2016 nMP

[Ph.D.]

Here's my not-very-well-informed take on the recent GPU news and how it relates to a nnMP:

First, Apple will not change the form factor, and hence will be working in the same power and space envelope. This isn't necessarily bad, as efficiency is quickly becoming a crucial factor in nearly every application of computing. Apple knows this better than most, and no-doubt the nMP was designed with this trend in mind.

Furthermore, as we've seen thus far, Apple will not make changes to the nMP until such time as those changes are substantial enough to be more than window-dressing. That is, at this stage in the game, I don't see Apple making a nMP with a revision using 28nm GPU chips, which due to the fixed thermal envelope would inevitably be a mediocre upgrade.

Next, my understanding is that HBM1 is limited to 4 GB per GPU in its normal configuration. This rules out use by Apple in their higher-end cards, since 16 GB would be a more likely target for a proper D700 replacement.

Proclamations by various suppliers indicate that HBM2 is coming this year. However, AMD has ruled out HBM2 until next year (Nvidia as well, I believe). This is probably due to a combination of the time needed to engineer compatible solutions, expectations of how quickly manufacturing can ramp-up HBM2 supply, and the expected cost of HBM2 in the intended higher-end configurations. That is, 16-32 GB per GPU of HBM2 may be too expensive for the broader market, including Apple, for some time.

So, with no HBM-2, this year is limited to a 4 GB HBM-1 iteration that won't suffice for Apple, or else a GDDR5 variant. Comments of uncertain reliability have been made elsewhere that Polaris GPUs and Nvidia's new GPUs may be made available in GDDR5 versions this year. Both companies are smart enough to avoid an outright dependency on HBM-2, so that makes sense. Thus, GDDR5/X or similar variant is probably the only hope for a timely 14/16 nm GPU update this year.

Where does this leave Apple? They are trapped waiting just as badly as we are, but the possibility of an update later this year is still alive. Barely.

(Personally, I don't see Apple going with Zen any time soon, if ever.)

 

[AidenShaw]

Mago, Vega will only be out next year.
I'm guessing another year without a new nMP, which is absurd, or availability with current tech (Broadwell and Polaris only).

Maybe Apple could adopt the "tick-tock" philosophy - and introduce an updated cMP with Broadwell and Maxwell at MacWorldSF in June.

In another three years, they could do a "tock" and upgrade the tube.

 

[Stacc]

Kaby Lake seems to be on track for H2.
Still no news from Broadwell-EP, strange. Intel should bring Skylake-EP already and skip it.
With a PCH like the Kaby Lake's 200 series would really make it awesome.

Your negativity seems misplaced. Broadwell is very much on track for release around June. Geekbench results are available for the 22 core version. It has started showing up in soon to be up and running supercomputers. Broadwell E (very similar to the smallest die size Xeon) is rumored to be released at Computex in early June.

In terms of GPUs I think Polaris 10 (note that Polaris 10 is bigger than Polaris 11) is the most likely chip to be in the mac pro. AMD has stated over and over that they are aiming for 2.5X performance per watt increase. If we dumbly apply this to the current lineup we get Fiji performance at 100 W. More or less a perfect card for the Mac Pro (remember, the current form factor is limited to about 125 W to 150 W). You may be bummed about the lack of a true big die GPU, but remember this is a new node (and a very challenging one).

So, with no HBM-2, this year is limited to a 4 GB HBM-1 iteration that won't suffice for Apple, or else a GDDR5 variant. Comments of uncertain reliability have been made elsewhere that Polaris GPUs and Nvidia's new GPUs may be made available in GDDR5 versions this year. Both companies are smart enough to avoid an outright dependency on HBM-2, so that makes sense. Thus, GDDR5/X or similar variant is probably the only hope for a timely 14/16 nm GPU update this year.

