Jump to content
euqlaog

2014 15” MBP Iris + 2X R9_280X@16Gbps-TB2 (Netstor NA211TB) + Win8.1/OSX [goalque]

Recommended Posts

I didn’t found any TB2 implementation utilising AMD card. Maybe because of the simple reason; Nvidia cards consume less power, are more quiet and generally faster in desktop computers, but what benefits AMD could give eGPU TB2 users? I wanted to find out and took a step further with the Netstor NA211TB as I saw a couple of used cards on the shelf of local computer store:

- MSI Radeon HD7970 3GB GDDR5 OC (1010Mhz)
- Asus Radeon R9 280X DirectCU II TOP 3GB GDDR5 (1070Mhz)

 

post-28870-14494999038249_thumb.jpgpost-28870-14494999038731_thumb.jpg

 

434€ for both of the cards.

Other components:

- Apple MacBook Pro 15” Retina Iris Pro, 2.2GHz Intel Core i7 (Mid 2014)
- Samsung 28” UHD LED Monitor (U28D590D)
- Display Port cable
- Thunderbolt cable
- ATX PSU, 450W-1000W depending on how many eGPUs you are going to run (up to 3 eGPUs is possible)

Will it Crossfire? R9 280X & HD 7970 Scaling Tested

I attached the cards in x8 and x4 slots of the NA211TB PCIe board, and very soon realised that I made the first ever Crossfire Thunderbolt eGPU implementation. @squinks, you inspired me to go this point as you succeeded with two Sonnet SEL enclosures and SLI. By choosing AMD, it turned out to be much easier project.

It worked out of the box: take the PCIe board off, use a powerful enough ATX PSU and do the cable modification as described here, then just plug in the cards install latest drivers from AMD.

Apple Boot Camp Software Graphics Drivers (release date 12/16/2014)

External monitor works always and both GPUs will be detected, no need for manual delay methods. Does the internal screen work? Yes, if you follow Evo*’s instructions here. No Optimus of course, but AMD provides something similar because the apps in the retina screen are accelerated as shown in the video. You can move windows normally between the screens. It seems that you have to plug in display cable after logged in Windows if you are outputting signal via 280X (SLOT1), otherwise the Intel Iris Pro shows error code. Mini DP - DP with the HD 7970 (SLOT3) should detect an external UHD monitor automatically. On OS X screen detection is completely plug-and-play.

One more thing, and for me the most important - on OS X both the cards enabled the external monitor via eGPU and accelerated OpenGL apps. This wasn’t possible with Nvidia cards. It works with the default AMD kexts that Apple provides, you don’t need additional web drivers. Just a couple of IOPCITunnelCompatible true modifications.

https://github.com/goalque/automate-eGPU
http://forum.techinferno.com/mac-os-...output-18.html

The procedure below is not up to date. Use automate-eGPU.sh script for both AMD and Nvidia cards

 

Spoiler

 

1. Open Terminal

2.
HTML Code:

sudo nano +287,32 /System/Library/Extensions/AMD7000Controller.kext/Contents/Info.plist

 

3. Press Enter in Nano and add the following lines before "/dict"

HTML Code:

IOPCITunnelCompatible

 

4. Press “Ctrl+X” and “Y”

Repeat steps 3-4 for AMDSupport and AMDRadeonX4000 kexts as well



 
HTML Code:

sudo nano +58,53 /System/Library/Extensions/AMDSupport.kext/Contents/Info.plist

 

HTML Code:

sudo nano +289,33 /System/Library/Extensions/AMDRadeonX4000.kext/Contents/Info.plist|

 

Finally,


 
HTML Code:

sudo kextcache -system-caches

 

and restart Mac.

If everything goes well, you should see the kexts in the list of loaded extensions.
(About This Mac -> System Report -> Software -> Extensions). AMDFrameBuffer.kext doesn’t need modification.

