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

[GrumpyCoder]

Well I assume the high-end GPUs (MI50/60?) are faster, but not aimed at the gaming market. That's what they focused a lot on today. So R7 is probably the best gaming choice they have, isn't it? And for gaming, I expected a little more. Not that I really need it, but it would have been nice.

If the R7 is somewhat silent I'd even be tempted to try one as a eGPU for a MacMini.

The 16GB is interesting for DL and the price is just right. Depending on the price of the MI50/60 one might be better off to do prototyping on a cheap card and then rent a MI60 (or multiple) in the cloud. They just have to get their heads out of their butts and deliver more on the software side of things. Nvidia's Drive platform is brilliant... where is the AMD version? I can't stress this enough, but they have to put up major resources to compete with CUDA.

As for CPUs, I've read some rumors 3600X (8c/16t) @4GHz base for ~$230, 3700X (12c/24t) @4.2GHz base for ~$330 and 3850X (16c/32t) @4.3GHz base for ~$500, all for AM4. Would be nice if true.

 

[deconstruct60]

The new GPU is a mid end card, it's not their high end. I.E. it's not the same thing as full Vega 20.

No this is AMD's new "top end" ( probably for large majority of 2019 ... in 2020 CES probably a new story. )


So a 25% gain on the new high end vs their old high end is pretty good.

The downside is that the power consumption is still pretty high, which may be why they're releasing this before a full Vega 20. I'd hope they could drive it down to 225w for an upper mid end card.

Given all of the 27% improvements they are talking about and only 25% improvement in power ( same power that much better improvement) the thermals of these are likely 300W. Hence I don't see Apple jumping up and down for these at "uber clock" settings.

This is AMD new high end system though.

1. It is priced above the Vega 64 ( so it is probably not being dropped. )
2. The thermals are probably higher than the Vega 64 ( so again the 64 is probably not being dropped )
3. No PCI-e v4 or Infinitiy Link So

The number of cores is "down" but as long as they throw power at it ( over clocking same old basic design) they get more performance.


"... Diving into the numbers a bit more, if you took AMD’s second-tier Radeon Instinct MI50 and made a consumer version of the card, the Radeon VII is almost exactly what it would look like. It has the same 60 CU configuration paired with 16GB of HBM2 memory. However the Radeon VII’s boost clock is a bit higher – 1800MHz versus 1746MHz – so AMD is getting the most out of those 60 CUs. ..."
... AMD will be employing some mild product segmentation here to avoid having the Radeon VII cannibalize the MI50 – the Radeon VII does not get PCIe 4.0 support, nor does it get Infinity Link support – ...
https://www.anandtech.com/show/13832/amd-radeon-vii-high-end-7nm-february-7th-for-699


About the only thing they did for graphics is crank up the raster processors ( 128 versus 64 ) and increased the bandwidth. ( more HBMv2 stacks ). Probably the ROPs are hooked to the same block covering the HBMv4 channels so two more channels gets another bank of 64 ROPs. ( I haven't checked the old arch reviews but wouldn't be surprised).

Not having this though shouldn't have been enough to be "show stopper' for a Mac Pro though. This in only going to be a high end option card. ( the $699 will probably get a 35-45% Apple+AMD tax to take it Pro and end up around the $949-1049 price zone which will probably be around a $699+ BTO option. )

Apple would still need a Vega 54 and Polaris to fill out the entry-midrange options for the Mac Pro.
[doublepost=1547062071][/doublepost]

Well I assume the high-end GPUs (MI50/60?) are faster, but not aimed at the gaming market.

Faster at long computational focused workloads perhaps. Faster at graphics tasks probably not. It is the exact same graphics pipeline just clocked faster. ( Perhaps a tad more adept at active sync high refresh rates) but the bsical graphics features pipeline is exactly the same.

The major changes that the MI50 and MI60 is to non graphics pipeline changes in other computation areas.

The MI50/MI60 can also probably work on "bigger" problems that span the VRAM workspace. That will be faster.

This Vega VII is probably useful to the MI50/MI60 in that it is probably a better card if shooting for developer workstation cards buidling code for MI50/MI60 grids. Working on code fragments and libraries with a far more affordable card so can get it into the hands of more users.

 

[koyoot]

Well I assume the high-end GPUs (MI50/60?) are faster, but not aimed at the gaming market. That's what they focused a lot on today. So R7 is probably the best gaming choice they have, isn't it? And for gaming, I expected a little more. Not that I really need it, but it would have been nice.

If the R7 is somewhat silent I'd even be tempted to try one as a eGPU for a MacMini.