Polaris will most likely be using GDDR5X, which doubles the bandwidth of standard GDDR5 and will allow however much VRAM AMD/Apple wants on the GPU. Given the small die sizes and the focus on value this clearly points to using a GDDR5 variant instead of HBM. AMD can't really afford another GPU release that can't be used in workstations, as Fiji/Fury didn't have enough VRAM.

If Apple announced a mac pro in the summer here would be the selling points:

CPU: Increased the max number of cores from 12 cores to 22 with improved single threaded performance.
GPU: Polaris 10 and 11 GPUs. Increased GPU power from 7 TFLOPS to 16 TFLOPS. VRAM is anyones guess but rest assured it would be higher than 6 GB.
SSD: Probably 2-4x performance increase here (compared to the current mac pro)
Thunderbolt 2 -> Thunderbolt 3

All the components here will likely be available by mid to late summer. I hoping for a WWDC announcement.

At next WWDC Apple has to reveal its cards about VR, AR and some of their base technologies, like Metal v2. Current Metal is buggy and incomplete, it doesn't support even Shader Model 5.0 let alone asynchronous graphics/compute with multi-GPU environment - current version is just a developers playground. Also, DSP audio support is essential for VR. AMD mentioned TrueAudio Next in their yesterdays slides.

I wouldn't hold out hope for anything VR related from Apple in the near future. I am sure Apple has prototypes floating around their labs but they are rarely the first to introduce technology like this. They will likely see how VR plays out for a year and then try and figure out where they want to jump in. If they do, it will likely be on mobile.

That said, I do agree that they need to do a better job of supporting gaming on the mac. If they have any interest at all in it on iOS or OS X they need to get their gaming and graphics libraries up to date. AMD stated yesterday that their plan forward is going to be heavily reliant on multi-GPUs. Apple has been selling the dual GPU mac pro for 2 years now and they are ok at leveraging the compute performance but terrible with graphics.

AMD is moving away from Crossfire by the sounds of last night. MultiGPU support it is?

Crossfire is basically a driver implementation for dual GPUs. The driver tries to hide the fact that each GPU renders alternating frames. DirectX 12 (and LiquidVR) tries to get away from these driver "hacks" by having the game interact directly with the hardware. Thus the game can have much better control over how dual GPUs works. Instead of alternating frames to each GPU, they could split the frame between GPUs or do other cool tricks. In DX12, this has led to pairing Nvidia and AMD GPUs together and other neat tricks. Apple doesn't like maintaining drivers so this sounds like something they would do and it would likely be implemented as part of metal.

 

[Mago]

If Apple announced a mac pro in the summer here would be the selling points:

CPU: Increased the max number of cores from 12 cores to 22 with improved single threaded performance.
GPU: Polaris 10 and 11 GPUs. Increased GPU power from 7 TFLOPS to 16 TFLOPS. VRAM is anyones guess but rest assured it would be higher than 6 GB.
SSD: Probably 2-4x performance increase here (compared to the current mac pro)
Thunderbolt 2 -> Thunderbolt 3

All the components here will likely be available by mid to late summer

Mostly agree, except cpu core count I doubt it will exceed 16 notwithstanding there are 20 and 22 cores versions, due TDP restrictions,leaving 6/10/16 cores to choose.

And about GPU I bet on Apple will simplify to only two options a base model based on Fury nano (or actually a fury nano renamed) and a higher performer based on Polaris ( I'm confused I mean Polaris 10 is for mobile devices and Polaris 11 for desktops).

 

[goMac]

I really doubt Apple is going after VR in any big way. Metal will get upgrades, but it's not going to get all the features people want. My guess is it will get SM5, but no multi GPU for now.

 

[Stacc]

Mostly agree, except cpu core count I doubt it will exceed 16 notwithstanding there are 20 and 22 cores versions, due TDP restrictions,leaving 6/10/16 cores to choose.

Yeah, it wouldn't surprise me either if we didn't see the highest core count versions. Maybe just because Intel would charge a ridiculous amount.

