US$189 AKiTiO Thunder2 PCIe Box (16Gbps-TB2)

[Tech Inferno Fan]

@Tech Inferno Fan: the graphics cards I have tested don’t utilise the specified 75W as shown in this table. It wasn’t just a single test, they are approximately same numbers after tens of tests. The interesting detail with a R9 280X was that the PCIe board power consumption numbers switched precisely 9.5W or 10.4W during 3DMark11 physics with apparently constant time period, no other value between those. With stable cards, such as EVGA GTX 780 OC, it was constant 7.4W the whole test period. The same 7.4W at idle when the GPU is not stressed. When R9 270X and HD 7970 passed the test, there was only 0.1W variation.

I am using a 3DMark11 physics test as a stability indicator, because it’s a 100% crash for AMDs with those two common powering methods: a back-powered riser, molex-to-barrel adapter or combination of these two. I tested also with two separate 6-pin to 4-pin molex cables from the PSU. All the test were done without soldering the AKiTiO board.

I have found keys to pass the physics twice in a row by isolating the PCIe board powering, so that the TB card and PCIe board power is not shared with that 4-pin yellow/black cable.

When I played with a multimeter a bit more, I found something interesting when testing the TB card alone. As you said, there is no direct connection from the 4-pin TB card socket +12V pin to the x4 slot (B1, B2, B3, A2, A3), but I can repeatedly get a short beep sound by switching red/black probes, indicating maybe a capacitor between those test points. When I did a diode test with the same test points, the multimeter showed “OL” (very high resistance) when the black proble was placed at x4 12V pin and the red probe was placed at 4-pin socket 12V. However, when I did the diode test by switching probes, the multimeter showed continuously 0.511 V (I guess the display number is the forward voltage drop). Does this mean the power is meant to flow only one direction and blocking from the other direction? Is there a risk that with back powering the TB card we may destroy the diodes on the board? (I remember one EE warned about it).

However, it seems to be possible with a back-powered riser to provide power firstly for the GPU, then for the TB card, and finally for the PCIe board via 4-pin cable. The other way is from the DC jack or J4 (where it splits for the PCIe board) -> TB card -> x4 slot -> x16 slot.

I have a possibility to test with either a back-powered riser (as most of the users here) or non-back-powered riser such as this:

8X PCIe 16X extension cable 19cm power

It’s expensive but was very stable with the Sonnet SEL. And the yellow wire is very thick. Some risers that I have got from eBay seemed good externally (two thick yellow/black wires + capacitor) but when I cut the wire, the inner part was ridiculously thin. Sometimes you cannot even trust to AWG markings.

One more thing, with the TB card attached to the x4 slot of the PCIe board, the multimeter behaved differently when testing the PCIe board reverse points (marked as red lines in your picture). The 4-pin cable was removed. And what was the difference? Well, I got those short beep sounds again (continuity test) when the TB card was attached, but when not attached, it was mute. The same result with a diode test, but this time no continuous value as with TB card 4-pin <-> x4. I am happy to continue with the aid of EE experience. Already did the electric tape test with the back-powered riser / non-back-powered riser + molex-to-barrel adapter, but unfortunately the same 3DMark11 physics crash with the R9 280X. Should I try that soldering? I have a plan to use pins at the end of the wire for x16 slot reverse board points.

Yes, once the PCIe board is attached to the TB board, testing the 12V points (A1..A3, B1..B3) for continuity against J4/J6 may get something back on your multimeter. I actually tested resistence and found it was not a direct link. Instead, probing a few caps on the TB board showed a direct link. Hence the incoming 12V at J4/J6 passes through a filtering stage before being presented at those PCIe slot points. That is why it's recommended to isolate those pins so as to not mix two different 12V sources if using PCIe riser, .

You mention isolating the 6 pins on the TB board as allowing the physics test to complete successfully. That's good news! As expected, you are then not mixing what are different 12V inputs (12V_ATX and 12V_filtered).

Yes, I'd suggest performing the 75W PCIe slot power mod outlined (solder 2 wires + isolate 6 pins) to see if you can get stability with your R9 280x. From your ATX PSU you also need a molex-to-DC barrel connector to attach to the DC input jack as well as the PCIe power leads to attach to the video card. Then it's all good to test.