The previous instructions works for Radeon HD 7900 series cards (Tahiti), and line numbers given as parameter to Nano are meant for OS X 10.10.1 and Mid 2014 15" MBPr Iris Pro. Note, that kext content might be different depending on the OS X version and hardware. Apple provides native AMD drivers for these series:

AMD2400Controller
AMD2600Controller
AMD3800Controller
AMD4600Controller
AMD4800Controller
AMD5000Controller
AMD6000Controller
AMD7000Controller
AMD8000Controller
AMD9000Controller

If something goes wrong, look at the Console errors after latest “bootlog” line. For example, if you happen to have 270X (not codename Tahiti), you will see this line in the log:

kernel: Driver “AMDPitcairnGraphicsAccelerator” needs “IOPCITunnelCompatible” key in plist

Meaning that you have to add IOPCITunnelCompatible true value for a block AMDPitcairnGraphicsAccelerator in AMDRadeonX4000.kext. If you let this kext untouched, you might get screen output, but it will be very slow.

Instead using Console you can type the following to search missing IOPCITunnelCompatible places:



 
Code:

grep IOPCITunnelCompatible /var/log/system.log

 

 

 

 



3DMark11: 17463 (graphics)
3DMark Fire Strike: 12244 (graphics)

Unigine Valley Benchmark 1.0:

Windows 8 (build 9200) 64bit/Direct3D11/1920x1080 8xAA fullscreen/Extreme HD: 2998 (FPS: 71.7, Min FPS: 25.9, Max FPS: 145.5)


Clpeak (R9 280X):

 

Spoiler

Device: AMD Radeon HD Tahiti XT Prototype Compute Engine
Driver version : 1.2 (Sep 28 2014 22:27:20) (Macintosh)
Compute units : 32
Clock frequency : 1070 MHz

Global memory bandwidth (GBPS)
float : 232.45
float2 : 240.18
float4 : 245.28
float8 : 129.71
float16 : 66.36

Single-precision compute (GFLOPS)
float : 3870.20
float2 : 3847.64
float4 : 3837.57
float8 : 3817.81
float16 : 3776.48

Double-precision compute (GFLOPS)
clCreateKernel (-46)
Tests skipped

Transfer bandwidth (GBPS)
enqueueWriteBuffer : 1.36
enqueueReadBuffer : 1.39
enqueueMapBuffer(for read) : 24.80
memcpy from mapped ptr : 5.57
enqueueUnmap(after write) : 2925.73
memcpy to mapped ptr : 6.43

Kernel launch latency : 8.66 us


 


Clpeak (HD 7970):

 

Spoiler

Device: AMD Radeon HD Tahiti XT Prototype Compute Engine
Driver version : 1.2 (Sep 28 2014 22:27:20) (Macintosh)
Compute units : 32
Clock frequency : 1010 MHz

Global memory bandwidth (GBPS)
float : 213.96
float2 : 219.56
float4 : 223.81
float8 : 118.71
float16 : 60.96

Single-precision compute (GFLOPS)
float : 3039.93
float2 : 2980.50
float4 : 2982.89
float8 : 2984.26
float16 : 2978.57

Double-precision compute (GFLOPS)
clCreateKernel (-46)
Tests skipped

Transfer bandwidth (GBPS)
enqueueWriteBuffer : 1.38
enqueueReadBuffer : 1.40
enqueueMapBuffer(for read) : 24.83
memcpy from mapped ptr : 5.59
enqueueUnmap(after write) : 3021.22
memcpy to mapped ptr : 6.44

Kernel launch latency : 8.21 us


Clpeak (Iris Pro):

 

Spoiler

Platform: Apple
Device: Iris Pro
Driver version : 1.2(Sep 25 2014 22:25:51) (Macintosh)
Compute units : 40
Clock frequency : 1200 MHz

Global memory bandwidth (GBPS)
float : 43.21
float2 : 51.03
float4 : 52.03
float8 : 25.53
float16 : 12.21

Single-precision compute (GFLOPS)
float : 514.61
float2 : 750.63
float4 : 753.93
float8 : 751.64
float16 : 746.71