And about GPU I bet on Apple will simplify to only two options a base model based on Fury nano (or actually a fury nano renamed) and a higher performer based on Polaris ( I'm confused I mean Polaris 10 is for mobile devices and Polaris 11 for desktops).

I don't know, Fury will be in a weird spot for the next generation of GPUs. Usually GPU cost is largely based on how large the die is. The nano is huge at 600 mm^2 and costs $500 retail right now. What do they do with this big expensive GPU if Polaris 10 matches its performance for similar or lower cost? Since Polaris 10 isn't going to be that big (rumored to be around ~230 mm^2) it probably won't be too much faster if at all compared to the Fury Nano. Its hard for me to even see any Fury card in the AMD 400 series lineup. The nano would also have to be downclocked so that two canfit in the mac pro (its currently a ~175-200 W card and needs to get to < 150 W).

From Apple's perspective it makes more sense to standardize all their upcoming machines on Polaris. In the Mac Pro have Polaris 11 as the low end option and Polaris 10 cut and Polaris 10 full as the higher end options. That way you get displayport 1.3, H265 encoding, and whatever other GCN optimizations are there. They can also put Polaris 11 in the macbook pro (unless they drop the discrete GPU altogether) and Polaris 10 in the iMac.

 

[goMac]

From Apple's perspective it makes more sense to standardize all their upcoming machines on Polaris. In the Mac Pro have Polaris 11 as the low end option and Polaris 10 cut and Polaris 10 full as the higher end options. That way you get displayport 1.3, H265 encoding, and whatever other GCN optimizations are there. They can also put Polaris 11 in the macbook pro (unless they drop the discrete GPU altogether) and Polaris 10 in the iMac.

Right. Fury Nano may be slower than mid end Polaris.

The high end is the problem right now. There doesn't seem to be a high end Polaris on deck in the next few months. Something like dual Fury/Radeon Pro will be faster than a single Polaris. But dual Polaris is a much tighter race.

I don't know if there is a great option for Apple until they can get their hands on high end Polaris. They could launch with dual high end Furys at the top, but it will only be a few months until they're outdated.

 

[dpny]

At next WWDC Apple has to reveal its cards about VR. . .

I don't think Apple has to address VR yet: it's still a very much overhyped technology, and anyone who buys a 1st gen VR headset is paying $800 to be a beta tester.

 

[Zarniwoop]

I don't think Apple has to address VR yet: it's still a very much overhyped technology, and anyone who buys a 1st gen VR headset is paying $800 to be a beta tester.

WWDC is a developer conference. Last year Apple started Metal for OS X hype, although the only ready product from day one that utilized Metal, was the OS X user interface. And still today, there's not a lot software supporting it. It just takes time. Soon they have to start pushing VR for developers in order to have software available when ever they bring their VR toys. MP will be the VR producing studio for Apple, not the current model, but v2 with duo Polaris 10.

If Final Cut or Logic won't get any support for VR, they will miss the train for good. I think this is the reason we have not seen updates for Pro software lately.. MP and Pro software have to start to support VR, and it needs Metal v2 and MP v2.

 

[Stacc]

Right. Fury Nano may be slower than mid end Polaris.

The high end is the problem right now. There doesn't seem to be a high end Polaris on deck in the next few months. Something like dual Fury/Radeon Pro will be faster than a single Polaris. But dual Polaris is a much tighter race.

I don't know if there is a great option for Apple until they can get their hands on high end Polaris. They could launch with dual high end Furys at the top, but it will only be a few months until they're outdated.

I think you are confused on your terminology here. Polaris 10 should be roughly as fast as Fury with lower power consumption. 2 Polaris 10 chips should more than double the GPU power compared to 2 D700s. This rumored Polaris 10 chip seems pretty high end to me. It probably won't be significantly faster than Fury but then again Fury is too hot to put in a mac pro.

