MP 7,1 Waiting for Mac Pro 7,1 A1991 (no more)

[ManuelGomes]

AMD Radeon Pro VII leaked. 16GB HBM2, 250W TDP, 13.1 TFLOPs (1/2 FP64). Wouldn't it make a little more sense than Navi for MP? Although, there are already Vega based cards of course, but if the cost would be similar to the Navi card maybe it would be a better choice. PCIe 4.0 but that wouldn't be much of a selling point here.
Let's wait for the Big Navi, or nVidia killer, call it what you will

 

[deconstruct60]

AMD Radeon Pro VII leaked. 16GB HBM2, 250W TDP, 13.1 TFLOPs (1/2 FP64). Wouldn't it make a little more sense than Navi for MP? Although, there are already Vega based cards of course, but if the cost would be similar to the Navi card maybe it would be a better choice. PCIe 4.0 but that wouldn't be much of a selling point here.

The cost is $1.8K

AMD Reveals Radeon Pro VII: A Workstation Card For When You Need It All

www.anandtech.com

Relative to a Vega II 'solo' card it is more affordable. It is no where near the current Navi cards. Even the W5700X ($1K) or the more plain W5700 ( ~$800). If trying to buy something better than a 580X then both of those are lower cost. Quite likely that even the upcoming "Big Navi" won't be priced higher than the W5700X ( if it even gets macOS drivers any time in 2020... probably not. )

Pro VII is basically a more affordable Pro Vega II. Cost discount because of no Thunderbolt infrastructure, no video switches , a largely recycled MI50 design , and only 4-hi HBMv2 memory ( so 1/2 the VRAM capacity). But the basic GPU die is exactly that of the Pro Vega II. AMD just gets to make more of those dies ( and can harvest the few that don't quite flush out to 64CUs. ).

Pro VII is somewhat the return of the original configuration of the MI50 at probably better pricing. ( Likely not an accident the day before Nvidia releases their "even more expensive" monster computational GPU card line up upgrades. ). That $1.8K will look modest next to those prices.

Let's wait for the Big Navi, or nVidia killer, call it what you will

Actually the "Navida killer" isn't Big Nvidia. It is actually the smallest o the three upcoming models.

".... He was told that Navi 23 is referred to internally as the “Nvidia killer”, which is an odd name for such a small GPU, smaller than even the 5700XT. However, it might not be a killer in the sense of performance, but in value, as RedGamingTech has speculated recently. ..."
https://adoredtv.com/big-navi-rumored-to-have-die-size-of-505-mm2/

The whole point of Navi was to more so go after the middle of the consumer market not the upper edge. Basically more of an overlap with Polaris that what Vega primarily addressed. It is a value proposition killer in part with what AMD has been doing for a long while in the GPU space. I doubt AMD is looking to top Navi on performance at the top end either. More $/performance is the more likely approach they'll take. "Big Navi" is likely to be leapfrogged by Nvidia either right before or relatively shortly after "Big Navi" ships at the top end. It isn't going to be some kind of 'killer' in pure performance ( and cost aside ) terms.


If it is smaller than the 5700XT that makes it more a good 580X candidate than mucking around at the "big" end of MPX modules.

It is a bit questionable whether Apple is going to jump on board with the Navi versions that threw lots of transistors and silicon at the real time rending stuff. All the more so if Metal really has other primary focus points in the 2019-2020 timeframe. Navi drivers haven't been spectacularly stable and chasing after a newly fledged subsystem in the GPUs doesn't seem likely to improve the stability. Won't be surprising if Apple passes on "Big Navi" for the CDNA , Arcturus model that is an extension of Vega lineage which they do have stable drivers for.

A couple of weeks again. ( there are other MI100 sites that have popped up over the last year. )
https://wccftech.com/amd-cdna-architecture-radeon-instinct-arcturus-gpu-120-cu-7680-cores/

If AMD has somewhat glued two Vega 20 computational dies together (with shared memory and un-core infrastructure ) that would make implementing the "Vega II duo" functionality simpler. And if can pack two into a full sized MPX modules then pragmatically have a 'quad' . If Apple could wrangle AMD into letting them price that new system at the current "solo"/"duo" pricing then those probably would work better in the market.

 

[AidenShaw]

the Vega II duo s are capable of 1:2 fp64 while very niche is the fastest fp64 card available

There's a mistake somewhere.... TechPowerUp specs the Vega II Duo at 1:16, an 880 Gflops theoretical performance (I think per GPU). They show a Titan V (V100 Volta GPU) at 1:2, and 7.45 Tflops theoretical.