No double precision support! Skipped

Transfer bandwidth (GBPS)
enqueueWriteBuffer : 8.80
enqueueReadBuffer : 10.44
enqueueMapBuffer(for read) : 1044.09
memcpy from mapped ptr : 5.77
enqueueUnmap(after write) : 4407.81
memcpy to mapped ptr : 6.40

Kernel launch latency : 30.61 us


https://github.com/krrishnarraj/clpeak

 

 

More photos:

 

Spoiler

post-28870-14494999039327_thumb.pngpost-28870-14494999039531_thumb.pngpost-28870-14494999039066_thumb.pngpost-28870-14494999039736_thumb.pngpost-28870-14494999040537_thumb.pngpost-28870-14494999041149_thumb.png

Share this post


Link to post
Share on other sites

That looks good, but I have some questions. Do you have any comparison of one GPU versus two GPU in CrossFire?

Because normally both GPU support X16 lanes of bandwidth, Thunderbolt 2 only offers X4 bandwidth, so the bandwidth of the first card is downgraded from X16 to X4.

Then your second card, runs on X1? Is this even possible?

I'm very curious about the comparison as I said above. Otherwise, would it be possible to use two Thunderbolt connections, and two Netstors/or whatever. And then CrossFire them together, so you will have 2x X4 bandwidth, right?

I'm just throwing my thoughts :)

Share this post


Link to post
Share on other sites
That looks good, but I have some questions. Do you have any comparison of one GPU versus two GPU in CrossFire?

Because normally both GPU support X16 lanes of bandwidth, Thunderbolt 2 only offers X4 bandwidth, so the bandwidth of the first card is downgraded from X16 to X4.

Then your second card, runs on X1? Is this even possible?

I'm very curious about the comparison as I said above. Otherwise, would it be possible to use two Thunderbolt connections, and two Netstors/or whatever. And then CrossFire them together, so you will have 2x X4 bandwidth, right?

I'm just throwing my thoughts :)

GPU-Z Bus interface values are incorrect in the pictures. When I clicked the question mark it said:

"Modern graphics cards come with extensive saving features. One of these functions reduces the PCI-Express link speed & width to lower levels to conserve power when the card is idle. This is why you might see undesired values in GPU-Z's Bus Interface readout."

When I did the render test, it showed x16@x8 2.0 and x16@x4 2.0. And by the way AMD catalyst said PCI-E 2.0, and x8 and x4 as it should be. The slot where MSI card is attached, is electrically x8. However, TB2 restricts the speed to ~x4. squinks was able to double the 3DMark11 score with two TB2 cables. But I have been thinking, is it really 32Gbps? I have to repeat the tests with a single card. If I remember correct, 3DMark11 graphics scores were something near 8000, so I got over the double score by using two GPUs in CrossFire, and a single TB2 cable. Or is it because of using two Crossfire bridges? I will confirm this.

I still have the AKiTiO box, but using two TB2 cable setup did not work perfectly in Win8.1 - the other card showed error code 35. On OS X 10.10.1 both cards worked, but you can utilise both of them only in OpenCL apps, and only one in OpenGL apps. Switching the display cable between cards is possible, and you can use multiple monitors.

EDIT: 3DMark11 Graphics scores for one card were ~9000, not 8000.

Share this post


Link to post
Share on other sites

goalque,

It should still be max X4. The benchmark results MUST be incorrect. Practically, and theoretically it can be x4 max. Even with bridges, the resultst should be the same. Because bandwidth is still going through one TB2 connection. Only with two TB2 connections you should be able to get better performance.

Or am I stupid?

Share this post


Link to post
Share on other sites
goalque,

It should still be max X4. The benchmark results MUST be incorrect. Practically, and theoretically it can be x4 max. Even with bridges, the resultst should be the same. Because bandwidth is still going through one TB2 connection. Only with two TB2 connections you should be able to get better performance.

Or am I stupid?

