2014 13" MBP + GTX750Ti@16Gbps-TB2 (AKiTiO Thunder2) + Win8.1/OSX10.10 [Flinch]

[Flinch]

Hi All,

 

I just wanted to share my experiences with setting up a Gigabyte GTX 750 TI eGPU powered only by the PCI slot in the Akitio. 

 

My setup:

 

- 2014 Macbook Pro 13", 8GB RAM, 500GB SSD, running OS X 10.10 with Win 8.1 installed via boot camp (EFI)

- Gigabyte GTX 750 TI - the slim one with the single fan

- Akitio Thunder 2

- 120W Power brick from Trust

- external monitor

This looked like a really attractive solution to me, since it's as close to plug-and-play as it gets with a cool running GPU which fits tidily in the case and can be closed.

 

I don't recommend that power brick though, it says "automatic voltage selection" on it which is a bit bogus. The thing has a shunt regulator in it, so "automatically selecting" the voltage means choosing the correct plug (which has a built-in resistor) to fit on the end which regulates the voltage, and the correct plug for 12v was NOT included in the packaging (for me at least). So I cut the barrel plug off the Akitio and soldered that along with a 400k potentiometer which I used to regulate the voltage down from 19v to 12.3v (note, this is not the same as simply wiring in a resistor in series, a shunt regulator has a feedback line and the resistor goes between V+ and the feedback).

 

First off, there seemed to be some lack of clarity as to whether simply adding a 120W power brick and powering the whole thing through the barrel plug actually worked in practice. Turns out it works. I was playing the Witcher 3 on high settings at 1080p all evening yesterday.

 

Initially I tried to get it running on my newer 2015 13" MBP 16GB RAM + 1TB SSD running Win 8.1/El Capitan, and I couldn't get it to boot reliably (had to do the hot-plug-after-chime rain dance, and racing a fast SSD was really hit and miss), and most of the time I'd get a black screen or get to watch the pretty spinning dots for hours.

 

I noticed, though, that rebooting worked more frequently than a cold start, so I disconnected the Akitio fan as I suspected that the torque when it starts up might cause a voltage drop which puts the card in a funny state.

 

I then switched to my older 2014 MBP13"  (the one on which this all works now), and installed Win 8.1 via boot camp. I also installed the Optimus EFI goodies from here which may have had the effect of inserting a tiny delay via the chain loading from GRUB and could have helped (the theory being that the card needs power a little BEFORE the PCI scan at boot time, but only gets power when the thunderbolt signal comes up, and that there's a race).

 

Once I had it booting reliably without any weird rituals, I had stability issues the first few days, where I couldn't run a game for more than a minute without the card crashing in various interesting ways.  I noticed using GPU-Z that the power consumption was spiking as high as 115% TDP and was getting a lot of "PerfCap Reason" "Pwr" notices, so I underclocked it so that it doesn't go over ~98% TDP. And that was the key to getting it stable. After some experimentation, I ended up with both the core and memory clock speeds reduced by 150mhz (I used MSI AfterBurner). Did that, and now it works like a charm

 

Summary:

- get a 120W power brick and a GTX 750TI without additional power connectors

- disable the Akitio built-in fan by unplugging it from the board (the GPU runs very cool anyway, haven't gotten it over 62 deg. C yet)

- install Win 8.1 via boot camp

- write the OPTIMUS.dmg to the EFI partition (speculative)

- install latest NVIDIA web drivers

- underclock the card until stable

 

Thanks for all your advice and discussions. Couldn't have done it without y'all.

[Tech Inferno Fan]

@Flinch, thank you for sharing your implementation. You are one of the first to use a 120W AC adapter to power a GTX750Ti. Something that has been discussed but not practically implemented.

So then your details are quite important.

 

Would you mind taking photos of all your gear and say uploading them to http://imgur.com as a album? The album can be embedded on here for easy perusal.

 

If you have a chance, would you mind running 3dmark11 and 3dmark13 as well and linking the results? That enabling me to add you to the appropriate spot on the leaderboard.

[Dschijn]

Awesome that you kept fighting all that problems! I've read about many GTX750Ti users that quit because of that stability problems.

Good to know that the GPU can go over the power limit of the parts in the system and make it unstable. Would be interessting to know if it is the PSU or the AKiTiO that can't handle it.

Are you 100% sure that your PSU can give you the full 120W at 12V? I've seen PSUs with different voltages but only a fix amp value. With that the overall power gets lower the lower the voltage is.

 

If you are from Germany, you should have tried this PSU: http://www.amazon.de/gp/product/B006Z9TQE6?keywords=120w 12v leike&qid=1450253739&ref_=sr_1_sc_1&sr=8-1-spell

 

Guess the eGPU will not be stable in OSX since you can't underclock in OSX.