If TechPowerUp is correct, then in the space of a Vega II Duo with aggregate 1.76 Tflops FP64 you could put a pair of Titan V's with the same amount of VRAM and 14.90 Tflops.

Other sources like Anandtech suggest that AMD is limited the GPUs to 1:4 or 1:8.

Even the most optimistic of the 1:2 reviews put the Vega II at 6.9 Tflops - still significantly less than the V100 GPUs.

And the goalposts may be moving - tomorrow is NVIDIA CEO Jensen Huang’s keynote at GTC Digital - so the new standard could be the RTX 3080 Ti and any related Quadro cards.

 

[mattspace]

Let's wait for the Big Navi, or nVidia killer, call it what you will

It'll be hand-delivered by Godot himself, the arrival date will be "tomorrow", etc.

 

[deconstruct60]

There's a mistake somewhere.... TechPowerUp specs the Vega II Duo at 1:16, an 880 Gflops theoretical performance (I think per GPU). They show a Titan V (V100 Volta GPU) at 1:2, and 7.45 Tflops theoretical.

1:16 would put it equivalent of the ratio the 5700 has (or the old Vega 10 implementation). The plain VII has 1:4 ratio (and it is also the same Vega 20 die implementation). Why AMD would dial the Vega II even further back than 1:4 would be odd. That is already a kneecaping. Apple throws out FP16 ( half precision ) as the alternative to single precision (FP32) in their specs.
Unless there is some deep discount Apple gets for knocking it down to 1:16 ( AMD can harvest some more or capping puts a more uniform thermal ceiling on the chip. ). It is there if AMD needed it but....

"...
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)' ..."

The AMD Radeon VII Review: An Unexpected Shot At The High-End

www.anandtech.com

Vega II is probably pushed into at least as many "creators and early adopters" contexts as the Vega VII was. Unless surfaces as defect (thermals) not sure why Apple would have them crank it down further.

 

[vel0city]

Nobody running a 12 year old Mac Pro is happy about that fact. Every single one of them wishes Apple hadn't fumbled the ball so badly on their desktop lineup. Every single one of them wishes that Apple had produced a viable upgrade at some point during the past decade.

Yeah, so much this. I know some people in here are disagreeing with you, but for me Apple's complete disregard and dismissal of their professional creative userbase has been the cause of much stress and frustration. They abandoned their original core users and made doing our work extremely difficult. I hate running a 2009 Mac in 2020 and all the compromises I have to put up with. It used to be so easy to buy a new Mac for work when you needed one, now it's a mess of compromises, halfway there specs that don't quite cut it, and a lot of missed opportunities in terms of what you can actually get done on a Mac.

I just don't get why there isn't an absolutely kickass Mac Pro that's really efficient for Adobe CC and the kind of 3D/motion graphics work (C4D/After Effects) that is common now. Mac Mini/iMac/iMac Pro/Mac Pro are all hobbled or ineffective in some way for the most popular creative software on the planet. It's crazy and stresses me out just thinking about it.

 

[mikas]

I know this is a dead horse to beat, and for quite a long time now.

Need a workstation for architectural practice, for design and documentation, and including presentations and everything there is to it.

I'd have a hard time persuading my business partners to purchase for me one of those new Mac Pro's. I have played that PC versus Mac game online many times, just like many of us already have done. For us a workstation could cost something between 5000 to 7000, we think. With 7000€ with finnish 24% VAT included, I can for example choose either of these (see below):

A PC build:
AMD Ryzen Threadripper 3970X 24 32c/64t
256GB RAM
1 TB NVMe
RTX 2080 Ti 11GB
Win 10 Pro
+ motherboard, powersupply, cooler, mouse+keyboard and such.
total: 7034,20€ (incl vat(

An Apple Mac Pro:
Intel Xeon W 8c/16t
32GB RAM
1 TB Flash
AMD Radeon 580X 8GB
Mac OS Catalina
+ mouse and keyboard
total: 7099€ (incl vat)

The difference is huge if you consider performance.

Personally I like MacOS better. Mojave behaves alright today. Catalina is not near stable enough for our apps. So it's not even advisable to buy a Mac Pro right now because you can't choose OS version (this is the situation with OS X almost every year though).

Ok, there is not so much to upgrade with that AMD build. Memory is the max. CPU is the second best there is. You CAN upgrade the GPU, power supply and SSD easily and at good price point.