The 16GB is interesting for DL and the price is just right. Depending on the price of the MI50/60 one might be better off to do prototyping on a cheap card and then rent a MI60 (or multiple) in the cloud. They just have to get their heads out of their butts and deliver more on the software side of things. Nvidia's Drive platform is brilliant... where is the AMD version? I can't stress this enough, but they have to put up major resources to compete with CUDA.

As for CPUs, I've read some rumors 3600X (8c/16t) @4GHz base for ~$230, 3700X (12c/24t) @4.2GHz base for ~$330 and 3850X (16c/32t) @4.3GHz base for ~$500, all for AM4. Would be nice if true.

How are ALUs from the same GPU, placed for ML market faster than ALUs in the same GPU for gaming/mainstream market?

P.S. I presume you haven't heard anything about TensorFlow on ROCm platform, from AMD?

 

[ssgbryan]

Vega 7 IS Vega 20 based. It is the same GPU

P.S. There are eng samples with 12 cores floating around, so maybe they will pull something magical for AM4.
Mac Pro should be based on HEDT and Server chips, at the least. If we are talking about using AMD CPUs on Apple computers: Threadripper, and EPYC.

A 16-core, 32 thread Ryzen 9 would be a good entry level Mac Pro CPU.

 

[koyoot]

A 16-core, 32 thread Ryzen 9 would be a good entry level Mac Pro CPU.

Dual channel memory with just 32 PCIe 4.0 lanes?

 

[ssgbryan]

I can't stress this enough, but they have to put up major resources to compete with CUDA.

AMD has something better than CUDA.

AMD ProRender Engine - will see the CPU, GPU, Video Ram and System Ram as 1 system. Doesn't care if the GPU is Team Green or Team Red.

Some of the apps I use already have a plug-in available (Blender, Maya, and Cinema4D).

 

[koyoot]

Doesn't care if the GPU is Team Green or Team Red...

...or Team Blue!

 

[GrumpyCoder]

How are ALUs from the same GPU, placed for ML market faster than ALUs in the same GPU for gaming/mainstream market?

Depends on the driver and library implementation and the level of optimisation you can expect from it. For the hardware itself, it doesn't matter. Unless of course you have different type of hardware that benefits different types of applications, see Turing.

P.S. I presume you haven't heard anything about TensorFlow on ROCm platform, from AMD?

It's not there yet, by far. Some stuff works, other stuff doesn't. The majority of stuff available for research requires significant amount of work to port it, despite the fact that it should be easy. The whole scientific world is CUDA based and with CuDNN very fast due to Nvidias excellent optimisation. Too much work. That's why AMD has to put more effort into it, hardware, software, support, etc. Mentioned it before, where is that Drive platform from AMD? They showed a little genetic application this time, but look at what Nvidia does. AMD has to focus on this.

This sums it up nicely: https://www.servethehome.com/tensorflow-1-8-with-amd-rocm-support/
More resources please AMD.

AMD has something better than CUDA.

AMD ProRender Engine - will see the CPU, GPU, Video Ram and System Ram as 1 system. Doesn't care if the GPU is Team Green or Team Red.

And that's going to help with ML and games how exactly?

 

[ekwipt]

Nope . It was 8C/16T eng sample .

Im pretty sure, that Engineering Sample was 65W, eng sample Apisak found with 3.7 base/4.0 GHz boost clock, few weeks ago.

This is the reason why AMD claims "minimum" performance of Core i9 9900K.

This has made my mind. My next build is based on Zen 2+Navi.

I think Apple will go 100% with AMD for the next Mac Pro, they just can't tease anything at this point or the AMD share price would go through the roof

 

[koyoot]

This sums it up nicely: https://www.servethehome.com/tensorflow-1-8-with-amd-rocm-support/
More resources please AMD.

Uhhh... TensorFlow started being supported since ROCm 2.0 version, not before...

I think Apple will go 100% with AMD for the next Mac Pro, they just can't tease anything at this point or the AMD share price would go through the roof

They have a choice either going with AMD x86-64, or their own ARM - CPU route, and so I think the painting on the wall is clear.

They will go their own CPU route!

 

[deconstruct60]

Dual channel memory with just 32 PCIe 4.0 lanes?

PCI-e v4.0 is a dual edge sword. It isn't the "generic PCs" solution some folks try to trot it out to be.