[Flinch]

3 hours ago, Tech Inferno Fan said:

@Flinch, thank you for sharing your implementation. You are one of the first to use a 120W AC adapter to power a GTX750Ti. Something that has been discussed but not practically implemented.

So then your details are quite important.

 

Would you mind taking photos of all your gear and say uploading them to http://imgur.com as a album? The album can be embedded on here for easy perusal.

 

If you have a chance, would you mind running 3dmark11 and 3dmark13 as well and linking the results? That enabling me to add you to the appropriate spot on the leaderboard.

 

Sure, I'll update with pics and benchmarks as soon as I get home (@work at the moment).

 

@Dschijn, thanks for the tip on the PSU, and yes I'm sure that it's 120W. It's a 10A supply with a range of 12V-19V wattages printed on the packaging with the wattage at maximum (120W) with the hypothetical 12V plug attached, and decreasing as the voltage increases which is exactly what you'd expect. It's basically the same thing as the one you linked, just a different brand.

 

Re: not enough power: at first, I did suspect that the PCI slot was really only delivering 25W and that this was the cause of the instability. I really don't think this is the case though. My clocks are still over 1GHZ with the memory clocked at the same frequency as the GTX 750 (not the TI), which is not an extreme underclocking, so I cannot imagine that it can run under load with only 25W.

 

Moreover, the 65W TDP isn't the same as the actual max power consumption (that's just what it can dissipate on average). The peak power consumption is often significantly higher in reality (20-30W is common), so my guess is that the GTX 750 TI under high load with factory settings can peak close to or over the 75W ceiling, and that the thunderbolt circuitry of the Akitio doesn't like it and becomes unstable.

 

PC's are probably more forgiving and have a bit more wiggle room in their PCI power circuitry, so you don't notice if you plug this GPU into a PC, but if you think about it, it almost sounds too good to be true that the GTX 750 TI gives you as much performance as it does for only 65W... because it probably really doesn't promise to *never* go close to or over 75W, but just to dissipate 65W on average with some clever scaling and fan control.

 

[Flinch]

EDIT: so the TL;DR version is that I'm pretty sure the Akitio does as advertised if you give it a 120W power supply by making 75W available, but the GPU is drawing more under load and either has the power throttled by the Akitio, so starving the GPU, or the Akitio thunderbolt daughter board throws a hiccup and the data pipeline stalls.

[Dschijn]

That is true. Every manufacturer is raising the power limit and clock speeds in the GPUs BIOS. That's why they always will exeed the stock power target. Lowering the power target should force the GPU back to it's older limits (like you did).

[Flinch]

Here are some benchmarks:

 

Sky Diver: http://www.3dmark.com/sd/3617810

3DMark 11: http://www.3dmark.com/3dm11/10676285

 

And the pics:

 

Spoiler

 

[Dschijn]

3d Mark Firestrike as well plz

[Flinch]

13 minutes ago, Dschijn said:

3d Mark Firestrike as well plz

I've swapped the Gigabyte card out with an EVGA GTX 750Ti SC (under-clocked the core clock only by 150Mhz to stabilize it), so this is with the EVGA:

 

http://www.3dmark.com/3dm/9758683

 

Hope that's okay, I don't feel like screwing in the Gigabyte again

Edited December 16, 2015 by Flinch

[sdedlus1]

On December 16, 2015 at 10:15 AM, Flinch said:

EDIT: so the TL;DR version is that I'm pretty sure the Akitio does as advertised if you give it a 120W power supply by making 75W available, but the GPU is drawing more under load and either has the power throttled by the Akitio, so starving the GPU, or the Akitio thunderbolt daughter board throws a hiccup and the data pipeline stalls.

 

On December 16, 2015 at 10:22 AM, Dschijn said:

That is true. Every manufacturer is raising the power limit and clock speeds in the GPUs BIOS. That's why they always will exeed the stock power target. Lowering the power target should force the GPU back to it's older limits (like you did).

Aha, this makes sense, and the 750 Ti has been benchmarked at peak 141 W, even if it averages more like 64 W. 

http://www.tomshardware.com/reviews/geforce-gtx-750-ti-review,3750-20.html

 

Thanks OP, this is very helpful. I was considering this setup but I'm looking for something that will work in OSX too, so I'll probably instead go with a card with 6-pin connector + Dell power supply, maybe GTX 950.

[kakarotto84]

 I myself own a Dell e6430 with ZOTAC 750 ti connected via PE4L Expresscard adapter. 

Isn't thunderbolt 2 supposed to be a tad faster vs mPCIe or Expresscard (20gbps / 5gbps)?

 

My 3dmark11: http://www.3dmark.com/3dm11/11169564 

 

I'm just referring to the graphics benchmark and not the overall result btw...

