MP 7,1 Need your input! What benchmarks and tests would you like to see on Mac Pro with 6900XT?

[OkiRun]

Thanks uploading the video. It's clear now why the short Pre render / Export time and high FPS.
You need to open Motion project file that includes layers, effects and so on and not the video file that is in the media folder (I attached a screen shot ) then follow the test process, this will lead to correct results

Sorry about that confusion.
FPS ranged consistently between 10 to 13 for last ten seconds; it fluctuates while playing
Export 1 min 33 seconds
RAM preview 1 min 22 seconds

 

[rondocap]

Sorry about that confusion.
FPS ranged consistently between 10 to 13 for last ten seconds; it fluctuates while playing
Export 1 min 33 seconds
RAM preview 1 min 22 seconds

Yeah that’s about the same result I get with my two Vega ii as well

 

[motion777]

I am surprised that in most of my tests, dual W5700x run pretty close to dual Vega ii GPUs. This is in Red Raw 6k, 8k, Canon raw 8k, Pro Res raw 6k, etc. Even in Octane X, and motion, they seem to run pretty even.

I wonder where the Vega ii advantage of more VRAM would come into play? Maybe I need to test noise reduction and in program affects, as export times and things like that are too close.

Now, dual 6900xt is definitely an improvement over both in the tasks I ran- seems to be at least 20% better in most cases than the other W5700x or Vega ii. Sometimes more.

Even in the candle test it's close, and the newer rocket test from here too, with blur and noise reduction. Once again the 6900xt is faster. But I'm more surprised how capable the dual W5700x is vs the Vega ii duo, much better bang for the buck for video.

 

[motion777]

I was surprised about this well, I thought that the difference will be more dramatic.
I wonder if that's the case in 3d software like Cinema 4D for example.
If your interested in VRAM monitoring tool check out I stat menu. It actually shows VRM usage. While using Motion or FCPX could really see it filling up.
There is also this video where VRAM in relationship with video editing is explained. I link it here if it interesting for you

How is your experience with the dual W5700X Regarding stability ? one of the reasons I was leaning towards Vega II
or 6800XT/6900XT was some negative experience that some users had with the W5700X and
obviously I was hoping for better performance

 

[motion777]

Yeah that’s about the same result I get with my two Vega ii as well

 

[motion777]

Sorry about that confusion.
FPS ranged consistently between 10 to 13 for last ten seconds; it fluctuates while playing
Export 1 min 33 seconds
RAM preview 1 min 22 seconds

No Worries ! Thanks for testing it, I'm happy it worked out

 

[rondocap]

I was surprised about this well, I thought that the difference will be more dramatic.
I wonder if that's the case in 3d software like Cinema 4D for example.
If your interested in VRAM monitoring tool check out I stat menu. It actually shows VRM usage. While using Motion or FCPX could really see it filling up.
There is also this video where VRAM in relationship with video editing is explained. I link it here if it interesting for you

How is your experience with the dual W5700X Regarding stability ? one of the reasons I was leaning towards Vega II
or 6800XT/6900XT was some negative experience that some users had with the W5700X and
obviously I was hoping for better performance

Yes I use istat menus as well. It's weird even about 3D rendering programs, where you think the Vega ii and duo should do better because of VRAM, but even there the 6900xt especially has a tremendous lead - and 5700x is very close too.

Maybe if you load up a huge project that's more taxing then the benchmarks, it might make a difference.

But at the end of the day - if you look at the real world, raw live result of the 6900/5700 vs the vega ii, they have either similar FPS performance, export performance, if not considerably more in the case of the 6900xt, which isn't even technically optimized yet as much as it can be with an MPX and proper module.

Imagine even a W6800 with 32GB of VRAM? Likely to make the Vega ii duo obsolete in pretty much every use case.

And what's infinity fabric doing for real world results? The dual W5700x does not have it, but seems to be on par with a Vega ii linked with IF to another Vega ii.

According to Apple, using 2 separate Vega ii is better than the Duo for video work, because it takes advantage of the 16x bandwith on each GPU, which can be important in 8k, stuff like that, otherwise they are similar. The Duo shares a PCIE 16x with both GPUs onboard, so in some cases it can be a bottleneck.

 

[diamond.g]

Yes I use istat menus as well. It's weird even about 3D rendering programs, where you think the Vega ii and duo should do better because of VRAM, but even there the 6900xt especially has a tremendous lead - and 5700x is very close too.

Maybe if you load up a huge project that's more taxing then the benchmarks, it might make a difference.

But at the end of the day - if you look at the real world, raw live result of the 6900/5700 vs the vega ii, they have either similar FPS performance, export performance, if not considerably more in the case of the 6900xt, which isn't even technically optimized yet as much as it can be with an MPX and proper module.

Imagine even a W6800 with 32GB of VRAM? Likely to make the Vega ii duo obsolete in pretty much every use case.

And what's infinity fabric doing for real world results? The dual W5700x does not have it, but seems to be on par with a Vega ii linked with IF to another Vega ii.

