Then everything is put through the Pixel Pipeline. I can't imagine those would be either limited or a bottleneck for the AMD RX48o. I would expect to see a graphics performance increase as well, but maybe the Thunderbolt 2 connection is limiting the GPU in heavy graphic computations.
I will run some benchmarks with both Metal as OpenGL, where Metal really has a huge performance increase for eGPUs (some games over 3 times the performance), probably because of a smaller bandwidth overhead.
Very nice find Goalque!
Pending: Add my system information and expected eGPU configuration to my signature to give context to my posts
What is certain at present is that the hack has an influence on CL_DEVICE_MAX_COMPUTE_UNITS value in OpenCL’s clGetDeviceInfo call. The R9 Nano has 64 ROPs/64 CUs, and RX 480 has 32 ROPs/36 CUs.
https://www.techpowerup.com/gpudb/2735/radeon-r9-nano
https://www.techpowerup.com/gpudb/2848/radeon-rx-480
AFAIK, TB2 doesn’t bottleneck much GPGPU tasks. I’m looking forward to your results!
I will run some benchmarks with both Metal as OpenGL, where Metal really has a huge performance increase for eGPUs (some games over 3 times the performance), probably because of a smaller bandwidth overhead.
Yeah, this is the kind of news i love to read!
MacBook Pro 2018 Touch Bar i7 quad-core 2.7Ghz - 16GB RAM - 512GB PCIe SSD
my Mantiz Venus extreme mod with Sapphire Nitro+ RX Vega 64
I took the time to benchmark some games with the unlocked CUs, it really made no difference at all neither negative nor positive. Tested Metal en OpenGL games/benchmarks, so I think the limitation of the CUs was only noticeable in computing performance itself. This is still a great mod for people doing GPU based 3D rendering or other GPU intensive computations.
Pending: Add my system information and expected eGPU configuration to my signature to give context to my posts
Intriguing! Were your Indigo tests with the Nano or 480? (sorry for bumping)
Pending: Add my system information and expected eGPU configuration to my signature to give context to my posts
I've executed the same test several times on my early 2013 15" rMBP + Node + Sapphire 4GB RX 480. The figures are largely the same:
- 2184...2247 (Bedroom) with the same eGPU also being the current display driver (in 4k60),
- 1628, 1628, 1629 (still Bedroom; three tests with the eGPU being NOT the current one but driving the 4k monitor from the dGPU)
- 7136...7213 in five cases and 6172 in one case (this NOT being the card non-current) (Supercar)
I've only tested the CPU once: 0.391 with Bedroom. The 650M tests weren't finished in over 30 minutes (they didn't even get to stage 1) so I stopped them (tried three times).
Incidentally, these figures are both somewhat larger than the ones (1.928, 6.261) at https://www.indigorenderer.com/benchmark-results?page=1&sort_desc=result_bedroom&filter=highlighted . This means these are pretty impressive results and also show
- even the TB1 bandwidth poses no problems for strictly OpenCL usage,
- not even when the external display is also driven using the same card.
Pending: Add my system information and expected eGPU configuration to my signature to give context to my posts
Hi @goalque.
I plugged in my WX5100 and tried to enable its 28 CU. Failed. LuxMark reports 14...
I started by editing the script and adding this option to it (ran it with sudo ./enable-baffin-CUs.sh ellesmere 36)
elif [[ "$second_argument" == 28 ]] then CU_count="\x04\3\x0E\5" fi
but not luck. Any suggestions?
Thanks!
late-2016 13" Macbook Pro + [email protected] (TB3 to TB2 adapter) (AKiTiO Thunder2) + macOS 10.12.3
late-2016 13" MBP + Quadro M2000/FirePro WX5100/GTX750-TB3 (AKiTiO Thunder3) + macOS 10.12.3
(Custom Thunder cables)
It seems that after reboot I do have 28 cores reported, but the scores are lower... Hmm...
late-2016 13" Macbook Pro + [email protected] (TB3 to TB2 adapter) (AKiTiO Thunder2) + macOS 10.12.3
late-2016 13" MBP + Quadro M2000/FirePro WX5100/GTX750-TB3 (AKiTiO Thunder3) + macOS 10.12.3
(Custom Thunder cables)
I can't help with the OS X related stuff, but how about giving it a shot in a different OS? I don't know if you have Win bootcamped, but you could try to boot Linux off a live-USB stick, load the driver and run some benchmark. Just to make sure your issue is indeed in the SW config of things.