I agree with the x4 speed, you brought very good aspects thank you! But just my thought if those two TB ports are sharing the same Bus or something and perform the same as when daisy chained. I might be wrong, maybe two TB2 has indeed effect. Benchmark results are valid, I suppose - 3DMark11 showed over 100FPS occasionally. I can do some other performance test as well if someone is interested. I don't know what are the real desktop scores of the 2 x 280X cards in CrossFire mode.

EDIT:

Here are the scores for single cards:

8729: AMD Radeon HD 7970 video card benchmark result - Intel Core i7 4702HQ,Apple Inc. Mac-3CBD00234E554E41

9149: AMD Radeon R9 280X video card benchmark result - Intel Core i7 4702HQ,Apple Inc. Mac-3CBD00234E554E41

Share this post


Link to post
Share on other sites
I agree with the x4 speed, you brought very good aspects thank you! But just my thought if those two TB ports are sharing the same Bus or something and perform the same as when daisy chained. I might be wrong, maybe two TB2 has indeed effect. Benchmark results are valid, I suppose - 3DMark11 showed over 100FPS occasionally. I can do some other performance test as well if someone is interested. I don't know what are the real desktop scores of the 2 x 280X cards in CrossFire mode.

The 13/15" Macbook Pro systems use a dual-port Intel® DSL5520 Thunderbolt2 controller. REF: https://www.ifixit.com/Teardown/MacBook+Pro+15-Inch+Retina+Display+Late+2013+Teardown/18696

The DSL5520 has 4 Thunderbolt channels, 2 across each port. Each channel is capable of 10Gbps data transmission. Thunderbolt2 allowing 2 channels to be aggregated to one thus providing 20Gbps. However, the electrical PCIe link is x4 2.0 so that limits bandwidth to 16Gbps. So we can summarize there are 2 x 16Gbps PCIe ports available on the 2013+ Macbook Pro systems.

You can see squinks' reported at http://forum.techinferno.com/implementation-guides/7326-%5Bguide%5D-2013-13-macbook-pro-gtx780ti-sli%4032gbps-tb2-sonnet-ee-sel-win8-1-a.html#post100318 the bandwidth to each of his GTX780Ti which is consistent with x4 2.0 performance (16Gbps).

<b>CUDA-Z Bandwidth (TB Device 1 and 2)</b></span>

<span style="font-family: Arial">Host to Device: ~1250 MiB/s</span>

<span style="font-family: Arial">Device to Host: ~1360 MiB/s</span>

<span style="font-family: Arial">Device to Device: ~135 GiB/s</span>

I was advised during email communications that test was performed simultaneously, hence 32Gbps combined bandwidth.

Share this post


Link to post
Share on other sites

@Tech Inferno Fan: Some observations:

1) I remember nesone telling that there was no difference in performance when using daisy chaining / two separate TB2 ports. This is also true with the 2014 15" MBPr + NA211TB + AKiTiO on OS X and LuxMark OpenCL numbers.

2) The new Mac Pro has up to 6 Thunderbolt 2 ports but ports 1-3, 2-4 and 6-5 share the same Bus. Apple says "Attach displays to different Thunderbolt busses when possible. Do not attach more than two displays to any Bus. This means that if you use the HDMI port, be sure to then only use one of the bottom two Thunderbolt ports (Bus 0)."

Mac Pro (Late 2013): Using multiple displays - Apple Support

So it does affect what TB ports you use to get the best performance in nMP, but apparently in 2013-2014 MBPr has two separate busses. Thanks for the technical links, I wanted to be sure if MBPr is capable to 32Gbps speeds via two TB2 ports. Now it seems so, but would be great to see bandwidth numbers when both GPUs are working together. In CUDA-Z, you can choose only one GPU from the list. AMD doesn't support CUDA and 3DMark scores don't tell the whole story. HD7970@x4.2 is said to perform 94% of x16 2.0 performance. Is this true for Win7 only or Win8.1 too?

