Separate names with a comma.
Discussion in 'Mac Pro' started by Prince134, May 8, 2014.
All reference design cards should be fine. The bios attached were for reference card.
Prince, would you please post results of Haven and Valley using your low clocked 7970 dual setup?
I'm very curious as to how it compares to a GTX 980. I have been setting aside funds for a MVC 980, but if your double 7970 setup is more powerful, I may go that way.
I'm also interested in how they compare in efi win81 since then crossfire can be used against the 980.
I know this may disappoint you. I have no access to the Mac Pro I had with 7970s at the moment. It got to be couple weeks before I can do that for you. But honestly the main advantage for the dual 7970 is for openCL under MacOS with apple apps, e.g. FCPX. I don't think D700s will outperform GTX980 in openGL setting or PC gaming purposes.
I have never run a bootcamp for that Mac Pro in windows since I used it for FCPX mainly.
As my post in #76 said, I will try this setup if my 7950 die. And it is dead about 2 months ago. However, I am still a bit hesitate to use dual 7970 without PSU mod (or draw too much power from the SATA port). So, I step down a bit and use dual 7950.
In fact, initially, I've just bought a R9 280 (same as 7950) to replace my faulty card. Since my faulty 7950 still under warranty. So, I call Sapphire to manage a RMA for me. And now, I have 2x 7950 on hand and able to try this setup.
I finished my dual 7950 installation few days ago. So far everything work as expected. Of course, the machine is not as powerful as the OP's dual 7970. However, the power management is much easier. By lowering the voltage (0.887V for 7950, and 0.912V for the 280, both card keep all clock speed at stock), each card only consume 69W (in Luxmark) from the 6pin cables. Which means each card can connect to a single 6pin source by a mini 6pin -> 2x6pin adaptor safely for normal ops.
The above is the total power draw from both cards. The 7950 is in slot 3 and both 6pin cables connected to BoostA, 280 is in slot 1 and both 6pins connected to BoostB. This is the peak loading through out that 2min benchmark. 5.61A x 12.3V = 69W, within the 75W limitation. (Update: By running more tests, I realise that Luxmark 2 is NOT a good test for power compsumption, Lexmark 3 is more demanding, and much closer to real world high demand max power draw. e.g. 4K video rendering)
Anyway, this is my original setup. Both GPU installed into the x16 slots. And let the USB 3.0 card to use the remaining x4 slot. This config should give out best performance.
However, as you can see, the fan on the white card (7950) is fully covered by the black card (280). End up the 7950 run very hot and noisy (max 88C @66% fan). Even though it seems that the HDD also block the 280's fan, but it isn't. Also, this particular fan can suck air from below, so cooling is not a problem for the 280.
Since I don't want to kill my 7950 again (by heat). I re-arrange it as follow.
By installing the 7950 in slot 3 and move the HDD to bay 4, now there is enough room for the fans to work properly. Both card now run very quiet under full load. Even though it's not necessary, this config also provide an extra gap for the 280's cooler to suck air from above.
However, with this config, obviously I have to scarifies some performance from the 7950. Because slot 3 is a x4 slot, but how much I actually lost? Here is the result.
From this result, the GPU in a x16 slot contribute 51.2% and the card in a x4 slot contribute 48.8%. That means total performance penalty is just 2.4%, not significant at all. I am more than happy to trade 2.4% performance for much much better cooling (max 71C @36% fan) and 10 times less noise (of course, this is just my feeling, but not scientific figure).
In fact, there is almost no real world penalty.
(Single GPU installed on the x16 slot)
1684x2 = 3368, which is actually a bit less than 3383. So, I would say the penalty is virtually not exist (for most real world application that cannot fully utilise the bandwidth).
From Barefeats, the dual D700 score around 3527, only 4% difference. I am sure that I can set the frequency to 850MHz and beat it. However, that's not my intention for this setup.
Anyway, as the picture shows. Crossfire bridge is installed (only extra $3.5, why not?). Even though it's completely useless in OSX, but the Crossfire works very well in Windows 10 (if both card selected to the non EFI ROM). Which means I am now having a good gaming machine as well. In fact, I am not sure what the bridge actually does. But I guess if the card is communicating via the bridge but not the PCIe slot, that means the x4 penalty is further reduced (if any).
In terms of performance and power consumption. I am sure the new 9xx card can do much better. However, I am very hesitate to deal with those drivers. Especially my wife will also use this computer. If the computer unable to boot when I was out of town, I will be in deep trouble. So, I am better to stick with the 79xx card
Just did more tests, and this time mainly focus on the 3D performance.
The result from the R9 280 which installed in the x16 slot.
The result from the HD7950 which installed in a x4 slot.
I only run this test for once due to time constrain. From the FPS or score, there is a 4.8% performance hit for the card installed in a x4 slot. So the overall performance hit is 2.4%, exactly the same figure shows in the Luxmark test.
And now, let's see how about the power consumption. Since there is no Crossfire in OSX and I only connect one monitor to the Mac at this moment. So I simply capture the power draw during the last Valley benchmark.
During the R9 280 test (slot 1), total max power draw from the card = 90.9+32 = 122.9W
During the HD7950 test (slot 3), total max power draw from the card = 90.6+23.9 = 114.5W
As expected, both card can draw more than 75W from a single 6pin source under high demand. And both card max at about 90W, which is within my own limit. I personally assume there is a 50% safety buffer for this 6pin cable from Apple, so 75 x 150% = 112.5W (AFAIK, the self shutdown protection kicks in at about 120W). I set my own limit as 95W max continuous power draw, and 105W absolute max. So far, only Furmark can pull 97.5W from the cable and still no way to shut down the Mac. And my normal ops is not running Furmark, so I am not worry about it. Besides that, I also set a hard limit 130W TDP in each card's VBIOS as a final safety precaution. And I believe that's why Furmark can only draw 97.5W, because when combining with the power draw from PCIe slot (38W), the card already hit the 130W limit (in fact, excess a bit), and stop pulling further.
When both card running really hard together (e.g. Crossfire in Windows), the PCIe fan can go up to 2200RPM. Which of course will produce some noise, but I can live with that. When I am gaming, I will sit few feet away from the Mac, and I almost can't hear it.
For other work like heavy 4K video editing in FCPX, the PCIe fan can go up to around 1800RPM, which is no big deal for me. All 6G VRAM utilise, max power draw from the cables are 85W and 71W respectively (usually no more than 65W, just the peak can go over 75W). Well within my own limit.
By running MacsFanControl, I can easily create another PCIe fan profile which is base on the PCIe ambient temperature. This can further lower the noise but didn't affect much the GPU's cooling.
N.B. It's just all my sharing. I am NOT recommending anyone to use this config, especially it will draw more than 75W from a single 6pin PCIe cable which is no doubt excess the limit declared by Apple. Again, IMO, this setup is a bit easier than the dual 7970 (if you are willing to take the risk without PSU mod) because of the overall power consumption is a bit less. Also, there is no need to sacrifice any SATA port (my bay 2 looks must be empty, but I actually instaledl a 2.5" SSD at there now which won't affect the GPU cooling).