https://www.amd.com/en/products/professional-graphics/radeon-pro-vii
[automerge]1589378317[/automerge]
[automerge]1589378317[/automerge]
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
MP 7,1 AMD Radeon Pro VII
- Thread starter Xenobius
- Start date
-
- Tags
- pro radeon radeon vii vii
- Sort by reaction score
Fewer streams and CU , slightly better compute and better memory bandwidth than the WX9100 / Vega F.E. It uses a smaller 7nm process but also the same power consumption ratings ( officially - we'll see what its like in reality in time ) . It should run cooler , but AMD probably bumped up performance to negate the thermal savings . Same GPU as the straight Radeon VII ( Vega 20 ) ... Waste of money to install this in a cMP . Might be fun to install in a MP7,1 , though .
Last edited:
It has 1/2 rate double precision. It's basically the MI50. It has 1TB/s memory bandwidth and ECC memory.
It has spades of compute compared to the cards you mentioned. Normal Radeon VII clocks in at 3.5 TFLOPS (1/4 rate) while this gets 6.5 TFLOPs (1/2 rate). The Radeon Pro W5700X has 593.8 GFLOPs and the VEGA FE has 819 GFLOPs for comparison.
The Radeon Pro Vega II Duo ($5,600) has 1/16th rate, so only has ~881 GFLOPS double precision.
It's also clearly binned better as it has a TDP of only 250W compared to 300W of the Radeon VII.
Last edited:
2019 Mac Pro - GPU options
real world speed test results for performance minded Macintosh users
barefeats.com
Pro W5700X versus other 2019 Mac Pro GPUs
real world speed test results for performance minded Macintosh users
barefeats.com
Lou
... and it uses Infinity Fabric to combine GPU memory (similar to Pro Vega II):
https://www.amd.com/system/files/documents/ifl-solution-brief-radeon-pro-vii.pdf
https://www.amd.com/system/files/documents/ifl-solution-brief-radeon-pro-vii.pdf
Largely a more affordable and simpler to make Apple Pro Vega II 'Solo'.
If specs here are right the "peak clock" is the same as the Vega II
AMD Reveals Radeon Pro VII: A Workstation Card For When You Need It All
Has the Infinity Fabric link that the plain commercial VII doesn't have. ( similar to MI/50 which may/may not be priced higher. ).
Apple may not have the 1/2 Double precision setting and didn't certify out the ECC mode but there shouldn't be a surprise here if the Pro Vega II is doing reasonably well. That means the demand is high enough to keep churning these out for now.
These are about $1K cheaper than Vega II. Loose about 16GB of HBM2 memory. ( so the VII model only needs 4-hi memory stacks instead of 8-hi . If can salvage some failed 8-hi as 4-hi then all the better).
But if can fit the two cards into a Mac Pro ( e.g., slot 1 580X , slot 2 (or 3 ) Pro VII , slot 4 (or 3) ) then could get into range of a Vega II Duo for about $2K less. And loose all the Thunderbolt , switching, MPX power elements of Apple's MPX module.
I suspect that the "Arcturus' MI100 may not be too far away and that this VII will serve as the more affordable alternative for workstations on a "close enough" GPU.
[automerge]1589390626[/automerge]
Well it should because it would be going "backwards" if it didn't. The MI50 and MI60 have had this for over a year.
AMD Radeon Instinct™ MI50 Accelerator (16GB)
The AMD Radeon™ Instinct™ MI50 server accelerator delivers high levels of performance for deep learning and HPC workloads.
this is just a variation on the MI50-MI60.
Diff between MI50 and VII ( much of the same differences. )
The AMD Radeon VII Review: An Unexpected Shot At The High-End
To some extent probably same product brought back at differnent price point with different name.
A while back AMD bumped the MI50 from 16GB to 32GB of VRAM. This is in part just bring the 16GB VRAM variant back and probably at a much lower price point (and possibility with a better tweak of the thermals to support the direct side by side placement of the double width cards and the "one slot gap" solution working better. )
[automerge]1589391177[/automerge]
On several of those benchmarks the 'regular' VII and the Pro Vega II 'Solo' are essentially tied. The additional 16GB of HBM2 VRAM isn't buying very much. If you trim that off the Vega II you'd get a more affordable card and get about a three way tie.
The thermals are down from a VII so if have some computation that has a substantively long "single thread control" block in it ( and just throw max (possibly overclocked) boost speed at it to plow through it won't do as well. But large heavyweight double precision simulations probably will do much better if have a very high "embarrassing" parallel workload.
Last edited:
The MP7,1 with macOS 10.15 and support for Infinity Fabric makes putting two of these in make this a substantive jump over what a cMP can do for several workloads with apps that are aware of the Infinity Fabric advantages.