So does it mean that if I use two TB ports linked to NA211TB + AKiTiO or two AKiTiOs, and somehow resolve the code 35 error, I can get even better 3DMark11 scores than 17463? Can I improve the results by Win7? Note, that I already almost doubled the 3DMark11 graphics score with a single cable.

Share this post


Link to post
Share on other sites
@Tech Inferno Fan: Some observations:

1) I remember nesone telling that there was no difference in performance when using daisy chaining / two separate TB2 ports. This is also true with the 2014 15" MBPr + NA211TB + AKiTiO on OS X and LuxMark OpenCL numbers.

2) The new Mac Pro has up to 6 Thunderbolt 2 ports but ports 1-3, 2-4 and 6-5 share the same Bus. Apple says "Attach displays to different Thunderbolt busses when possible. Do not attach more than two displays to any Bus. This means that if you use the HDMI port, be sure to then only use one of the bottom two Thunderbolt ports (Bus 0)."

Mac Pro (Late 2013): Using multiple displays - Apple Support

So it does affect what TB ports you use to get the best performance in nMP, but apparently in 2013-2014 MBPr has two separate busses. Thanks for the technical links, I wanted to be sure if MBPr is capable to 32Gbps speeds via two TB2 ports. Now it seems so, but would be great to see bandwidth numbers when both GPUs are working together. In CUDA-Z, you can choose only one GPU from the list. AMD doesn't support CUDA and 3DMark scores don't tell the whole story. HD7970@x4.2 is said to perform 94% of x16 2.0 performance. Is this true for Win7 only or Win8.1 too?

So does it mean that if I use two TB ports linked to NA211TB + AKiTiO or two AKiTiOs, and somehow resolve the code 35 error, I can get even better 3DMark11 scores than 17463? Can I improve the results by Win7? Note, that I already almost doubled the 3DMark11 graphics score with a single cable.

You raise a very valid point here that can be easily answered. Your 15" Macbook Pro has two thunderbolt ports. The question is, when you connect to either, does the root port remain the same?

The root port will be listed at the top of the tree when do a "View devices by connection" in Device Manager. When you click on it's details it will be something like "PCI bus 0, device 1, function 1".

If it happens that there are different top level PCIe ports listed for either port, eg: port1 = "PCI bus 0, device 1, function 1" and port2="PCI bus 0, device 1, function 2", then indeed we would have two separate x4 2.0 (16Gbps) links at the top level and connecting both ports to their own Thunderbolt enclosure would give a combined 32Gbps. If the top hosting PCIe port remains the same, then there is a limitation of x4 2.0 (16Gbps) that is shared by the underlying ports.

I certainly hope Apple has provided 32Gbps. It could almost be false advertising if running both concurrently dropped each down to 8Gbps, giving only 16Gbps total.

Share this post


Link to post
Share on other sites
You raise a very valid point here that can be easily answered. Your 15" Macbook Pro has two thunderbolt ports. The question is, when you connect to either, does the root port remain the same?

The root port will be listed at the top of the tree when do a "View devices by connection" in Device Manager. When you click on it's details it will be something like "PCI bus 0, device 1, function 1".

If it happens that there are different top level PCIe ports listed for either port, eg: port1 = "PCI bus 0, device 1, function 1" and port2="PCI bus 0, device 1, function 2", then indeed we would have two separate x4 2.0 (16Gbps) links at the top level and connecting both ports to their own Thunderbolt enclosure would give a combined 32Gbps. If the top hosting PCIe port remains the same, then there is a limitation of x4 2.0 (16Gbps) that is shared by the underlying ports.

I certainly hope Apple has provided 32Gbps. It could almost be false advertising if running both concurrently dropped each down to 8Gbps, giving only 16Gbps total.