The difference there as opposed to using a powered PCIe riser you are already using is you are getting a better quality Gen2 signal between the TB bridge and the AMD card. If the signal is noisy and can't maintain Gen2 speed then I suspect the TB-TB driver will crash in a similar way to what you are describing. That could be a signal integrity issue.

Why does a riser introduce signal degradation? This being for the same reason as why adding a socketted connection degrades our expresscard/mPCIe connections. Wherever a riser (or socket) is introduced along the Gen2 5Ghz signal path, any mismatch in impedance causes signal reflections. The ideal situation for best Gen2 signal integrity is to NOT cause any reflections and so avoid PCIe risers (or sockets for mPCIe/EC situations) altogether.

[Morv]

Thank you @Tech Inferno Fan.

It's too bad I've already sold my GTX 780, I'd probably have tried the mod.

I hope someone else can try that and hopefully it really fixes the issue because otherwise the AKiTiO is awesome

Also, if Akitio knows about us using their box as an eGPU adapter they hopefully address this issue by themselves for the next generation with Thunderbolt 3, once it's released. Do you think they'd react to our needs as far as possible?

[euqlaog]

You mention isolating the 6 pins on the TB board as allowing the physics test to complete successfully. That's good news! As expected, you are then not mixing what are different 12V inputs (12V_ATX and 12V_filtered).

Yes, I'd suggest performing the 75W PCIe slot power mod outlined (solder 2 wires + isolate 6 pins) to see if you can get stability with your R9 280x. From your ATX PSU you also need a molex-to-DC barrel connector to attach to the DC input jack as well as the PCIe power leads to attach to the video card. Then it's all good to test.

I have to clarify the sentence that you bolded. By isolating PCIe board powering I meant using one ATX PSU for the DC jack input, and one for the TB card 4-pin, and not isolating with electronic tape those 6-pins of the x4. That way, I was able to measure the power consumption of both inputs separately. GTX 780 took much more wattage via TB 4-pin than R9 280X.

I did the proposed 6pin taping trick with a powered riser, one ATX PSU and molex-to-barrel adapter, but it didn’t seem to give improvement to stability regarding physics test. I think the reason might be the powered riser. I will try the soldering method too and report the results.

[Tech Inferno Fan]

I have to clarify the sentence that you bolded. By isolating PCIe board powering I meant using one ATX PSU for the DC jack input, and one for the TB card 4-pin, and not isolating with electronic tape those 6-pins of the x4. That way, I was able to measure the power consumption of both inputs separately. GTX 780 took much more wattage via TB 4-pin than R9 280X.

I did the proposed 6pin taping trick with a powered riser, one ATX PSU and molex-to-barrel adapter, but it didn’t seem to give improvement to stability regarding physics test. I think the reason might be the powered riser. I will try the soldering method too and report the results.

The 6pin taping trick can be applied to the powered riser instead - the male side that slots into the AKiTiO. Then 12V input at the PCIe riser and the 12V input via the barrel adapter are kept separate. Easier than dismantling the AKiTiO.

Applying the soldering method means you can get rid of the powered riser and instead, slot the video card directly into the AKiTiO Thunder2.

[euqlaog]

The 6pin taping trick can be applied to the powered riser instead - the male side that slots into the AKiTiO. Then 12V input at the PCIe riser and the 12V input via the barrel adapter are kept separate. Easier than dismantling the AKiTiO.

Applying the soldering method means you can get rid of the powered riser and instead, slot the video card directly into the AKiTiO Thunder2.

Ok, I only taped the 6pins of x4 but taping the male side of x16 riser may be worth to try. However, the result might be equal to using non-back-powered riser as it only feeds power to the direction of the GPU. Earlier, I didn’t know that there exists also that kind of risers, until I did a continuity test with a multimeter.

[Tech Inferno Fan]

Ok, I only taped the 6pins of x4 but taping the male side of x16 riser may be worth to try. However, the result might be equal to using non-back-powered riser as it only feeds power to the direction of the GPU. Earlier, I didn’t know that there exists also that kind of risers, until I did a continuity test with a multimeter.

I mean if you are using a powered PCIe riser like the one shown below I refer to cellophane taping the left (male) side that slots into the AKiTiO. Apply cellophane to the first 6 pins on the left. That's 3pins on each side. That doing the same thing as applying cellophane to the AKiTiO board's 6pins as I described earlier but is easier to.

That then separates the 12V input via the molex on the riser and the 12V input to the AKiTiO Thunder via the DC barrel adapter to increase power stability.