But what workflows will benefit from double precision ? Not likely for video editing , which is where a card like this could wind up in a macOS System . Even if a single Radeon Pro VII could match the performance of Dual GPU WX9100 / Vega F.E. configuration for video editing ( both FCPX and Resolve love dual GPU configurations ) , there is significant cost savings going with the later . You can pull used Vega F.E.s off of eBay for $400 each and rebuild them . WX9100 is a different story .... for now .
Single precision compute performance of the above cards is roughly the same - what most users would use .
Same with video ECC , nice but hardly necessary .
Memory bandwidth of the RPVII would be about the only exciting feature here . It's where paper specs have a hard time showing fair comparisons . And Infinity Fabric , but ...
Infinity fabric is useful only in a multiple GPU System , so you need to plunk down extra cash for two+ of these babies .
I have little doubt both the WX9100 and the RPVII are top binned . But with proper maintenance , GPUs last a long time anyways . There might be the power savings thingy , but AMD has shown in the past they'll just make things hot and power hungry to boost performance to match nVidia anyways . We'll see when somebody deploys one .
[automerge]1589393326[/automerge]
Nearly $4000 of cards in a cMP ??? Seriously ? And I build custom workstations , too . They'd be better installed in a MP7,1 where they can stretch their legs .
Last edited:
First, AMD didn't make this solely for a macOS system. Probably not even primarily for a macOS system (the PCI-e v4.0 doesn't do much additional for the Mac Pro 2019). You pulled the ProVII versus WX9100 (and friends) comparison chart from the Anandtech article. Well, there is a AMD slide in the article that clearly outlines AMD's primary targets.
AMD Reveals Radeon Pro VII: A Workstation Card For When You Need It All
HPC (e.g., supercomputers) largely are double precision workloads when it comes to simulations. AMD has some substantive 'wins' for some supercomputer deployments. Not everybody has left Macs that do HPC ( more than a few but not everybody). Physical system design. Again, likely not everyone has left "the building". The major point is that a card like this lets the Mac Pro trend water is some of these markets outside of video/audio editing. ( it is a Dell/HP workstation killer but should be able to retain more than a few folks who haven't left and get a few new ones over time. )
Broadcast & media probably does have some implicit presumptions about leveraging Infinity Fabric to "move the needle" in terms of performance.
Second, even macOS isn't solely video and audio editing. That probably is the biggest subset of Mac Pro 2019 users but it isn't the whole set. "Who" the users are doesn't matter as much as how aligned this is with the other GPU with supported drivers ( Vega II ). This is basically the same thing with a bit less memory, so the long term support by Apple is likely to last at least as long as the Mac Pro 2019 service life.
If Apple happened to not choose to track "big navi" for some reason that card may or may not light up in a future macOS version.
For software that leverages the Infinity Fabric in dual it wouldn't be "match", it is more a question of how far behind the WX9100 / Vega F.E. will be.
Most users that you are sub selecting. Two of the three targeted markets it both of these make a difference ( high fidelity physical modeling and accurate data. )
Again you are mainly missing the point. It is software, not hardware where the substantive value add is. The cMP doesn't have the software ( new Metal support for Infinity Fabric) so it is being left being. It has very little to do with its abilities as a hardware box with slots.
Unlikely config but out of curiosity, any chance this would work on a Mac Pro 5,1 with existing Radeon VII drivers? Under Mojave ideally?
(understanding the need for supplementary power)
(understanding the need for supplementary power)
It is really going to depend upon what workflow and those app's ability to utilize what is present.
If you are looking for a simplistic unidimensional metric then single precision (TFLOPs ) :
| card | TFLOPS | delta from Vega II | VRAM (GB) | DP TFLOPs |
| VII | 13.6 | -0.5 | 16 | 3.5 |
| W5700X | 9.4 | -4.7 | 16 | ~0.618 |
| Pro VII | 13.1 | -1.0 | 16 | 6.5 |
| Vega II | 14.1 | 0 | 32 | ? 3.5 |
| Vega II Duo | 28.3 | +14.3 | 2 x 32 | ? 2 x 3.5 |
The price points of all of the above are vastly different from top to bottom. Vega II has four more CU (compute units: 64 shader processors per CU. so if counting shaders 256 more) over the Pro VII.
If your workload only concurrently lights up 55 CU's at any one time then that extra 4 usually won't make that much of difference. If have significant single threaded sections the plain VII overclocked may be more competitive.
If data working set is 10-20GB big then the Vega II can handle that. If it is just 8GB then not much of a difference across the whole line up above.
So if money is no object and all care about is maxed out single precision on a single GPU die then the Vega II 'solo' is a winner if software can squeeze everything out of the chip possible. Putting money back on the table. it is about $1K more for less than a 10% gain. For some that less than 10% gain is worth $1K. For others, the over 100% gain is worth extra $3.8K
Not much binning necessary actually. They just need to lower their insane voltages. I am yet to see a Vega20 chip that doesn't manage to stay below 1050mV on stock frequency. On 250W TDP there probably is still way too much voltage. They could easily lower that even further.