For me it looks like a jungle of PCI-to-PCI bridges :D So here some screenshots:

post-28870-14494999042125_thumb.png

post-28870-14494999041774_thumb.png

Share this post


Link to post
Share on other sites
For me it looks like a jungle of PCI-to-PCI bridges :D So here some screenshots:

[ATTACH=CONFIG]13518[/ATTACH][ATTACH=CONFIG]13519[/ATTACH]

OK.. I can see the Thunderbolt chip is attached to a single x16 root port. Now I need to find out what width they've used. The way to do that is download PCI Scope Trial from PCIScope Overview , allow it to detect hardware, Save the configuration as a .BPD file, upload it to mediafire and post a link for me to review. It will list the link width of that questionable bridge so we'll know for sure. It would need to be x8 2.0 to provide 32Gbps across two ports. If it's x4 2.0 then Apple/Intel have some explaining to do.

Share this post


Link to post
Share on other sites
on OS X both the cards enabled the external monitor via eGPU and accelerated OpenGL

Always after each restart? Because i sometimes dont get my monitor detected and i have to restart several times before it gets recognized with my NVidia. Thanks.

Share this post


Link to post
Share on other sites
OK.. I can see the Thunderbolt chip is attached to a single x16 root port. Now I need to find out what width they've used. The way to do that is download PCI Scope Trial from PCIScope Overview , allow it to detect hardware, Save the configuration as a .BPD file, upload it to mediafire and post a link for me to review. It will list the link width of that questionable bridge so we'll know for sure. It would need to be x8 2.0 to provide 32Gbps across two ports. If it's x4 2.0 then Apple/Intel have some explaining to do.

-removed-

Share this post


Link to post
Share on other sites
Always after each restart? Because i sometimes dont get my monitor detected and i have to restart several times before it gets recognized with my NVidia. Thanks.

Yes. AMD is the key to get eGPU screen output with 2014 15" MBPr Iris Pro only on OS X. It seems that MVC's custom eEFI is the only way to get this feature for Nvidia cards. As far as I know, Late 2013 13" MBPr supports Nvidia cards properly, and Late 2013 15" MBPr with 750M sometimes - not aware of any other TB2 Mac that is able to detect an external monitor via Nvidia eGPU.

@Tech Inferno Fan: Check your PM, I sent the link.

Share this post


Link to post
Share on other sites
@Tech Inferno Fan: Check your PM, I sent the link.

Bad news. The top PCie port is set to run at x4 and the 2 underlying buses then share that 16Gbps amongst themselves. Means that you'd get 8Gbps per port if running them concurrently. Daisy chaining additional eGPUs would give the same performance as if using the two ports on the Macbook. Only difference between the two being different buses used to host the eGPUs. The exact layout being:

0:1.0 (x4) -> 5:0.0 (x4); which then separates into two buses:

-> 6:3.0 (x4) -> 08:0.0 (x4) -> 09:3.0 (x4) -> 0a:0.0 (x4) -> b:4.0 (x8) -> e:0.0 (x8) [AMD card]
-> 6:5.0 (x4) -> 3A:0.0 (x4) -> 3B:3.0 (x4) -> 3c:0.0 (x4) [AMD card]

Apple should have hosted each TB2 port off on it's own x4 2.0 bus to yield an unshared 16Gbps per port. Hope they fix this in their <strike>Broadwell 15" Macbook Pro</strike> TB3 Skylake Macbook Pro.

Share this post


Link to post
Share on other sites

Can you tell us which AMD Kexts need to be modified to get the cards recognized and to get full acceleration? Trying something similar with my 7950.

Share this post


Link to post
Share on other sites
Can you tell us which AMD Kexts need to be modified to get the cards recognized and to get full acceleration? Trying something similar with my 7950.

See the first post, I just added the instructions. As you have the Tahiti series card as well, the modifications should be exactly the same.

Share this post


Link to post
Share on other sites

Great, thanks. Does your 4K display works with full resolution @60hz under OSX ? Screenshot only shows 2560x1440.

Share this post


Link to post
Share on other sites
Great, thanks. Does your 4K display works with full resolution @60hz under OSX ? Screenshot only shows 2560x1440.

