MP 7,1 Vega II + WX5700.. possible?

[chfilm]

Hey guys,

I'm wondering if it’s gonna be possible to combine two different GPUs if the WX5700 is gonna come with infinity fabric as well...? Unlikely, right?

I’m really kind of wondering if we were lured into some kind of trap with the single Vega II as opposed to two WX5700 which are gonna be more powerful and will probably not lost a lot more than the single Vega..
On the other hand the Vega should be powerful enough for pretty much anything for now and we can always add a second one later..

really looking forward to afterburner Tests to understand how effectively it can take a load off the rest of the system And if it’s actually a must have...

 

[profdraper]

As I understand it: the 16Gb WX5700 is less powerful than the 32GB Vega II, likely why its been introduced given the enormous gap between the Radeon Pro 580X & the Vega. They offer an option for 2 x WX5700s but don't indicate if that's in a MPX Bay or not. If it were, then no doubt this would use infinity fabric to 'bind' the two cards, just like for the Vega II Duo. That'd be a 32GB rig but would need to be lower priced than the Vega II to make it worthwhile. Otherwise, don't mix GPU card types, the NLEs make best use of matched pairs (FCPX, Resolve Studio).

 

[bsbeamer]

All depends on the app and tasks. Premiere Pro is getting better at multiple GPU and eGPU with dGPU, but it’s not used for all GPU/METAL tasks. Absolutely accelerates hardware encode (export). The added VRAM sometimes helps, even though it’s split between two cards. Makes a big difference for VR, not as much for regular video when just working on timelines. SOMETIMES using a headless adapter tricks the system to provide better performance with multiple GPU, if you don’t have a monitor connected to each GPU.

 

[Korican100]

As I understand it: the 16Gb WX5700 is less powerful than the 32GB Vega II, likely why its been introduced given the enormous gap between the Radeon Pro 580X & the Vega. They offer an option for 2 x WX5700s but don't indicate if that's in a MPX Bay or not. If it were, then no doubt this would use infinity fabric to 'bind' the two cards, just like for the Vega II Duo. That'd be a 32GB rig but would need to be lower priced than the Vega II to make it worthwhile. Otherwise, don't mix GPU card types, the NLEs make best use of matched pairs (FCPX, Resolve Studio).

but wouldn't one single GPU be better for NLE's than two? For instance if the Vega II = 2 x WX5700s

 

[chfilm]

As I understand it: the 16Gb WX5700 is less powerful than the 32GB Vega II, likely why its been introduced given the enormous gap between the Radeon Pro 580X & the Vega. They offer an option for 2 x WX5700s but don't indicate if that's in a MPX Bay or not. If it were, then no doubt this would use infinity fabric to 'bind' the two cards, just like for the Vega II Duo. That'd be a 32GB rig but would need to be lower priced than the Vega II to make it worthwhile. Otherwise, don't mix GPU card types, the NLEs make best use of matched pairs (FCPX, Resolve Studio).

It states now that it’s an MPX module!

 

[shuto]

From Apple's tech spec page...

AMD Radeon Pro W5700X
40 compute units, 2,560 stream processors
16GB of GDDR6 memory with 448GB/s memory bandwidth
Up to 9.4 teraflops single precision or 18.9 teraflops half precision
Four Thunderbolt 3 ports and one HDMI 2.0 port on card
Two DisplayPort connections routed to system to support internal Thunderbolt 3 ports
Support for Display Stream Compression (DSC)
Support for up to six 4K displays, three 5K displays or three Pro Display XDRs
Full-height MPX Module fills an MPX bay and uses extra power and PCIe bandwidth
AMD Radeon Pro Vega II
64 compute units, 4,096 stream processors
32GB of HBM2 memory with 1TB/s memory bandwidth
Up to 14.1 teraflops single precision or 28.3 teraflops half precision
Infinity Fabric Link connection enables two Vega II GPUs to connect at up to 84GB/s
Four Thunderbolt 3 ports and one HDMI 2.0 port on card
Two DisplayPort connections routed to system to support internal Thunderbolt 3 ports
Support for up to six 4K displays, three 5K displays or two Pro Display XDRs
Full-height MPX Module fills an MPX bay and uses extra power and PCIe bandwidth



So no mention of infinity link on with W5700X, and W5700X performance is less than Vega II. Think your Vega II purchase is going to be a beast for a while!