Mac Pro can be upgraded too of course. Choices are very limited with GPU. RAM upgrade is ok and the max ram is huge enough even with those lower xeons. CPU upgrades could be really pricey later on, because this xeon W series is outdated already, and a new socket is coming from intel again.

I believe I have no choice here. An iMac is not quite enough, and iMac Pro is a completely sealed box I hate, no user RAM upgrades, jeez. I don't warm to eGPUs either, though it is an option. A little bit clunky and pricier tha PCIe GPU.

Attached example builds from a local shop and from Apple.

 

[AidenShaw]

And the goalposts may be moving - tomorrow is NVIDIA CEO Jensen Huang’s keynote at GTC Digital - so the new standard could be the RTX 3080 Ti and any related Quadro cards.

Ampere A100

FP64 up from 7.46 to 9.7 Tflops, and 19.5 Tflops of TensorCore FP64 .

 

[deconstruct60]

Ampere A100

FP64 up from 7.46 to 9.7 Tflops, and 19.5 Tflops of TensorCore FP64 .

Pragmatically for now, at the low , low price of $119K. Different dimension from the Mac Pro. In fact, big chunk of the presentation is really aimed at all the Xeon SP and AMD EPYC can throw away once move your Spark load off of those and onto the A100. ( Yeah there is an AMD EPYC Rome (or two) in the DGX box, but major pointing at lowering the need for more in other rack mounted systems. )

No generic PCI-e card in sight. ( eventually but not coming soon. With or without macOS drivers this instantiation has no impact on Mac Pro line up. ). Once not using about the maximum die size that can be fabricated, that 9.7 will likely shrink back much closer to the 7.56. ( # of SMs increased 35% .... DP increased 30%. There wasn't a real 'breakthrough' there. Just more cores that scaled at a slope less than one. ) .
The Tensor variant of FP64 isn't portable.


The die size is unwieldy...

"...Something else to note: The full A100 chip has up to 128 SMs and 8192 FP32 CUDA cores, but only 108 SMs are enabled in the initial version. That says a lot about yields on such a massive chip, as well as TSMC's N7P node. Nvidia is probably saving up 'good' die for future products, but the initial A100 will only have about 85% of the SMs enabled (compared to 95% of SMs being enabled on the V100). ..."

Nvidia Unveils Its Next-Generation 7nm Ampere A100 GPU for Data Centers, and It's Absolutely Massive

7nm and 54 billion transistors

www.tomshardware.com

that is likely a bit generous. Probably, they have been saving up "good" dies just to get to 108 SMs in reasonable numbers. Whether AMD does with CDNA this year , Nvidia will probably counter with the "ludicrous speed" version to step on it if they can.

We'll see AMD's Arcturus may also bee a 600-700mm^2 die and have similar issues. But it will be able to layer on "more cores" in the same way to ramp TFLOP peak counts.

 

[AidenShaw]

No generic PCI-e card in sight.

Just like when the V100 came out. The A100 is a data center/HPC product.

Down the road a bit will be a graphics version with less FP64 performance but it will have RTX cores for ray tracing (no RTX on the A100).

 

[Mago]

TechPowerUp is correct, then in the space of a Vega II Duo with aggregate 1.76 Tflops FP64 you could put a pair of Titan V's with the same amount of VRAM and 14.90 Tflops.

Other sources like Anandtech suggest that AMD is limited the GPUs to 1:4 or 1:8.

No, adding to Dec's reply, the Vega II duo it's an ecc-less mi60 core on custom board, AMD actually puts the same GPU core across mi50/60/Vega II/R7 only with firmware changes (disabling fp64 EU as required) maybe to increase yield or as market segregation tactic, fact is on Linux Vega II are seen as mi60 , I cannot quote a web reference as the numbers comes from a chat with someone who benched The mp7,1 with opencl/ROCm.

Given the latest surprises from AMD and nVidia, all this topic is moot.

tomorrow is NVIDIA CEO Jensen Huang’s keynote at GTC Digital - so the new standard could be the RTX 3080 Ti and any related Quadro cards.

You're right, AMD has no chance against nVidia this year. Further Huang said something sweet: compute and consumer lines to share architecture, while unlikely we see the A100 in q Quadro or a Titan, it's very likely we see products like was the Titan Volta with generous FP64/TF32 (BF16?) Performance.