[euqlaog]

I mean if you are using a powered PCIe riser like the one shown below I refer to cellophane taping the left (male) side that slots into the AKiTiO. Apply cellophane to the first 6 pins on the left. That's 3pins on each side. That doing the same thing as applying cellophane to the AKiTiO board's 6pins as I described earlier but is easier to.

That then separates the 12V input via the molex on the riser and the 12V input to the AKiTiO Thunder via the DC barrel adapter to increase power stability.

That did the trick! 3DMark11 test passed twice in a row with a single 450W PSU + taped 6pins of the powered riser + molex-to-barrel adapter. This has never happened before with a single PSU. The third time almost went through, struggling with very low FPS during physics, the whole eGPU system took over 110W but at the end of the physics test, it showed blank white screen. When the physics test was passed those two times, energy meter displayed only 47-49W. I rebooted to confirm that this wasn’t just a coincidence, and it passed physics again, but only once this time. However, that’s significant improvement, because without taping, it was 100% crash. Maybe we can achieve 100% stability for the R9 280X by soldering

[longgee]

Just want to let you guys know that I am closely following this thread and i am from Singapore I want to buy the AkiTiO for 189 but I don't think they ship it here.

[Tech Inferno Fan]

Just want to let you guys know that I am closely following this thread and i am from Singapore I want to buy the AkiTiO for 189 but I don't think they ship it here.

Contact the email address on the first post. They ship it out of Taiwan.

We can do ship overseas using DHL to ship. I can list out some estimated quotes for you for the following country but prices vary upon region within the country as well.

Asia - $68.46 – $116.16

Europe - $61.75 - $109.68

As you mention, it may be better for them to buy from one of our resellers if there is duty and high shipping costs involved. We can ship to customers in Asia from our Taiwan office and that should lower the shipping costs quite a bit.

[euqlaog]

I decided to try Nando’s suggested slot power method with the taping trick. R9 280X was attached directly into the x16 slot. Instead of direct soldering of the other end, I used the existing 2-pin fan/led power socket of the AKiTiO PCIe board and a bit thicker cables for 12V. Did that made it better than a powered riser + taping? Unfortunately no. This mod passes 3DMark11 physics the first time, but the second run will crash. I know this is not an AMD driver or operating system related problem, because NA211TB with the same card is very stable - just a moment after AKiTiO tests, 4 sequential 3DMark11 tests passed and at this point I didn’t try more because it has never shown stability issues.

The last thing to try is those additional capacitors. By the way, x16 slot pins are very very small…

[Tech Inferno Fan]

I decided to try Nando’s suggested slot power method with the taping trick. R9 280X was attached directly into the x16 slot. Instead of direct soldering of the other end, I used the existing 2-pin fan/led power socket of the AKiTiO PCIe board and a bit thicker cables for 12V. Did that made it better than a powered riser + taping? Unfortunately no. This mod passes 3DMark11 physics the first time, but the second run will crash. I know this is not an AMD driver or operating system related problem, because NA211TB with the same card is very stable - just a moment after AKiTiO tests, 4 sequential 3DMark11 tests passed and at this point I didn’t try more because it has never shown stability issues.

The last thing to try is those additional capacitors. By the way, x16 slot pins are very very small…

[ATTACH=CONFIG]14106[/ATTACH]

Are you using the same PSU for both tests? AFAIK, the NA211TB comes with it's own PSU. Can you use that PSU to do testing with the AKiTiO to make it an equivalent test? After the mod the only reasons I can see for your instability are (1) poor PSU causing power instability/fluctuations (2) the AKiTiO + AMD card having somewhat different PCIe slot impedances causing signal degradation.

Since you are seeing the problem at the same spot during 3dmark11 testing, I'm more inclined to suggest this is a power issue. Eg: during that particularly moment in the test a large power surge being the cause of the crash.

[benchallenger]

Hi - I just bought a unit in sim lim square 2 days ago - 5th floor, place called storage solutions I think. 400 SGD. They have in stock.

[Dschijn]

Hi - I just bought a unit in sim lim square 2 days ago - 5th floor, place called storage solutions I think. 400 SGD. They have in stock.

Singapore?