My Radeon VII runs rock solid on stock frequency and a GPU voltage of 972mV. Stays well below the 160W mark under full load. Overclocked to 1903MHz I need 1032mV which still results in lower consumption than stock with lower temps and more silent fans. It's a shame AMD doesn't utilize the full potential of their chips OOB.
Yeah, I did some tests as well. If I downclock my Radeon VII to match the Pro VII's 1700MHz clock speed. I only need 917mV to run the card stably (factory setting is 1093mV). That's way below 200W power consumption even under stress.
Those work station card's clock speed usually quite close the optimum speed.
Similar situation for RX580. If downclock to 1243MHz to match the WX7100's setting, and downvolt the card accordingly. The power draw under stress will also reduce to about 130W. Way below a normal RX580 will draw, but just about 7% performance penalty.
Boost clock actually isn't the maximum clock you see in the PowerPlay Table. I guess the Radeon Pro VII has the same 1801MHz maximum clock as the regular Radeon VII.
The Apple cards only have 1/16th rate double precision as far as I am aware.
I know, but I expect that will with a lower boost clock speed to stay within 250W typical power draw.
And TBH, that extra little bit boost clock doesn't make much difference in real world, but can draw quite a lot of extra power.
that would be odd. That was AMD's initial intention with the 'mainstream" Vega VII , but they backed out of that relatively quickly before shipping it.
"...
Looking to clear things up, AMD put out a statement:
' The Radeon VII graphics card was created for gamers and creators, enthusiasts and early adopters. Given the broader market Radeon VII is targeting, we were considering different levels of FP64 performance. We previously communicated that Radeon VII provides 0.88 TFLOPS (DP=1/16 SP). However based on customer interest and feedback we wanted to let you know that we have decided to increase double precision compute performance to 3.52 3.46 TFLOPS (DP=1/4SP)' ..."4
The AMD Radeon VII Review: An Unexpected Shot At The High-End
The Pro Vega II is aimed at "creators and early adopters" at least as much as the VII was. At almost +$2K in price increase to be dragging way behind the VII would need some really good justification. ( better thermals or better binning or something. )
Apple has made some odd ball decisions. "Give us something worse than your dramatically less expensive card" may be the call they made. ). If one of the motivators for AMD to drop this card now is that Apple isn't consuming as many "Vega 20" dies as their max projection and AMD needs another outlet for the dies they have queued up then perhaps that wasn't a great idea.
For folks who need period DP (FP64) they are going to drop all the MPX options if the Vega II is deliberately deeply kneecap ( the ability is there in the die to set the level 1:4 which is still off the top mark.)
And yet it is suppose to be a complete OpenCL replacement...... Apple's inattention to detail in the pro workstation space continues. [ Pretty good chance this is a gap (or at best weakness) the Apple GPU also shares. ]
I think it probably is compatible with Mojave. There are a couple Vega 20 device IDs in the drivers that I haven't been able to identify yet. It's likely this is one of them. It may not identify it properly in About This Mac though. It would probably say something generic like "Radeon Vega."
This GPU looks like a nice upgrade over the consumer Radeon VII.
Ant
Anyone have a cMP and a Radeon Pro VII to try? I’m running an undervolted kext version Radeon VII in my own CMP 12 x 3.05. Very nice in DaVinci Resolve and FCPX. But there’s no more new ones available at less than 1000 euros these days.
Anyone have a cMP and a Radeon Pro VII to try? I’m running an undervolted kext version Radeon VII in my own CMP 12 x 3.05. Very nice in DaVinci Resolve and FCPX. But there’s no more new ones available at less than 1000 euros these days.
Answering your call a whopping 4 years later, I am delighted to report my upcoming project. About A week ago, I was able to snatch a brand new sealed in box Radeon Pro VII workstation card for 300 bucks. I did sent it to my friends in Poland to get it flashed with boot screens and modified device ID for a perfect fit for my 2012 Mac Pro 5.1.
More to come with a short performance review towards the end of October.
As promised, attached you see my new Vega VII workstation makeover with blue accents this time. (My dual tray is in storage at the moment)
Since I had many different GPUs in this thing I am used to the Vega VII level of performance or higher. What really really surprised me is the low noise duct fan. The reference Vega 64 / 56 back in the day was really a very bad designed and noisy card. Because of this I was very sceptical. However, since the 6800XT is not supported in the Sonoma, I sold my trusted mutant card and fell back to my 580 with AIO water block. But this time AMD surprised me full heartedly. The Radeon Pro Vega VII is very calm and the duct fan is not very noisy. I think that is the most important and surprising information. Big kudos to AMD Pro Series, they really did their homework when it comes to noisy duct fan solutions. Another surprise was the weight of that card and the fact that it's shroud is metal. It is really next level when it comes to AMD build quality. The hefty 1899,00 Dollar price tag did go somewhere.
Just for kicks, you will find my past RX 580 AIO setup attached as well.
Attachments