 

[Mago]

AMD does with CDNA

AMD doesn't learned anything from nVidia, despite progress with cdna/rdna, you can't code ROCm or HYP targeting navi rdna, as best you can target Vega and pray for it to run stable on Navi, no official report for ROCm/hyp on rdna card's seems AMD will provide it with cdna card's, this is unfortunate as you can't develop a compute product targeting rdna user's (ok with opencl under windows it's possible), this kick off small Linux users/developers, even frustrates market for those applications which could work fine on RX5400 on Nas or Soho servers.

Meanwhile intel seems well laser focused around oneAPI not closed neither restricted, maybe AMD will concede it's 2nd place in gpgpu market to Intel, it's ironic.

 

[deconstruct60]

AMD doesn't learned anything from nVidia, despite progress with cdna/rdna, you can't code ROCm or HYP targeting navi rdna, as best you can target Vega and pray for it to run stable on Navi, no official report for ROCm/hyp on rdna card's seems AMD will provide it with cdna card's, this is unfortunate as you can't develop a compute product targeting rdna user's (ok with opencl under windows it's possible), this kick off small Linux users/developers, even frustrates market for those applications which could work fine on RX5400 on Nas or Soho servers.

CDNA is likely an incremental refinement of Vega. So any Vega56 , 64 , VII , Pro VII card is a quite useable baseline for ROCm development until can get hands on this next iteration. All of those are better computational compute cards than a RX5400.

If pointing at NAS servers than the embedded GPU in the Ryzen 4000 APU series is the incrementally updated Vega also. So ... again highly aligned.

Is there going to be an affordable CDNA card? Probably not at the RX5400 level. Something at the "Vega 20" 60-64CU sized range.... probably but not the first iteration.

AMD has lots of stuff to optimize here and splitting focus at this point in time probably wouldn't help. They can merge back in at RDNA2 or maybe whatever RDNA3 is with the computational stack when the initial part is working and more mature.

Meanwhile intel seems well laser focused around oneAPI not closed neither restricted, maybe AMD will concede it's 2nd place in gpgpu market to Intel, it's ironic.

Depends. The Xe-LP is a laptop focused implementation so it probably won't win major prices on compute throughput to mid level desktop and up CDNA ( or even the older Vegas of that class). Xe-HP looks like its compute odd and somewhat narrow in being competitive. ( still largely is the server media stream GPU that Intel pitched. It probably isn't going to be a top end 3D compute card. ) .

There is lots of software that Intel has to get right for oneAPI to work. Intel has the bodies. Time will tell if they get the time and resources to get it right.

AMD also could tap in there if their OpenCL is up to snuff , but again unitl oneAPI has gotten to some stability , trying to completely cover oneAPI is probably secondary effort ( follow enough not to be too far of a gap but not pulling resources away from what AMD is already doing. )

 

[ssgbryan]

Vega isn't Navi. If you don't understand that part, why anyone read this post?

CDNA is a branch of RDNA - Vega is the last line of GCN architecture.

 

[ZombiePhysicist]

Vega isn't Navi. If you don't understand that part, why anyone read this post?

CDNA is a branch of RDNA - Vega is the last line of GCN architecture.

I honestly can’t nor do I care to keep up with all these silly code names and what follows what. Is there an easy to follow roadmap somewhere with cliff notes.

 

[ssgbryan]

I honestly can’t nor do I care to keep up with all these silly code names and what follows what. Is there an easy to follow roadmap somewhere with cliff notes.

That is the downside of being technically illiterate.

AMD, Nvidia, & Intel have all published road maps. Most have foot notes that explain the charts.

 

[ZombiePhysicist]

That is the downside of being technically illiterate.

AMD, Nvidia, & Intel have all published road maps. Most have foot notes that explain the charts.

right, knowing code names is key.

 

[ManuelGomes]

Possible Navi 21 Mac Pro cards spotted it seems (at least Apple branded), along with regular 21, 22 and 23. Salt required though

 

[Executor]

sorry, for my videowork my MPs are working very well and faster
then a imacpro does in my workflow - so i be really happy!

I am on the cusp of updating to the new Mac Pro 7,1. So I went to a colleague's place to check it out. And yes it was faster, but was it $20000 faster? for my application, the answer is no. Luckily my employer is gonna get me one anyway, but even then, I will always cherish my old 5,1 and will keep it for as long as it works.

 

[koyoot]

AMD doesn't learned anything from nVidia, despite progress with cdna/rdna, you can't code ROCm or HYP targeting navi rdna, as best you can target Vega and pray for it to run stable on Navi, no official report for ROCm/hyp on rdna card's seems AMD will provide it with cdna card's, this is unfortunate as you can't develop a compute product targeting rdna user's (ok with opencl under windows it's possible), this kick off small Linux users/developers, even frustrates market for those applications which could work fine on RX5400 on Nas or Soho servers.