There is always something faster being released in the future. I think you are referring to Vega, the AMD GPU to be released in 2017 with HBM2. Sure, Apple could wait around and use this chip but between broadwell, Polaris and Thunderbolt 3 they already would have a great upgrade. Who knows, maybe Apple would actually make an effort to keep the Mac Pro competitive and release a Broadwell/Polaris Mac Pro this year and then a Skylake/Vega Mac Pro next year (we can dream...).

 

[tuxon86]

I think you are confused on your terminology here. Polaris 10 should be roughly as fast as Fury with lower power consumption. 2 Polaris 10 chips should more than double the GPU power compared to 2 D700s. This rumored Polaris 10 chip seems pretty high end to me. It probably won't be significantly faster than Fury but then again Fury is too hot to put in a mac pro.

There is always something faster being released in the future. I think you are referring to Vega, the AMD GPU to be released in 2017 with HBM2. Sure, Apple could wait around and use this chip but between broadwell, Polaris and Thunderbolt 3 they already would have a great upgrade. Who knows, maybe Apple would actually make an effort to keep the Mac Pro competitive and release a Broadwell/Polaris Mac Pro this year and then a Skylake/Vega Mac Pro next year (we can dream...).

My understanding is that HBM2 will come out late on AMD because of Hynix. NVidia is using Samsung HBM2 and Pascal equiped with HBM2 should launch on time. This doesn't change anything for the nnMP since Apple seems commited to AMD at the moment.

 

[jonisign]

Yeah, it wouldn't surprise me either if we didn't see the highest core count versions. Maybe just because Intel would charge a ridiculous amount.

Mostly agree, except cpu core count I doubt it will exceed 16 notwithstanding there are 20 and 22 cores versions, due TDP restrictions,leaving 6/10/16 cores to choose.

The high core count chips have nearly equivalent TDPs for several models to Apple's current Ivy Bridge-EP offerings. For example, the Xeon E5-2698 v4 is a 20 core chip with a TDP of 135W. Compare that to the current top of the line Xeon E5-2697 v2, 12 cores and a TDP of 130W.

If Apple doesn't include the high end 20+ core chips it will be because of expense, rather than TDP restrictions.

 

[Mago]

I don't think Apple will get in VR market with am Mac based solution, they ever ignored gaming on Mac and Metal on OSX is far behind Metal on iOS.
Further iOS as gaming and entertainment platform has important presence and compared with OSX (dismal) traditionally ignored it's an clear sign on where Apple VR ambitions should go.

The Mac could stay as VR development platform but not beyond debugging as on iOS development.

Further Apple chip Ax (specially the one on iPad pro) has more than enough power to deliver an rich experience (compared to samsung gear VR) also to rival some Pc based VR proposal.

 

[koyoot]

It looks like, Zen CPUs have roughly the same power consumption on 14nm LPP as Intel CPUs on 14 nm with similar core configuration. But AMD CPUs might have lower core clocks. We are talking about 16 core version.

Also server parts are quite different than consumer parts, with twice the amount of cores, and... cache.

 

[Zarniwoop]

I don't think Apple will get in VR market with am Mac based solution, they ever ignored gaming on Mac and Metal on OSX is far behind Metal on iOS.
Further iOS as gaming and entertainment platform has important presence and compared with OSX (dismal) traditionally ignored it's an clear sign on where Apple VR ambitions should go.

The Mac could stay as VR development platform but not beyond debugging as on iOS development.

Further Apple chip Ax (specially the one on iPad pro) has more than enough power to deliver an rich experience (compared to samsung gear VR) also to rival some Pc based VR proposal.

iOS devices might be for the VR consumption, but Mac (pro) is the content creating tool. Both will share the VR business.. Games for iOS are made with Macs. VR for iOS certainly as well.

And who knows what kind of ideas Apple has for AR.

 

[koyoot]

And one another thing. GloFo ditched Samsung LPP process completely, and are using IBM SOI 14 nm process.

http://forums.anandtech.com/showpost.php?p=38100014&postcount=334
http://semiaccurate.com/forums/showpost.php?p=257667&postcount=122
And now we have this slide from VeriSilicon.
http://imgur.com/68y9WUg
What makes me apprehensive to this info is the leakage. Nothing, apart from Silicon-on-Insulator Process would allow that low leakage. So this is not LPP, but SOI process, IMO.

