How to create a Mac Pro 5.1 that is more powerful than a Mac Pro 2013?

[rueyloon]

Last month I picked up a 4.1 for around $100 USD which started me on this black hole to trying to work on it.

I did the following
updated it to 5.1
upgraded to 12 core 3.46
bought a PCI adapter and used the 1gig from my macpro
bought 2 X R9-390x (I can put in 3 if I sacrifice all PCI slots)

The computer is still slower than my NMP 12core 2.5ghz

Any other ideas?

 

[h9826790]

Last month I picked up a 4.1 for around $100 USD which started me on this black hole to trying to work on it.

I did the following
updated it to 5.1
upgraded to 12 core 3.46
bought a PCI adapter and used the 1gig from my macpro
bought 2 X R9-390x (I can put in 3 if I sacrifice all PCI slots)

The computer is still slower than my NMP 12core 2.5ghz

Any other ideas?

Slower for what?

 

[rueyloon]

Slower for what?

Pretty much everything that I do that is, fcpx photoshop etc...

 

[h9826790]

If your nMP is the 12 cores + dual D700. I won't be surprised that the nMP is faster than your cMP.

The nMP has better CPU performance on both single thread and multi thread operation (even though not much). And that's just the raw power, the nMP has newer and better architecture. It's highly likely that it can finish the same job faster.

Dual 390X should be significantly stronger than dual D700. However, very likely FCPX is so optimised for the D700. Therefore, D700 can perform better actually on most simple task. And unless you really doing something that beyond the D700 can handle, the 390X may not able to show off it's power.

For photoshop, it's mainly CPU single thread, RAM size + speed, and hard drive speed limiting. And nMP sure has better CPU single core performance. As long as the RAM size is reasonable, the nMP's RAM should also performance better than the cMP. For SSD, the nMP SSD sure is fast enough to handle photoshop, so no big difference between cMP and nMP.

So, unless you are doing something that clearly favour the cMP. e.g. a software that can utilise CUDA. Or something that can benefit from >5000MB/s read / write speed. Or utilise the 10Gb network. etc. For most simple task, the nMP should work better.

In fact, the nMP is a FCPX orientated machine. I will be very surprised if a maxed out nMP still slower than the cMP in FCPX. So far, AFAIK, the fastest Mac (real Mac) in BruceX test still the nMP.

I rarely heard that the cMP can finish some normal real world jobs that significantly faster than the nMP (basically except the software that optimised for Nvidia GPU). The advantage of the cMP is that we can build a machine that comparable to the nMP with just 1/3 of it's cost. Or we can config / upgrade the cMP to make it more fit (or cost effective) for our work.

 

[thats all folks]

how are you powering two 390Xs? are yours flashed/resistor modded? is that card even fully supported in OS-X?

I believe the R9 280X is a near equivalent drop in for the D700s. though only 3GB RAM.

 

[rueyloon]

how are you powering two 390Xs? are yours flashed/resistor modded? is that card even fully supported in OS-X?

I believe the R9 280X is a near equivalent drop in for the D700s. though only 3GB RAM.

power with external PSU, currently I'm just experimenting so the door to the Mac Pro is left open.
The AMDs can work straight out of box, does not need to be flashed but no boot up screen.

The device ID is already part of the KEXT so no kext editing is even required.

 

[DPUser]

Last month I picked up a 4.1 for around $100 USD which started me on this black hole to trying to work on it.

I did the following
updated it to 5.1
upgraded to 12 core 3.46
bought a PCI adapter and used the 1gig from my macpro
bought 2 X R9-390x (I can put in 3 if I sacrifice all PCI slots)

The computer is still slower than my NMP 12core 2.5ghz

Any other ideas?

No. As noted, the architecture is several generations older.

But look at how much you spent compared to the nMP.

 

[Plato65]

How would one maximize the performance of the MacPro5,1 (whether it will beat MacPro6,1 or not)?
1) 12 x 3.46GHz processor seems obvious
2) 6 x 16GB (=96GB) memory seems the fastest, though 8 x 16GB (=128GB) may be better for some
3) Amfeltec PCIe card + 4 x HyperX 960GB SSDs should give ~5GB/s speeds for ~3.8TB storage
4) NVidia Titan X Maxwell seems to be the fastest card with drivers
5) 10 gigabit ethernet and/or USB3.1 cards for networking/communication speed

For maximal CUDA performance one could use 2 x Titan X, but then one would need an external power supply (I think) and either one Titan X would have to be in an x4 slot or the Amfeltec card would have to be moved to an x4 slot with slower transfer rates.

Any other optimizations?

 

[rueyloon]

how are you powering two 390Xs? are yours flashed/resistor modded? is that card even fully supported in OS-X?

I believe the R9 280X is a near equivalent drop in for the D700s. though only 3GB RAM.

Is it possible for the 280x to be recognized as a d700?