EDIT: It looks to me that the WX 5100 gets identified as a RX 460, and seems to be used as such. 14 is the number of CUs in the cut-down Baffin dies used for RX 460s.
Want to output [email protected] out of an old system on the cheap? Read here.
Give your Node Pro a second Thunderbolt3 controller for reliable peripherals by re-using a TB3 dock (~50$).
"Always listen to experts. They'll tell you what can't be done, and why. Then do it."- Robert A. Heinlein, "Time Enough for Love."
The IT Sage referenced this script in his blog showing how install a 470/480 in my 2010Mac Pro 5,1. My question is will this script work or can it be modified for the 470 which is has 32 CUs?
tia
Bill
Pending: Add my system information and expected eGPU configuration to my signature to give context to my posts
Might be the dumb one in the room, but after enabling this, my Mac Pro will immediately power down whenever I try to run a Geek Bench or touch 4K footage in Premiere - was totally fine with both the above before this. Any ideas how to DISable the Baffin CU script? Or get it working?
Pending: Add my system information and expected eGPU configuration to my signature to give context to my posts
Wow thank you! That works great in my MacPro 5.1 and my RX480. I would like to put a RX460 which lies around in there too. But is that possible? Or does that Compute Unit Hack mean the RX460 would not work? Thanks
Pending: Add my system information and expected eGPU configuration to my signature to give context to my posts
This sounds like what I need. How do I execute this script?
Pending: Add my system information and expected eGPU configuration to my signature to give context to my posts
RX480 and RX460 do work parallel in MAcPro 5.1. DaVinci Resolve ist faster. Final Cut Pro X is glitchy 🙁
Pending: Add my system information and expected eGPU configuration to my signature to give context to my posts
Maybe this is a dumb question. Do I have to restart after running the script? My clpeak reports for my RX 480:
Device: AMD Radeon HD Baffin Unknown Prototype Compute Engine
Driver version : 1.2 (Dec 9 2016 21:43:55) (Macintosh)
Compute units : 16
Clock frequency : 1266 MHz
You can see it says 16 CUs.
By the way, would this script improve both graphical and compute performance? Edit: I read through the thread more closely.
Pending: Add my system information and expected eGPU configuration to my signature to give context to my posts
Yes. Restart your Mac and you’ll see 36 compute units.
I find that performance of the GPU is much better in Windows. Games have higher FPS.
I tried searching about performance but didn't find much comment. Upon restarting, the number of CUs went up to 36, but as previously discussed, graphical performance didn't improve much.
Is it because the OSX drivers that Apple writes for the GPUs are not up to date?
Pending: Add my system information and expected eGPU configuration to my signature to give context to my posts
How do I actually make use of the hack? Just put the commands into Terminal? Sorry for the newbie questions.
I have a RX460 and also a RX580 I want to test this hack with.
Pending: Add my system information and expected eGPU configuration to my signature to give context to my posts
How do I actually make use of the hack? Just put the commands into Terminal? Sorry for the newbie questions.
I have a RX460 and also a RX580 I want to test this hack with.
You don’t need this script for RX 460. It’s beneficial for Ellesmere and Fiji GPUs such as the RX 580 and R9 Fury in macOS 10.12 only.
To make use of the enable-baffin-CUs.sh script, navigate to the file location (assuming you saved or downloaded it in ~/Downloads folder), bless it with executable permission then run with these commands in Terminal:
cd ~/Downloads chmod +x enable-baffin-CUs.sh sudo ./enable-baffin-CUs.sh ellesmere 36
Might be the dumb one in the room, but after enabling this, my Mac Pro will immediately power down whenever I try to run a Geek Bench or touch 4K footage in Premiere – was totally fine with both the above before this. Any ideas how to DISable the Baffin CU script? Or get it working?
This happens when the GPU gets insufficient amount of juice. It didn’t crash your Mac Pro when only 16 CUs were running. When all 36 CUs are running, the card pulls a lot more power. I’d recommend using both PCIe booster power ports with a dual 6-pin to 8-pin adapter.
Is there any use in using this script in 10.13?
Pending: Add my system information and expected eGPU configuration to my signature to give context to my posts
Are you sure this script is not required on macOS 10.12.6? Maybe I don't remember right but when I had set up my eGPU on this release, I needed the patch for all CUs. Will test and check again.