Also that tells why it would have 0.29 die are conversion compared to 28 nm GloFo. I will not be surprised if that P10 GPU is Fury X on 14 nm with 100W TDP and HBM1, without any of bottlenecks that were on Fiji.

 

[Zarniwoop]

Next summer is going to be pretty interesting in this regard.. Polaris 10 and 11 are like built for Apple.

it would be very odd if we don't see any new or updated Pro Mac's at WWDC.

- Macbook Pro with Skylake Xeon and P11
- Mac Pro with Broadwell Xeon and P10

 

[Ph.D.]

With 2 GPUs, the nMP could make a good VR development machine, wouldn't it? I seems natural that many VR systems would (eventually?) employ dual GPUs, with one per eye.

Also:

What makes me apprehensive to this info is the leakage. Nothing, apart from Silicon-on-Insulator Process would allow that low leakage.

SOI helps, but I don't think that the choice of SOI or not is the most relevant factor in leakage in this case. Most leakage in deeply scaled transistors would be tunneling through the gate oxide, not substrate related. That's what drove finfet, after all. Wouldn't you agree?

 

[Bubba Satori]

Next summer is going to be pretty interesting in this regard.. Polaris 10 and 11 are like built for Apple.

it would be very odd if we don't see any new or updated Pro Mac's at WWDC.

- Macbook Pro with Skylake Xeon and P11
- Mac Pro with Broadwell Xeon and P10

Lot's of great CPUs and GPUs coming out in the next year or so, but...
If there is no word from Apple about the Mac Pro at WWDC I give it a 50/50 chance the MP is EOL.
No word by August I'd say the chance the MP is EOL at 90%.
Treating their pro/prosumer customers like mushrooms* will have it's desired effect
and Apple will trot out the infamous truck in a prius world analogy and wave goodbye to Mac Pros.

I hope I'm wrong.

*Kept in the dark and fed BS.

 

[koyoot]

It looks like 14 nm SOI is completely built for Apple. The reality turns out to be far more interesting than it was looking to this day...

SOI helps, but I don't think that the choice of SOI or not is the most relevant factor in leakage in this case. Most leakage in deeply scaled transistors would be tunneling through the gate oxide, not substrate related. That's what drove finfet, after all. Wouldn't you agree?

Yes, I would agree. The thing is not about choice here, and relevance. I was talking about validity of the slide . Im having hard time believing it 100% .


P.S. I know that most of you do not have account there, but lets take hints of big picture and look at few articles:
http://semiaccurate.com/2013/03/25/what-is-apple-doing-at-14nm/
http://semiaccurate.com/2013/07/12/apple-has-their-own-fab/
http://semiaccurate.com/2013/01/02/apples-silicon-design-capabilities-increase/
http://semiaccurate.com/2013/07/12/apple-has-their-own-fab/
http://semiaccurate.com/2013/12/17/apple-samsung-intel-foundry-plans/
http://semiaccurate.com/2013/08/26/a-third-player-emerges-apples-foundry-plans/
http://semiaccurate.com/2013/12/30/nvidia-lose-apple-macpro-line/
http://semiaccurate.com/2014/03/25/details-apples-gpu-emerge/
Now. GloFo bought Fab and patents from IBM. GloFo synchs their own process with Samsung. GloFo throws away Samsungs technology, when they got enough experience and knowledge, and finishes work on 14 nm SOI process.
About technology, Apple works on their own GPU. Raja Koduri, chief engineer in AMD in 2013 goes from Apple to AMD. Jim Keller was at Apple, and went back to AMD, to create Zen. All of the GPUs and CPUs will be produced on GloFo 14 nm process.

And we now are in Low-Level API world, world of VR, 4K and higher, etc. Upcoming months will be at least very interesting.