"... One of the differences with PCIe 4.0 is that it can only handle PCB traces up to 7 inches before needing a redriver/retimer, so these extra ICs are needed for ports lower down the board. But, the first PCIe slot on most motherboards is in that limit ... motherboards, assuming the traces adhere to signal integrity specifications ..."
https://www.anandtech.com/show/13829/amd-ryzen-3rd-generation-zen-2-pcie-4-eight-core

It given the heat and infrastructure around mode CPU sockets that's going to pragmatically going to be about one slot away.

In desktops more like one "first slot" distance and then into a PCI-e switch ( e.g. x16 PCI-e v4 into two x16 PCI-e v3 ) or something like that on many boards. And to just a singular x16 PCI-e v4 on others boards as the GPUs roll out later 2019-2020 time frame.

Wouldn't count on many "race to affordable" boards to be adhering to signal specs 1-2 genations in advance of what they were deployed on. [ Nor would I expect Apple's to either. Doubt they are "betting the farm" on PCI-e v4 in 2019 and most of 2020. ]

I don't think PCI-e v5 gets back any of this loss in distance. It is more bandwidth, but still trading off distance.


In the keynote there was a slide that said " Threadripper is the #1 best selling workstation processor " ...

https://www.anandtech.com/show/1382...live-blog-looking-ahead-starts-9am-pt-5pm-utc

Not sure how they counted that ( vs Xeon W only or what ), but if true it might bring at the very least some relief on the next iteration of Xeon W price point. Not total collapse but closer to matching the high end Core i9's which basically use the exact same die. Xeon W drifted off into the non competitive weeds a bit based on Intel not acknowledging they had a problem.

 

[ssgbryan]

And that's going to help with ML and games how exactly?

I don't care about games - I create.

 

[Boil]

Man, a Radeon VII with a 16 core / 32 thread Ryzen 9 3850X CPU is going to make one heck of a personal workstation...!!!

Now if they would just up the RAM amount for ITX to 64GB...

 

[goMac]

They have a choice either going with AMD x86-64, or their own ARM - CPU route, and so I think the painting on the wall is clear.

They will go their own CPU route!

I don’t think it’s cost effective for them to do dedicated, very very short production runs of their own workstation chips. Most CPU lines make it up in volume, and they can’t make it up in volume with the iMac Pro/Mac Pro alone.

For the Macs they sell a whole ton of, like the smaller laptops, it makes a lot of sense. Especially in since those laptops aren’t far off the iPad.

 

[koyoot]

It will be funny when Apple will be criticised for using Intel, outdated CPUs, instead of switching to AMD for Zen 2...

 

[askunk]

I own two companies (needless to say, less capitalized than Apple ), but if I had to judge with an entrepreneurial approach, I wouldn't change supplier for two reasons:

1) I guess Intel provides small quantities of Xeons at an affordable price (don't forget how the iMP costs much less than every single piece put together), therefore I think you should see the bigger picture: the whole contract may be more convenient than AMD's, plus providing EEC capable chips for the high end.

2) Even more important: if I were thinking about moving to proprietary CPUs and APUs, I wouldn't negotiate a new contract just for one or two generations of machines. The internal overhead costs in terms of resetting teams and establish a new workflow wouldn't pay back in time, imho.

Let's be clear. I would love to see AMD's chips in Macs

 

[ssgbryan]

Xeon = Higher Cost, Lower Performance.

Intel's only advantage is single core performance (and that should disappear with the release of the next gen Ryzen CPUs). But it isn't 2010 anymore; most programs are multi-threaded.

AMD's CPUs would be perfect for Apple - throw the Ryzen 3000 with APUs into the Mac Mini, Ryzen 5 and 7 series in the Imac, and Ryzen 9 in the Mac Pro.

 

[Mago]

Let's be clear. I would love to see AMD's chips in Macs

I mostly agree with you, I'd like to see the new ThreadRipper 2 or Epyc Rome in the new MP, mostly due superior corecount and FP performance and efficient 7nm process, but I dont see a business sense (unless intel sells to Apple its CPU at AMD prices -even losing-) to Apple to stay on intel on the iMac Pro and Mac Pro, Intel's dominance in server market imposes a huge markup in their silicone, while on desktop is where AMD punches hard surely Intel is more flexible to provide good pricing, in high end mobile is where AMD is even more weak, so unlikely to see AMD in laptops, and mainstream desktops (as long Intel is capable to provide competitive CPUs), but on iMP/MP duo its another history, is where Apple could save the huge server markup just by switching to AMD, also providing a more attractive desktop.