[benchallenger]

Correct

[euqlaog]

Are you using the same PSU for both tests? AFAIK, the NA211TB comes with it's own PSU. Can you use that PSU to do testing with the AKiTiO to make it an equivalent test? After the mod the only reasons I can see for your instability are (1) poor PSU causing power instability/fluctuations (2) the AKiTiO + AMD card having somewhat different PCIe slot impedances causing signal degradation.

Since you are seeing the problem at the same spot during 3dmark11 testing, I'm more inclined to suggest this is a power issue. Eg: during that particularly moment in the test a large power surge being the cause of the crash.

The PSU for AKiTiO testing was a Silverstone Strider ST45SF-G 450W 80+ Gold. Yes, NA211TB has its own PSU with apparently multiple 12V rail, and it was possible to test it with the AKiTiO as well, but didn’t pass the physics part. Then I tried Silverstone’s PSU for NA211TB, and at least two sequential 3DMark11 tests were successful. NA211TB seems to be very stable no matter which PSU you use.

The biggest difference is how NA211TB’s PCIe board is powered - it has a straight 24pin connector whereas AKiTiO is now powered by 6-pin -> 4-pin molex cable -> molex-to-barrel adapter. I wonder why 4-pin molex connectors still exist… the pins can easely move and refract. Maybe I should make a better barrel adapter.

I noted that AKiTiO (if it doesn’t crash) gives sligthly better score for R9 280X than NA211TB after many runs of 3DMark11. Maybe due to more simple PCIe board.

Yep, 3DMark11 physics is not power consuming at all, and when it comes to near the end, that’s the critical point as suddenly much more energy is needed for the last combined test.

Does it matter if I tape the A1 pin too? I taped all the 6 pins of the x4 slot. Taping 6pins of the male side of the powered riser with a capacitor has shown better results. Maybe stability can be improved with some extra capacitors.

[Tech Inferno Fan]

The PSU for AKiTiO testing was a Silverstone Strider ST45SF-G 450W 80+ Gold. Yes, NA211TB has its own PSU with apparently multiple 12V rail, and it was possible to test it with the AKiTiO as well, but didn’t pass the physics part. Then I tried Silverstone’s PSU for NA211TB, and at least two sequential 3DMark11 tests were successful. NA211TB seems to be very stable no matter which PSU you use.

The biggest difference is how NA211TB’s PCIe board is powered - it has a straight 24pin connector whereas AKiTiO is now powered by 6-pin -> 4-pin molex cable -> molex-to-barrel adapter. I wonder why 4-pin molex connectors still exist… the pins can easely move and refract. Maybe I should make a better barrel adapter.

I noted that AKiTiO (if it doesn’t crash) gives sligthly better score for R9 280X than NA211TB after many runs of 3DMark11. Maybe due to more simple PCIe board.

Yep, 3DMark11 physics is not power consuming at all, and when it comes to near the end, that’s the critical point as suddenly much more energy is needed for the last combined test.

Does it matter if I tape the A1 pin too? I taped all the 6 pins of the x4 slot. Taping 6pins of the male side of the powered riser with a capacitor has shown better results. Maybe stability can be improved with some extra capacitors.

A1 isn't connected, so is OK to tape over. Yes, if the molex-to-barrel is flimsy then it could be a problem with power delivery. Consider bypassing the barrel connector and attach a molex to the tracks instead as done at http://forum.techinferno.com/diy-e-gpu-projects/8882-%5Bguide%5D-2014-13-macbook-pro-gtx970%4016gbps-tb2-akitio-thunder2-handle-win8-1-a.html#post121260

The single-slot AKiTiO Thunder2 will be faster than any multi-slot Thunderbolt enclosure such as the NA211TB or Sonnet EE III-D. That's because the latter uses PCIe bridges to share the out the PCIe link, adding latency.

[Dschijn]

I wonder why 4-pin molex connectors still exist… the pins can easely move and refract. Maybe I should make a better barrel adapter.

True! 4-Pin Molex is the worst connector ever. Maybe it is better to connect the blank wires by soldering them together?!

[Oscar J]

If I haven't got the tools or the time to mod the Thunder2, and instead buy the OWC Mercury Helios, what would I need to do to run an AMD R9 card (170W+)? Or would it be unstable just like the Akitio is with AMD cards?

[Morv]

No one will know unless someone tries it. You'd still need an ATX PSU(or any other compatible kind of external PSU), though. Helios provides 120W and you obviously can't run a 170W+ card with it while expecting it to be stable.