 

[tuxon86]

I don't think Apple will get in VR market with am Mac based solution, they ever ignored gaming on Mac and Metal on OSX is far behind Metal on iOS.
Further iOS as gaming and entertainment platform has important presence and compared with OSX (dismal) traditionally ignored it's an clear sign on where Apple VR ambitions should go.

The Mac could stay as VR development platform but not beyond debugging as on iOS development.

Further Apple chip Ax (specially the one on iPad pro) has more than enough power to deliver an rich experience (compared to samsung gear VR) also to rival some Pc based VR proposal.

But VR isn't just about gaming. It is also about media, science, engineering, medical research...
[doublepost=1458139961][/doublepost]

With 2 GPUs, the nMP could make a good VR development machine, wouldn't it? I seems natural that many VR systems would (eventually?) employ dual GPUs, with one per eye.

Also:



SOI helps, but I don't think that the choice of SOI or not is the most relevant factor in leakage in this case. Most leakage in deeply scaled transistors would be tunneling through the gate oxide, not substrate related. That's what drove finfet, after all. Wouldn't you agree?

Presently the second GPU is only used for gpgpu task and not for display. It doesn't look as if the next version will support dual gpu for display either.

 

[Mago]

But VR isn't just about gaming. It is also about media, science, engineering, medical research...

Beyond entertainment, other VR applications are very niche markets usually high vertical I doubt Apple is interested on these markets and will put its strength on a Mac based VR, o don't deny Apple will show some support at least for a device like HTC Vibe but don't hope it to be mainstream on Apple, as I said Apple VR will be focused arround some portable device based on iOS either having a Google cardboard like setup or with an dedicated headset with dual screens one for each eye, also a flat 3d screen mounted in something like Google cardboard w/o any lenses.

 

[koyoot]

Beyond entertainment, other VR applications are very niche markets usually high vertical I doubt Apple is interested on these markets and will put its strength on a Mac based VR, o don't deny Apple will show some support at least for a device like HTC Vibe but don't hope it to be mainstream on Apple, as I said Apple VR will be focused arround some portable device based on iOS either having a Google cardboard like setup or with an dedicated headset with dual screens one for each eye, also a flat 3d screen mounted in something like Google cardboard w/o any lenses.

VR in fact is the next "input device". It is much, much more than anyone can imagine right now. It IS everything.

 

[Zarniwoop]

But VR isn't just about gaming. It is also about media, science, engineering, medical research...
[doublepost=1458139961][/doublepost]

Presently the second GPU is only used for gpgpu task and not for display. It doesn't look as if the next version will support dual gpu for display either.

The current situation is just a result of that Apple didn't want to invest on combining openGL and openCL because they already had Metal coming up. That's why they stopped at openGL 4.1 and openCL 1.2. After when Metal is complete, they can offer GL and CL as a layer on top of Metal. The end of separate drivers from Intel, AMD and nVidia for oGL and oCL. Or they can sack both to the land of legacy. M$ could bring DirectX for Metal if they saw any benefit from it - just to annoy Apple could be the reason.

For 4k video editing, the current MP is good. It is still one of the best 4k tools out there and separate GL and CL cards works fine. But, Metal will combine both GL and CL worlds in to one or more asynchronous pipelines where both data flows to any GPU there is in the system. This feature has to be part of the Metal before Apple can do any serious VR, so developers can choose how many percent of GPU time they want to give for compute and graphics.

 

[Ph.D.]

Presently the second GPU is only used for gpgpu task and not for display.

Yes, but that's probably not locked into the hardware, and a nnMP or more likely an OS update, or both, could change that. Intuitively, it does seem attractive to have one GPU per eye. Rendering alternate frames in time simultaneously on 2 GPUs is an ugly process in comparison to rendering the same frames in time simultaneously from two different perspectives. While not strictly necessary, I believe a lot of demonstration hardware already uses 2 GPUs for what I'd imagine is this exact reason.