Amd Just launched its Vega 20 (Radeon VII / Instinct M60), hands down coming to the next MP/iMP, but why not to buy the whole package? a FULL AMD iMP/MP (with outstanding FP64 performance, 32x faster than latest nVidia/ Intel combo)? If Apple is doing business, it has no sense not to offer FULL AMD iMP/MP, going to Intel will be more expensive and inferior for FP64 performance -nVidia still FP32 King by long-).

But, this is a minor issue when you see the negative impact in developers is having Metal, Metal is great, but no one likes it, ask the community everybody wants is Vulkan, and OpenCL 2.x, kicking off from the Apple macOS/iOS ecosystem was a huge mistake, despite solutions to compile Vulkan/OpenCL into Metal, I considered it, but has little sense, since you re not away from new bugs/tuning.

Apple's big error it to believe they are Almighty and Irreplaceable, sometimes hurts to switch, but now I'm very Happy on Ubuntu Linux doing everything I used to do in a Mac, (except Autocad/OmniGraffle ) and all Apps/IDEs I needed have now native Linux/Ubuntu implementations (Pycharm, AndroidDev, CUDA, mROC OpenScad, LibreOffice etc), so i should care on the Mac Pro just to code for Metal ? NO

 

[askunk]

EPYC is very cool for HPC, but has very low base frequencies for ST ops. I still prefer a Xeon for my workload. I would compromise for a Threadripper, but I don't think Apple would for the mMP. It wouldn't make sense with respect to the iMP.

 

[Mago]

This Vega VII is probably useful to the MI50/MI60 in that it is probably a better card if shooting for developer workstation cards buidling code for MI50/MI60 grids. Working on code fragments and libraries with a far more affordable card so can get it into the hands of more users.

Preliminary info on the Vega VII suggest it to be the same MI60 with video output, no PCIe4 no Infinity Link, while seems VegaVII to have a custom xGMI bridge inferior to the one on MI60, or at least coming later as the MI60 market gets cold (driven by FP64 performance, a small niche)

 

[Mago]

EPYC is very cool for HPC, but has very low base frequencies for ST ops. I still prefer a Xeon for my workload. I would compromise for a Threadripper, but I don't think Apple would for the mMP. It wouldn't make sense with respect to the iMP.

ST performance maybe inferior with integer, AMD Epyc has far superior FP performance than Intel, it levels things.

I Believe the Mac Pro if goes full AMD, it should be based on ThreadRipper 2/3 having upto 32 cores is enough, offering a 64 core EPYC solution, has little sense in macOS ecosystem, even few apps will take full on 32c TR, meanwhile TR2 has fair ST performance, unbeatable pricing points, and native ECC(optional) support across all configurations .

Me as Mac Product Manager, will move iMac to Ryzen 5/7/9 + AMD RX6xx and iMP/mMP to TR2 12c-32c and single Radeon VII upto dual MI60 (plus nVidia BTO option), Sell a similar pricing point as former generation but with much higher performance.

 

[theroaringbadger]

ST performance maybe inferior with integer, AMD Epyc has far superior FP performance than Intel, it levels things.

I Believe the Mac Pro if goes full AMD, it should be based on ThreadRipper 2/3 having upto 32 cores is enough, offering a 64 core EPYC solution, has little sense in macOS ecosystem, even few apps will take full on 32c TR, meanwhile TR2 has fair ST performance, unbeatable pricing points, and native ECC(optional) support across all configurations .

Me as Mac Product Manager, will move iMac to Ryzen 5/7/9 + AMD RX6xx and iMP/mMP to TR2 12c-32c and single Radeon VII upto dual MI60 (plus nVidia BTO option), Sell a similar pricing point as former generation but with much higher performance.

A mac pro will never run AMD if there is no Thunderbolt 3 support.Too many pros depend on it.

 

[GrumpyCoder]

TB3 already works with AMD, it's probably just a licensing issue. I don't see AMD happening for the MacPro though. The future is ARM and it doesn't really make sense to switch to AMD and then to ARM within 5 years later.

 

[koyoot]

The future is ARM

For servers, low powered ones? Sure. For everything performance related - nah..

 

[askunk]

Low and mid range will use ARM APUs, high end will keep CISC CPUs and discrete GPUs for another while, I think. Servers and HPCs... I honestly do not know. ARMs can be tailored for vertical applications, which can be a big plus for HPC.

Since it may take Apple a bit before they will be able to cover the whole line with proprietary chips, I see very far the hypothesis of Apple transitioning to AMD chips.

Unless the close bond between Apple and AMD on GPUs helped to keep prices low and get a good contract, instead of having two upset suppliers for watching small orders quickly fading away. Maybe there is room for surprises...