 

Can someone explain this anomaly? 

 

Update: OC'd to 1165mhz core clock with 90w laptop AC charger. Steady as a rock! http://www.3dmark.com/3dm11/11175775

 

 

Edited April 16, 2016 by kakarotto84

[Dschijn]

Could be that 3dmark11 doesn't need that much bandwidth. Try Firestrike!

Could also relate to different clock speeds of the cards (overclocks).

[kakarotto84]

On 4/13/2016 at 5:19 PM, Dschijn said:

Could be that 3dmark11 doesn't need that much bandwidth. Try Firestrike!

Could also relate to different clock speeds of the cards (overclocks).

 

Seems to be the case. http://www.3dmark.com/3dm/11641939

It finally showed Thunderbolt 2's advantage over Expresscard, bandwidth wise.

Don't know why the benchmark detected NVS 5200m as primary GPU though.

 

TIL: use firestrike for more accurate results

 

Thanks Dschijn!

 

update: overclocked a bit while still using a 90w AC charger since card will be restricted by the 75w pci-e slot anyway and no crashes (or accidents) so far.

http://www.3dmark.com/3dm/11655636

Edited April 15, 2016 by kakarotto84

[Xaularis]

On 12/16/2015 at 4:55 PM, Flinch said:

Hi All,

 

I just wanted to share my experiences with setting up a Gigabyte GTX 750 TI eGPU powered only by the PCI slot in the Akitio. 

 

My setup:

 

- 2014 Macbook Pro 13", 8GB RAM, 500GB SSD, running OS X 10.10 with Win 8.1 installed via boot camp (EFI)

- Gigabyte GTX 750 TI - the slim one with the single fan

- Akitio Thunder 2

- 120W Power brick from Trust

- external monitor

This looked like a really attractive solution to me, since it's as close to plug-and-play as it gets with a cool running GPU which fits tidily in the case and can be closed.

 

I don't recommend that power brick though, it says "automatic voltage selection" on it which is a bit bogus. The thing has a shunt regulator in it, so "automatically selecting" the voltage means choosing the correct plug (which has a built-in resistor) to fit on the end which regulates the voltage, and the correct plug for 12v was NOT included in the packaging (for me at least). So I cut the barrel plug off the Akitio and soldered that along with a 400k potentiometer which I used to regulate the voltage down from 19v to 12.3v (note, this is not the same as simply wiring in a resistor in series, a shunt regulator has a feedback line and the resistor goes between V+ and the feedback).

 

First off, there seemed to be some lack of clarity as to whether simply adding a 120W power brick and powering the whole thing through the barrel plug actually worked in practice. Turns out it works. I was playing the Witcher 3 on high settings at 1080p all evening yesterday.

 

Initially I tried to get it running on my newer 2015 13" MBP 16GB RAM + 1TB SSD running Win 8.1/El Capitan, and I couldn't get it to boot reliably (had to do the hot-plug-after-chime rain dance, and racing a fast SSD was really hit and miss), and most of the time I'd get a black screen or get to watch the pretty spinning dots for hours.

 

I noticed, though, that rebooting worked more frequently than a cold start, so I disconnected the Akitio fan as I suspected that the torque when it starts up might cause a voltage drop which puts the card in a funny state.

 

I then switched to my older 2014 MBP13"  (the one on which this all works now), and installed Win 8.1 via boot camp. I also installed the Optimus EFI goodies from here which may have had the effect of inserting a tiny delay via the chain loading from GRUB and could have helped (the theory being that the card needs power a little BEFORE the PCI scan at boot time, but only gets power when the thunderbolt signal comes up, and that there's a race).

 

Once I had it booting reliably without any weird rituals, I had stability issues the first few days, where I couldn't run a game for more than a minute without the card crashing in various interesting ways.  I noticed using GPU-Z that the power consumption was spiking as high as 115% TDP and was getting a lot of "PerfCap Reason" "Pwr" notices, so I underclocked it so that it doesn't go over ~98% TDP. And that was the key to getting it stable. After some experimentation, I ended up with both the core and memory clock speeds reduced by 150mhz (I used MSI AfterBurner). Did that, and now it works like a charm

 

Summary:

- get a 120W power brick and a GTX 750TI without additional power connectors

- disable the Akitio built-in fan by unplugging it from the board (the GPU runs very cool anyway, haven't gotten it over 62 deg. C yet)

- install Win 8.1 via boot camp

- write the OPTIMUS.dmg to the EFI partition (speculative)

- install latest NVIDIA web drivers

- underclock the card until stable

 

Thanks for all your advice and discussions. Couldn't have done it without y'all.

I got exact the same issue.

 

What the frequency of the clock have you reduced to ?

 

Many thanks.