According to Apple, using 2 separate Vega ii is better than the Duo for video work, because it takes advantage of the 16x bandwith on each GPU, which can be important in 8k, stuff like that, otherwise they are similar. The Duo shares a PCIE 16x with both GPUs onboard, so in some cases it can be a bottleneck.

If you need bandwidth the Vega is going to be faster. The infinity cache isn't useful in the cases where you need bandwidth.

 

[rondocap]

If you need bandwidth the Vega is going to be faster. The infinity cache isn't useful in the cases where you need bandwidth.

What are some use cases where the bandwidth advantage of the Vega come into play? Real time effects or color grading in video editing? Or more towards the 3d rendering side?

There are some things that are still very much CPU bound, like some raw video formats, so the GPU isn't as effective, but other things do utilize the GPU considerably more as well.

 

[diamond.g]

What are some use cases where the bandwidth advantage of the Vega come into play? Real time effects or color grading in video editing? Or more towards the 3d rendering side?

There are some things that are still very much CPU bound, like some raw video formats, so the GPU isn't as effective, but other things do utilize the GPU considerably more as well.

Data sets in memory that exceed the (128MB) Infinity Cache size on the 6900XT operate at the slower .5 TB speed, where on Vega, using HBM2, sees transfer rates stay at 1TB. Basically 3D rendering should take advantage of the cache, video transcoding will not.

Unlike the graphics-oriented AMD RDNA™ family, the AMD CDNA family removes all of the fixed-function hardware that is designed to accelerate graphics tasks such as rasterization, tessellation, graphics caches, blending, and even the display engine. However, the AMD CDNA family retains dedicated logic for HEVC, H.264, and VP9 decoding that is sometimes used for compute workloads that operate on multimedia data, such as machine learning for object detection.7 Removing the fixed-function graphics hardware frees up area and power to invest in additional compute units, boosting performance and efficiency

I've mentioned to @leman that I thought CDNA would be better for the tasks that folks using a MacPro would need, but there was some kickback saying that Apple actually recommends developers use those fixed function units.

 

[rondocap]

Data sets in memory that exceed the (128MB) Infinity Cache size on the 6900XT operate at the slower .5 TB speed, where on Vega, using HBM2, sees transfer rates stay at 1TB. Basically 3D rendering should take advantage of the cache, video transcoding will not.


I've mentioned to @leman that I thought CDNA would be better for the tasks that folks using a MacPro would need, but there was some kickback saying that Apple actually recommends developers use those fixed function units.

Thank you for the insight. Best GPU selection on the Mac Pro is actually somewhat unclear to each use case, Apple doesn't provide much information and since it's a small user base, we have to do what we are doing here to hash out what works best, unlike on the PC workstation side where there is a lot more information and clearer guidelines.

I think most people would be surprised that a $999 msrp 6900xt or even 2 W5700x can beat or keep up with a $5600 Vega ii duo in most video tasks, aside from some 3d rendering use cases. But that's Apple structure of products - they keep older GPUs up at the same price unlike on the PC side, I guess.

 

[Somian]

Can try if it works on an 8k tv with HDMI 2.1 if you have access to one? (7680*4320@60hz with 4:4:4 color) on macOS?

 

[rondocap]

Here is the beauty of the 28 core cpu and 4 GPUs in action, dual W5700x and dual 6900XT connected via eGPU

 

[OkiRun]

View attachment 1797845

Here is the beauty of the 28 core cpu and 4 GPUs in action, dual W5700x and dual 6900XT connected via eGPU

I don't think this will be possible combination on the upcoming AS Mac Pro Mini Tower.

 

[leman]

Data sets in memory that exceed the (128MB) Infinity Cache size on the 6900XT operate at the slower .5 TB speed, where on Vega, using HBM2, sees transfer rates stay at 1TB. Basically 3D rendering should take advantage of the cache, video transcoding will not.

That's a tricky one. For transcoding, most of your data is probably coming from outside the GPU, so you will be limited by the PCIe connection. For the cache to be ineffective, you are either doing true random memory access (which will kill your performance anyway), or you are processing massive amounts of data in a computationally shallow workload that hits memory only once. My expectation is that the Infinity Cache will work well is most relevant scenarios.

I've mentioned to @leman that I thought CDNA would be better for the tasks that folks using a MacPro would need, but there was some kickback saying that Apple actually recommends developers use those fixed function units.

If your app is doing image processing, using the rendering pipeline for it can make things more efficient, especially on Apple hardware. If you are doing pure compute, you don't need the rendering fixed function. AMD's approach makes a lot of sense. then again, Apple GPUs are a bit different — their compute units are already streamlined, and since they are using a more dense process, they have less of a space issue.

 

[rondocap]

That's a tricky one. For transcoding, most of your data is probably coming from outside the GPU, so you will be limited by the PCIe connection. For the cache to be ineffective, you are either doing true random memory access (which will kill your performance anyway), or you are processing massive amounts of data in a computationally shallow workload that hits memory only once. My expectation is that the Infinity Cache will work well is most relevant scenarios.