Meanwhile intel seems well laser focused around oneAPI not closed neither restricted, maybe AMD will concede it's 2nd place in gpgpu market to Intel, it's ironic.

You can't code ROCm for Navi, because AMD has more compute focused GPU architecture which is CDNA/Vega.

Navi will always be targetting highest graphical output, over compute capabilities.

CDNA GPUs likely to have 128 CUs, 8192 ALUs, over 25TFLOPs of Compute FP32 performance, and 12.5 TFLOP FP64.

There is a reason why those GPUs are called Radeon MI100, and MI200. It will simply have 100 TFLOPs of INT8 performance, which translates to 25 TFLOPs FP32, which means: 128 CUs/8192 ALUs clocked at around 1.55 GHz.

MI200 likely to be dual GPU setup, with insane 50 TFLOPs of compute power FP32, and 25 TFLOPs FP64.

 

[Mago]

That's All, transition times, for me too, about a year now working at(remote) or from Linux, makes non sense for me at least in short time to adopt Mac except for building mac/ios Apps, not part of business line.

Actually I'm in love with CUDA ecosystem, and almost every tool I use long ago transitioned to Linux or has an solid sustitute, Macos will stay for long for soho/artistic users, but at least for a while HPC/STEM rules out Macs (unless you want to do everithing remote).

About remote I find its viable and practical to work with CUDA from a Mac by means Jupyter notebooks or Julia (remote juno), also some people reports success with VSCode/nSight, but the current stable Ubuntu Desktop (20.04) renders trivial any macOS excuse to work from a Mac, and actually if you do Tensorflow, DataScience you dont think macOS.

BTW, now iOS is more android-look-alike than ever, iPadOS is more iOS than ever, macOS is more iPadOS-like than ever, you see the trend?

see you in 2 years (if we are alive then)

 

[ManuelGomes]

Glad I didn't go with the MP2019 bandwagon. Will it be short lived?
OK, back to waiting for Late 2022 aMP (armMac Pro).
No small feat replacing Xeons. But I'm pretty sure in that undeground secret lab they must have something by now, maybe very early ES, in 2 1/2 years it will be awesome.
Intel must be pi$$ed as hell.
Back to RISC (pleaee no more x64 bull)

 

[filmak]

Glad I didn't go with the MP2019 bandwagon. Will it be short lived?
OK, back to waiting for Late 2022 aMP (armMac Pro).
No small feat replacing Xeons. But I'm pretty sure in that undeground secret lab they must have something by now, maybe very early ES, in 2 1/2 years it will be awesome.
Intel must be pi$$ed as hell.
Back to RISC (pleaee no more x64 bull)

Yes, but I don't know, for the high end it seems a bit difficult to switch right now to ARM cpus,
What kind of GPU they will have? right now they 're talking only about the integrated one.
How will high end apps run with only the integrated GPU?
What about PCIe expansion? how will this be implemented? what kind of slots? thunderbolt? chipsets? etc...

I think that (you can also say that I wish or hope), that intel will stay for some time inside the high end Macs.
Also losing the ability to run VMs with every possible OS and software (ok the majority of them) is very very disappointing, at least for me.

 

[ZombiePhysicist]

Glad I didn't go with the MP2019 bandwagon. Will it be short lived?
OK, back to waiting for Late 2022 aMP (armMac Pro).
No small feat replacing Xeons. But I'm pretty sure in that undeground secret lab they must have something by now, maybe very early ES, in 2 1/2 years it will be awesome.
Intel must be pi$$ed as hell.
Back to RISC (pleaee no more x64 bull)

I feel the opposite. I may have gotten the last intel Mac Pro. I doubt there are that many of them. A bit of a unicorn. PC video cards should continue to work well for a good long while. Who knows how that will work out on arm.

It's a relatively tiny market for intel, so while not thrilled, I doubt they care too much.

I do think the opportunity for many core ARM chips is great. 80 core arm chips are out already. Overall, I'm Psyched!

 

[ManuelGomes]

First ARM Macs (late this year) will use integrated graphics (A12Z), that's MacBook and maybe Mini?
The higher end machines will use discreet probably. Is Apple developing also a GPU? MAybe, based on Imagination IP? Or maybe stick with AMD.
Regular PCIe slots, TB should be no issue.

It might be a small count for Intel, the Xeons, but overall it's the whole lineup they're loosing.