 

[avkills]

Yes but the Navi cards look to have better hardware H.264/H.265 encoders. It will be an interesting benchmark once the W5700X is released. That is why I basically stuck with the 580x for now. Wait and see what the real verdict is. If encoding is way faster on the W5700X, then that is the way I will probably go.

 

[gazwas]

Yes but the Navi cards look to have better hardware H.264/H.265 encoders. It will be an interesting benchmark once the W5700X is released. That is why I basically stuck with the 580x for now. Wait and see what the real verdict is. If encoding is way faster on the W5700X, then that is the way I will probably go.

The exact reason why I'm waiting to order my Mac Pro.

The Vega II is on paper is a much more powerful GPU and comes with a lot more memory hence its high cost but the W5700X has all the latest bells and whistles that the Navi architecture brings over the older Vega and wouldn't be surprised it the differences were not that great and even favouring the newer GPU is some specific tasks.

However, what we don't know is the cost and considering its going to be a full MPX module like the Vega II it might be a lot higher than we're all hoping, especially if it uses the infinity link stuff for dual GPU.

 

[chfilm]

I’m wondering what will happen if we put a regular off the shelf 5700 inside, just as some sort of h264 encoding afterburner because of the new video encoding Hardware that the vega is missing.

Hopefully those fans won’t spin up in sleep mode... less vram ok, but I already have 32gigs, so... ah it will be Interesting.

 

[avkills]

I think I would want the 16GB VRAM that the "Apple" branded WX7500 is offering. Of course price is going to be a big consideration. If it is only a few hundred off from a 32GB Vega II and the benchmarks favor the VegaII then the Vega II will be my choice.

 

[gazwas]

I’m wondering what will happen if we put a regular off the shelf 5700 inside, just as some sort of h264 encoding afterburner because of the new video encoding Hardware that the vega is missing.

Hopefully those fans won’t spin up in sleep mode... less vram ok, but I already have 32gigs, so... ah it will be Interesting.

I'd wait until 10.15.3 before doing anything personally as I'd imagine this is what Apple are waiting for to release the W5700X.

 

[chfilm]

I'd wait until 10.15.3 before doing anything personally as I'd imagine this is what Apple are waiting for to release the W5700X.

Yea software is for sure the holdup right now.
Not doing anything right now anyways, I’m happy with my Vega, I’m just curious how much this card would speed up h264 encoding.

 

[jasonmvp]

As I wrote in another thread, I'll probably get two of the W5700X modules and sell my OEM Vega II. The editing software I use will make use of two GPUs, and the hardware encoder on the 5700 is newer and more capable.

 

[jasonmvp]

As I wrote in another thread, I'll probably get two of the W5700X modules and sell my OEM Vega II.

Whoops, got a bit ahead of myself. I mistakenly thought there was one more x16 slot in the Pro than there actually is. I already have two x16 cards in the machine for my two HighPoint cards. That makes having two MPX modules installed impossible.

A potential work-around (but suboptimal one) would be a W5700X MPX, and then another aftermarket 5700XT card that can squeeze into a 2-slot form factor. And there are only a couple of those out there; most are 2.75-slot. That would give me two independent 5700s, but one of them would have half the VRAM as the other.

Hm.

 

[bsbeamer]

If you physically can't fit and power everything in the tower, PCIe expansion may be necessary. The machine (theoretically) should have enough slots/bandwidth. MP5,1 and previous could run multiple GPU setups via PCIe expansion back in the day. All of that being said, MPX needs to stay within MP7,1 and don't see that ever being used in an external chassis the right way.

 

[jasonmvp]

If you physically can't fit and power everything in the tower, PCIe expansion may be necessary.

The only way I could theoretically do an external solution and have two MPX modules would be to move one of the HighPoints out. I just don't see that ending well. At all. In fact, I suspect that'll suck vaccuously as far as storage throughput is concerned, even though TBolt3 is (supposedly) capable of 40Gbits/sec.