If your app is doing image processing, using the rendering pipeline for it can make things more efficient, especially on Apple hardware. If you are doing pure compute, you don't need the rendering fixed function. AMD's approach makes a lot of sense. then again, Apple GPUs are a bit different — their compute units are already streamlined, and since they are using a more dense process, they have less of a space issue.

Thank you for the insight- very interesting how these GPUs work with the Mac specifically, i think unless one is highly technically informed, which a lot of the Mac Pro base that do video work may not be - it takes a while to figure the optimal hardware for ones use case.

Data sets in memory that exceed the (128MB) Infinity Cache size on the 6900XT operate at the slower .5 TB speed, where on Vega, using HBM2, sees transfer rates stay at 1TB. Basically 3D rendering should take advantage of the cache, video transcoding will not.


I've mentioned to @leman that I thought CDNA would be better for the tasks that folks using a MacPro would need, but there was some kickback saying that Apple actually recommends developers use those fixed function units.

if the infinity fabric linking two Vega ii makes the gui and communication faster - what do you think happens if a Vega ii duo is installed along with a single Vega i? Does the internal IF on the duo still give the same benefits, or is it negated because of the presence of the single Vega i?

 

[rondocap]

That's a tricky one. For transcoding, most of your data is probably coming from outside the GPU, so you will be limited by the PCIe connection. For the cache to be ineffective, you are either doing true random memory access (which will kill your performance anyway), or you are processing massive amounts of data in a computationally shallow workload that hits memory only once. My expectation is that the Infinity Cache will work well is most relevant scenarios.



If your app is doing image processing, using the rendering pipeline for it can make things more efficient, especially on Apple hardware. If you are doing pure compute, you don't need the rendering fixed function. AMD's approach makes a lot of sense. then again, Apple GPUs are a bit different — their compute units are already streamlined, and since they are using a more dense process, they have less of a space issue.

I may have both a Vega ii duo and two solo Vega ii to test soon - what tests can I run to reveal the differences between them, specifically the pcie bandwidth limitations?

im thinking 8k raw, 4444, 30p, 60p, etc. anything else that will stress the bandwidth? DenLise or other effects?
‘’I’m assuming 3D rendering not much difference since it’s not as pcie bound.

 

[jerrygladh]

Which brand is your 2x6900 xt?

Jerry

 

[rondocap]

Which brand is your 2x6900 xt?

Jerry

They are both the AMD Reference editions, so fit perfectly fine

 

[rondocap]

Some results:

 

[cobra521]

rondocap,​

Thank you for what clearly involved a lot of time and effort!

One interesting sidelight if I interpret that part of the experiment correctly: W5700X * 2 is in general more than twice as fast as one card by itself. If true, impressive for just those two cards.

Adding a couple of 6900XT cards sometimes buys a fairly good improvement, sometimes maybe not as much as one would hope. Obviously depends on the task at hand.

Tom

 

[rondocap]

rondocap,​

Thank you for what clearly involved a lot of time and effort!

One interesting sidelight if I interpret that part of the experiment correctly: W5700X * 2 is in general more than twice as fast as one card by itself. If true, impressive for just those two cards.

Adding a couple of 6900XT cards sometimes buys a fairly good improvement, sometimes maybe not as much as one would hope. Obviously depends on the task at hand.

Tom

yeah, it can be a bit weird sometimes With multi gpus, but almost always 2 of the provide good results in most tasks.

Dual w5700x is pretty amazing, makes me question the much more expensive dual Vega ii - the more vram doesn’t seem to matter nearly as much.

I have a Vega ii duo coming to test, I want to see if a Vega ii duo + single Vega ii work well together In final cut.

i also want to try and find a bandwidth limitation of using a Vega ii duo vs 2 separate Vega ii, possibly 12k braw may show some as it’s heavy gpu usage

Dual w5700x (dual Vega ii is similar for more money) and the best choice for most until the w6800 mpx comes, I’m expecting them to be impressive like the 6900xt is.

 

[DFP1989]

@rondocap the dual W5700X results are very interesting, especially since combining the 6900 and W5700X doesn’t look to do much for me in FCPX.

Do you notice any improvement in the dual W5700X with HEVC 10-bit 422 footage?

 

[rondocap]

@rondocap the dual W5700X results are very interesting, especially since combining the 6900 and W5700X doesn’t look to do much for me in FCPX.

Do you notice any improvement in the dual W5700X with HEVC 10-bit 422 footage?

Not in terms of playback, but with 10 bit 422 I didn’t do any other tests so far. M1 macs handle those much better at the moment, but w5700 has decent enough hardware for exporting it

 

[demitri]

They are both the AMD Reference editions, so fit perfectly fine

Would you mind telling me where you bought them? And how loud is it when idle vs full usage? The near dead silence of the W5700X is one of my favorite things about it? And lastly, would there be a noticeable advantage to using both my current W5700X installed alongside the 6900XT?

Thank you for your time!