 

[pat500000]

If your nMP is the 12 cores + dual D700. I won't be surprised that the nMP is faster than your cMP.

The nMP has better CPU performance on both single thread and multi thread operation (even though not much). And that's just the raw power, the nMP has newer and better architecture. It's highly likely that it can finish the same job faster.

Dual 390X should be significantly stronger than dual D700. However, very likely FCPX is so optimised for the D700. Therefore, D700 can perform better actually on most simple task. And unless you really doing something that beyond the D700 can handle, the 390X may not able to show off it's power.

For photoshop, it's mainly CPU single thread, RAM size + speed, and hard drive speed limiting. And nMP sure has better CPU single core performance. As long as the RAM size is reasonable, the nMP's RAM should also performance better than the cMP. For SSD, the nMP SSD sure is fast enough to handle photoshop, so no big difference between cMP and nMP.

So, unless you are doing something that clearly favour the cMP. e.g. a software that can utilise CUDA. Or something that can benefit from >5000MB/s read / write speed. Or utilise the 10Gb network. etc. For most simple task, the nMP should work better.

In fact, the nMP is a FCPX orientated machine. I will be very surprised if a maxed out nMP still slower than the cMP in FCPX. So far, AFAIK, the fastest Mac (real Mac) in BruceX test still the nMP.

I rarely heard that the cMP can finish some normal real world jobs that significantly faster than the nMP (basically except the software that optimised for Nvidia GPU). The advantage of the cMP is that we can build a machine that comparable to the nMP with just 1/3 of it's cost. Or we can config / upgrade the cMP to make it more fit (or cost effective) for our work.

I have to agree with this. Nmp is optimized for fcpx.

 

[mastermamo]

I did a side by side test using the following:
5.1 12 Core 64 gig RAM GTX780Ti
6.1 3.5Ghz 12gig RAM

Rendered out a scene in AE with tons of particles

Scene Duration : 30 secs
Codec : Prores 422 HQ (Lossless)
Resolution : 1920x1080 25fps

5.1 = 1min 28s
6.1 = 1min 03s

Now 20 odd seconds might not seem a big margin but for longer duration renders etc it's a lifetime. Remember with more RAM on the 6.1 times will drop even more. Not bad performance at all but does it justify the huge capital outlay over a 5.1.
Nope not in my book.

 

[h9826790]

Is it possible for the 280x to be recognized as a d700?

Yes, but won't make any difference. It's just cosmetic, you can flash the 280X with a Mac EFI that ident itself as D700. Or simply hex edit the kext to make it ident all unflashed 7970 / 280X as D700.

 

[rueyloon]

Yes, but won't make any difference. It's just cosmetic, you can flash the 280X with a Mac EFI that ident itself as D700. Or simply hex edit the kext to make it ident all unflashed 7970 / 280X as D700.

I asked this because you pointed out fcpx is optimized to work with the d700, so a more powerful card might actually perform slower. So if the 280x can be recognized as a d700 will it work tighter than say the 390x that I'm using?

 

[h9826790]

I asked this because you pointed out fcpx is optimized to work with the d700, so a more powerful card might actually perform slower. So if the 280x can be recognized as a d700 will it work tighter than say the 390x that I'm using?

I said that because it's easier to understand. The performance has nothing to do with the GPU's name, but its device ID.

If you want, I can amend my statement to "FCPX is very well optimised to the GPU that has device ID 6798".

I am not sure if the 280X can do better than 390X in FCPX, but it's totally possible. The 390X is not a popular card in the cMP circle. Even though it's a kind of OOTB card, but the driver was buggy (I have no idea how's the driver quality now).

 

[PhotoGUY73]

I am interested in this particular set up. I already have a quote for the Amfeltec PCIe card.
The HyperX are the alternative to the Samsung SM951 512GB,Internal (MZHPV512HDGL-00000) SSD AHCI ?

3) Amfeltec PCIe card + 4 x HyperX 960GB SSDs should give ~5GB/s speeds for ~3.8TB storage


My current setup:
MacPro 5,1 2.93ghz 12core
MacVidCards Nividia Gtx 980 Ti 6GB
Samsung SSd 850 PRO 256GB boot drive Sierra OS
Various HHD.

I'm also looking for suggestions for RAID hard drives into internal bays of Mac Pro?

Anyone?

 

[Daniel Reed]

The computer is still slower than my NMP 12core 2.5ghz

Any other ideas?

The nMP 12-core is E5 2697 v2 and is listed by Intel as 2.7 GHZ part.
BUT, it will turbo mode all 12 active cores at 3GHz if you keep it cool (replace the crap factory thermal paste).
It will also turbo at 3.5GHz w/ 6 cores active.

Not possible to beat the top nMP CPU performance on the Cheesegrater.
BUT with multiple GTX Titan-X 12GB Maxwell cards and an external power supply,
It can kill it as a CUDA beast in apps!
BUT, those 12GB card are at least $1K each, or a bit less used on eBay.