I can keep everything internal, doing a single MPX and an aftermarket 5700XT by moving one of the HighPoint cards down to the second MPX slot; it's otherwise just a x16 slot. That will leave the "top" most x16 slot open with an empty x8 next to it. The aftermarket GPU could pop into that x16. But again, it would have to be a 2-slot card; most of the currently available 5700XT cards are 2.75-slot.

 

[deconstruct60]

The only way I could theoretically do an external solution and have two MPX modules would be to move one of the HighPoints out. I just don't see that ending well. At all. In fact, I suspect that'll suck vaccuously as far as storage throughput is concerned, even though TBolt3 is (supposedly) capable of 40Gbits/sec.

I suspect there is a decent chance the Highpoint card works in the x8 slot(s).

Install PCIe cards in your Mac Pro 2019
"...

• Slot 6: up to x8 lanes (single-wide)
• Slot 7: up to x8 lanes (single-wide) ..."

I'm pretty sure Apple is describing the electrical PCIe connection, not the physical slot. x8 electrical is going to be better than Thunderbolt. It isn't x16 but it also is just as relatively far from x4. As long as the SSD card has a real PCIe switch there it should adapt.

The only problem more so would be if the cards are not single-wide. 1.25 or 1.5 wide would be an issue if also need to use the other slots in that area for something else.


A x16 and a x8 would be more than putting a x16 extender adapter card in slot 5 and then having the external box have two x16's share that one x16 path if doing concurrent read/writes on both. Pragmatically would get two x8's out of that when both in full contention.

 

[jasonmvp]

I suspect there is a decent chance the Highpoint card works in the x8 slot(s).

Work: yes. Work at full speed: no. My setup is a bit unique in that I have two full cards RAID01'd together. So having one card at x16 and the other x8 would provide for some "fun" FS experiences, I suspect. It's not really worth the risk.

 

[deconstruct60]

Work: yes. Work at full speed: no. My setup is a bit unique in that I have two full cards RAID01'd together. So having one card at x16 and the other x8 would provide for some "fun" FS experiences, I suspect. It's not really worth the risk.

Then both could be x8 if the gap is an issue (but software RAID of those two should be able to deal with that). Depends on trade-offs want to make with multiple GPUs versus the RAID cards.

If want to lock down two x16 SSD add-in-cards local to full bandwidth the CPU PCI-e lanes then you're only choice to punt the "compute" GPU(s) off of a full x16 electrical slot. (or possibly outside the box if want more.)

Apple labels four of the slots as x16 , but slot 4 is highly likely sharing bandwidth with other slots and/or devices. There is only 64 CPU PCI-e lanes. So three hard linked, non sharing x16 bundles means there is only one x16 left on the CPU. Can't have a x16 and other slots without there being some sharing going on. I suspect slot 4 is only close of x16 only if don't do lots of other stuff with almost all of the other non-x16 slots.

So it doesn't really matter if remove the MPX modules... there are probably only three full pass through, no sharing x16 slots available.

Slot 1 MPX (full size) , slot 3 RAID-a , slot 4 double wide compute GPU (doesn't go too wide over double) , slot 5 RAID-b ( and not quite max full length to give power for slot 4 room). That will essentially move the compute GPU to something sub x16 but probably "close enough" for embarrassingly parallel work that fits in the GPUs VRAM as a large cache. And even if Slot 1 wasn't an MPX module you'd still have the same number full x16 slot constraints.

It is a width of the compute GPU card issue more than an MPX specific issue.

 

[jasonmvp]

Then both could be x8 if the gap is an issue (but software RAID of those two should be able to deal with that). Depends on trade-offs want to make with multiple GPUs versus the RAID cards.

I had briefly considered that. The storage is currently pushing 10GB/sec read and +9GB/sec writes in simple benchmarks. Obviously that would be capped at 8 and 8 respectively if I moved the cards to x8 slots. The question is: would I notice the loss of the 1-2GB/sec?

 

[avkills]

So I have decided to wait some more; the WX5700 isn't the performance I was hoping for. Going to see what RDNA2 is all about first. Going to be tough to beat the 3090 though.