Be the first to try Helios with AMD card and tell us if it works. No risk, no experience

[Oscar J]

What modifications do you reckon I'd have to do except for a more powerful PSU? Surely it can't be that simple?

[Morv]

You may also want to only use one PSU(the ATX or whatever one), so you'd need a barrel plug.

I bought such a thing -> 2.5mm DC Power Socket with Screw Terminal Block (5pcs) and removed the molex connector from my old powered riser, so I could connect the cables with the block. You may take any other cheap molex-to-xyz adapter,

This way you can power the board itself and your gpu with the ATX PSU.

[Oscar J]

You may also want to only use one PSU(the ATX or whatever one), so you'd need a barrel plug.

I bought such a thing -> 2.5mm DC Power Socket with Screw Terminal Block (5pcs) and removed the molex connector from my old powered riser, so I could connect the cables with the block. You may take any other cheap molex-to-xyz adapter,

This way you can power the board itself and your gpu with the ATX PSU.

I googled molex-to-xyz adapter, but didn't seem to get any relevant results. Could you be a bit more exact? I'm an absolute rookie with these things.

[euqlaog]

Yes, if the molex-to-barrel is flimsy then it could be a problem with power delivery. Consider bypassing the barrel connector and attach a molex to the tracks instead as done at http://forum.techinferno.com/diy-e-gpu-projects/8882-%5Bguide%5D-2014-13-macbook-pro-gtx970%4016gbps-tb2-akitio-thunder2-handle-win8-1-a.html#post121260

I got excited and bypassed the barrel connector by soldering 4 thick wires (two yellow 12V + two black GND) from a 4-pin molex connector. Heat shrink tubing tightened them together. I still use the x4 slot taping trick + 2-pin socket on the rear side of the J6 to provide 75W for the x16 slot.

By doing this, I have the most stable AMD eGPU system so far and no need for a barrel adapter or a powered riser. However, the maximum physics pass count with R9 280X is 2. I also disabled the ULPS (Ultra Low Power State) by modifying windows registry values, but maybe this only concerns CrossFire setups.

Furthermore, I soldered an additional capacitor at C3 (used one from my powered risers). Not sure how much that improved stability.

My conculsion is that the taping trick is the most important modification - not mixing 12V_filtered and 12V_ATX. I suppose the problem is how the R9 280X utilises the PCIe slot power when the GPU has nothing to do during physics. Earlier, without taping trick, R9 280X crashed often at idle, but now it doesn’t seem to happen.

With this setup, HD 7970 has no problem going through 4 sequential rounds of 3DMark11 test (didn’t try more). I just need a new and quieter cooler as the blower fan is very loud during graphically intensive tasks. Another good thing about HD 7970 is that Iris Pro is always detected without error code. If you want the same with R9 280X, you have to plug in the display cable a little more later after starting the computer.

[Dschijn]

@goalque: That sounds good. Could you post a picture of the PCB and the soldered cables?

Is the instability only present in 3dMark? Can you play games without that problem?

Modern GPUs are crazy with power management. They can switch between power states really fast and have crazy power consumptions for a short peroid of time.

I can recommend the Raijintek Morpheus cooler! Very beefy, but very powerful as well!

[euqlaog]

@goalque: That sounds good. Could you post a picture of the PCB and the soldered cables?

Is the instability only present in 3dMark? Can you play games without that problem?

Modern GPUs are crazy with power management. They can switch between power states really fast and have crazy power consumptions for a short peroid of time.

I can recommend the Raijintek Morpheus cooler! Very beefy, but very powerful as well!

Yes, I can post a picture later on today. Have to be very careful with the soldering, already melted a corner of the 4-pin socket

I have only focused on 3DMark11 physics because most Nvidia cards pass that very well, but R9 280X and R9 270X struggle with it, suddenly the whole eGPU power consumption may inrease over 110W resulting in crash. R9 280X has never crashed during other parts of 3DMark11, only the physics. If the power consumption stays low and does not vary a lot, it will pass the physics.

The AMD instability is related to low power consumption situations, such as using a browser or just opening windows. With the normal powering methods AMD cards often crashed after a few minutes when browsing the web. Taping trick seemed to give a solution for this, but I haven’t done comprehensive “normal use” testing. Before this setup, I once played BF4 and ran Valley benchmarking on OS X for an hour and no sign of crash.

I will take a look that cooler, thanks!