For FCPX performance no idea, i don't use it.

AND for the trashcan, you can add up to two additional eGPUs.
I doubt AMD drivers were updated in 10.12.3 for Polaris 10 (i)Mac support.

When does support Polaris 10 (most likely in 10.12.4), the nMP w/ 12-core + Dual D700s + Dual RX480s may kill it for OpenCL/FCPX w/ 136 compute units

 

[ActionableMango]

Generally speaking I'd imagine that the newer computer is typically going to be faster. However, there are some limited scenarios where a 12-core, 64GB, dual D700 nMP was slower than a similarly equipped cMP (12-core, 64GB, dual R9280x). Barefeats has a list of some of them here:
http://barefeats.com/tube16.html

Basically some Resolve and FCPX benchmarks.

It's a pretty old article though, and I don't know if it is still relevant.

 

[rueyloon]

AND for the trashcan, you can add up to two additional eGPUs.
I doubt AMD drivers were updated in 10.12.3 for Polaris 10 (i)Mac support.

When does support Polaris 10 (most likely in 10.12.4), the nMP w/ 12-core + Dual D700s + Dual RX480s may kill it for OpenCL/FCPX w/ 136 compute units

That's part of what I'm trying to do now. I added an egpu through an akitio Casing. The set up isn't stable.
[doublepost=1485228836][/doublepost]I think with the stagnation of the 2013 macpro, there is quite some interest in how the 2010 compares. I'll be posting some benchmarks and real world comparisons soon.

 

[Daniel Reed]

That's part of what I'm trying to do now. I added an egpu through an akitio Casing. The set up isn't stable.
[doublepost=1485228836][/doublepost]I think with the stagnation of the 2013 macpro, there is quite some interest in how the 2010 compares. I'll be posting some benchmarks and real world comparisons soon.

I'm using the RX 480 in a bisonbox 2S, which is just an akitio2 backplane with the power rewired for a 400watt psu in a custom box - it actually identifies itself as Akitio Thunder 2 in system report!

It's stable for me, but the system won't restart with it plugged in. I have to reconnect it after the chime when the apple logo apears but before the loading bar. I haven't updated to 10.12.3 yet

An akitio node w/ TB3>TB2 adaptor may be the best way to go w/ nMP moving forward

 

[rueyloon]

I'm using the RX 480 in a bisonbox 2S, which is just an akitio2 backplane with the power rewired for a 400watt psu in a custom box - it actually identifies itself as Akitio Thunder 2 in system report!

It's stable for me, but the system won't restart with it plugged in. I have to reconnect it after the chime when the apple logo apears but before the loading bar. I haven't updated to 10.12.3 yet

An akitio node w/ TB3>TB2 adaptor may be the best way to go w/ nMP moving forward

are you able to confirm if the internal and external GPUs will work in unison?

 

[Synchro3]

An akitio node w/ TB3>TB2 adaptor may be the best way to go w/ nMP moving forward

Afaik no. http://forum.netkas.org/index.php/topic,13580.msg36465.html#msg36465

 

[Daniel Reed]

Afaik no. http://forum.netkas.org/index.php/topic,13580.msg36465.html#msg36465

There's a work around for it, basically have to disable the Apple blacklist.
I forget the filename, its something like IOTHUNDERBOLTFAMILY
(my notes are on a different workstation then the one i'm sitting atm)

[doublepost=1485287400][/doublepost]

are you able to confirm if the internal and external GPUs will work in unison?

YES, they work together and are selectable in benchmark tools & video applications I've tested

While the RX 480 gets a higher OpenCL compute score then a single D700 on GeekBench,
LuxMark shows the RX 480 contributing about 70-80% of a single D700 for shared compute jobs.

So in theory, while using the modified Baffin/Polaris 11 macOS driver w/ 36 CU enabled,
adding two RX 480 eGPUs is roughly the same as one-and-a-half D700s.

Once Polaris 10 is native to macOS, I suspect RX 480 performance to increase.

 

[Daniel Reed]

There's a work around for it [to enable Akitio Node on macOS],
basically have to disable the Apple blacklist.
I forget the filename, its something like IOTHUNDERBOLTFAMILY

Here's reference to it on netkas: http://netkas.org/?p=1464

AND here: http://forum.netkas.org/index.php/topic,11654.msg34141.htm

User netkas wrote:
For our less skilled friends:

For IOThunderboltFamily.kext (for binary):

find these hex bytes

55 48 89 e5 41 57 41 56 41 55 41 54 53 48 81 ec 38 01

and replace 41 57 41 56 with 31 c0 5d c3


User rominator wrote:
Can confirm this enables Razor Core at 40GB/Sec


As an aside, I recall reading, but it could be wishful thinking, Akitio stating support macOS OOTB will happen when they update Node w/ a supported thunderbolt chipset in an upcoming revised model release - and currently advertise Windows support only at this time.