You are welcome :) I prefer 2560x1440 HiDPI (if not directly selectable, you can use SwitchResX) because full UHD native resolution is too small for me. HD7970 max resolution @60Hz is 2560x1600 via mini display port or DL-DVI-I, and checked that 3840x2160@60Hz works with the 280X and U28D590 monitor on OS X.

Share this post


Link to post
Share on other sites

I added a YouTube video into the first post. It demonstrates what the NA211TB can be at its best:

Finally succeeded to run 3 eGPUs (see the video till the end), making it a very high-performance OpenCL computing device.

Share this post


Link to post
Share on other sites

I was curious after reading a couple of reviews of new R9 390 / R9 390X cards, and that made me thinking how they would perform in eGPU use even though we know that R9 200 series AMDs turned out to be unstable with AKiTiO. I took a risk and chose the same manufacturer as previously, Asus.

And the result?

3DMark11 physics looped at least 10 times in a row. Then…

AMD Radeon R9 390 video card benchmark result - Intel Core i7 4702HQ,Apple Inc. Mac-3CBD00234E554E41

NVIDIA GeForce GTX 980 video card benchmark result - Intel Core i7 4702HQ,Apple Inc. Mac-3CBD00234E554E41

AMD fights back, ~356€ R9 390 8GB beats my GTX 980 4GB.

P12774 (graphics 15114) vs P11863 (graphics 14162). Latest drivers from Nvidia and AMD. No overclocking. Does it work on OS X El Capitan? Yes. What would be the result of two R9 390s in CrossFire? Or 390X, Fury, Fury X, Nano…?

Share this post


Link to post
Share on other sites
P12774 (graphics 15114) vs P11863 (graphics 14162). Latest drivers from Nvidia and AMD. No overclocking. Does it work on OS X El Capitan? Yes. What would be the result of two R9 390s in CrossFire? Or 390X, Fury, Fury X, Nano…?

Sweet!

Share this post


Link to post
Share on other sites
Sweet!

I was too optimistic :/ Crashing again... but more stable than R9 200 series. R9 390 seems to give the best performance per dollar. No doubt that it's faster than GTX 970. If AKiTiO was stable with newer AMDs, we would see many implementations at the top of the leaderboard. I really wish we could run R9 Nano stable with AKiTiO...

NA211TB stays stable and can show even better results by +50% power target, +9% GPU clock, and +700 MHz memory clock:

post-28870-14495000595018_thumb.png

AMD Radeon R9 390 video card benchmark result - Intel Core i7 4702HQ,Apple Inc. Mac-3CBD00234E554E41

AMD Radeon R9 390 video card benchmark result - Intel Core i7 4702HQ,Apple Inc. Mac-3CBD00234E554E41

The scores would be ~200-300 points better with a single slot enclosure.

post-28870-14495000594544_thumb.png

Share this post


Link to post
Share on other sites

Thanks for the guide! I have a Radeon HD 7950 that I'm trying to get to work with my 2015 Macbook Air 11". I followed your steps, and the egpu shows up in the system report under Graphics/Displays, but under About this Mac>Displays, it still indicates that Intel HD6000 GPU is being used. I cleared the IOPCITunnelCompatible errors in system.log. And the kext seem to be loaded under System Report>Extensions. Is there anything else I should check or some way to get the display to use the egpu? I even attached an external monitor, and Displays still indicated that the Intel HD6000 was being used vice the egpu, even though the monitor was connected directly to the egpu. Thanks for any help.

Share this post


Link to post
Share on other sites

My desk is only 50cm*100cm and I want to build a eGPU system like you but with external keyboard and monitor(24"). Do you think it possible ? or it takes more place than I thought...

It would be extremely stupid to find room not enough after I built my eGPU system

Share this post


Link to post
Share on other sites

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!

Register a new account

Sign in

Already have an account? Sign in here.

Sign In Now


×

Important Information

By using this site, you agree to our Terms of Use. We have placed cookies on your device to help make this website better. You can adjust your cookie settings, otherwise we'll assume you're okay to continue.