2016 nMP

[Stacc]

As we where close to WWDC every day seems more likely at least for the Mac Pro a comeback to nVidia on Apple, I think also we could see Apple offering both AMD Polaris 10 Pro and XT at lowee end Mac Pro (as d300/d500 replacement) s and some nVidia Pascal (102/104) at Mac Pro higher end (as D700 replacement and option for those dependant on CUDA) .

Its certainly possible we could see Nvidia in the Mac Pro. I don't think we will see both AMD and Nvidia in the Mac Pro at the same time though. Apple is not going to go to the trouble of designing custom cards from two different manufacturers. A possibility is that they could bring support for GPUs over Thunderbolt 3.

It's curious that CUDA 8 which requires special kernel access having support from Apple in this requirement (to enable Pascal's unified memory) on OSX. I think this it's an cue on nVidia comeback to Apple, at least the Mac Pro.

Nvidia has shown they are very serious about having the fastest possible compute product. Apple could leverage both unified memory and NVLink to create a really unique and powerful product. NVLink could help offset some of the bandwidth constraints it is facing trying to get multiple GPUs, a PCIe SSD and multiple thunderbolt 3 controllers sharing the same 40 PCIe lanes. Unified memory would help if Apple wants to use GP104 and its limited to 8 GB of GDDR5(X) VRAM.

 

[ManuelGomes]

I don't see both AMD and NVidia cards on anything Apple also. Either one or the other, not both.
NVLink will not be coming anytime soon to the desktop parts, I would imagine. it will be HPC only, for a while at least I'd reckon. And I doubt it will be on the low and mid range cards as well, maybe on the higher end further in time.
It would be a distinctive factor for Apple to have a workstation with "different" hardware from everyone else, and stop the HP/Dell/Lenovo comparisons of late. And that would be a Power based CPU (I still believe they will be ditching ARM and go OpenPower) with NVidia GPUs all networked with NVLink.
But that would be a long time from now, and it would surely set them apart from all the other workstation vendors. Thing is, that tech is server territory only for now. Time will tell.

 

[Mago]

Its certainly possible we could see Nvidia in the Mac Pro. I don't think we will see both AMD and Nvidia in the Mac Pro at the same time though. Apple is not going to go to the trouble of designing custom cards from two different manufacturers. A possibility is that they could bring support for GPUs over Thunderbolt 3.



Nvidia has shown they are very serious about having the fastest possible compute product. Apple could leverage both unified memory and NVLink to create a really unique and powerful product. NVLink could help offset some of the bandwidth constraints it is facing trying to get multiple GPUs, a PCIe SSD and multiple thunderbolt 3 controllers sharing the same 40 PCIe lanes. Unified memory would help if Apple wants to use GP104 and its limited to 8 GB of GDDR5(X) VRAM.

Could nvLink run on the existing (and unused) lanes for cross fire x on the Mac Pro.

 

[koyoot]

Could nvLink run on the existing (and unused) lanes for cross fire x on the Mac Pro.

AMD develops their own Fabric that can be used with any x86 CPU.

 

[Mago]

AMD develops their own Fabric that can be used with any x86 CPU.

This is for Zen cpu on HPC (as infiniband or Intel's omni path) applications nothing similar to nvLink (which is more related to SLI on steroids) and it's restricted to one node.

I think SLI to be substituted eventually by nvLink at least on workstation or compute nodes.

 

[koyoot]

I just got used that my words even if turn out reality after few months, are generally ignored...

 

[Zarniwoop]

It's curious that CUDA 8 which requires special kernel access having support from Apple in this requirement (to enable Pascal's unified memory) on OSX. I think this it's an cue on nVidia comeback to Apple, at least the Mac Pro.

Nvidia's "unified memory" is just another term for HSA (part of it) - which (the term) they don't use openly, because it has a red tag on it. Even today OS X could run Nvidia and AMD card in the same machine without a hiccup. And that could be one of the futures of nMP. ("Always changing is the future") Nvidia for GPU, AMD for GPGPU - or vice versa. They could be even used together to render same screen as of today.

To put it short: OS X will not have problems supporting unified memory.
[doublepost=1460150042][/doublepost]I've mentioned it before, but AMD and Apple are most likely "Cooking" Apple specific APU. It could be a combination of Zen cores, ARM die, DSP/ISP and Apple GPU and AMD GPGPU. All integrated on the same chip. And it wont exclude dGPU from Nvidia used in same machine.

Or.. it could be IBM Power die with Apple ARM and AMD GPU/GPGPU integrated on same chip. But I think Apple will keep x86 compatibility for a while. (or not, because we have not seen Apple Pro apps been updated for a while too... just small tweaking)

Anyway, thanks to HSA (and its unified memory architecture) all this is possible. And with Metal and its next version, will improve.

 

[koyoot]

The APU is for AppleTV.

P.S. HSA on Nvidia hardware is done through CUDA.

 

[Zarniwoop]

The APU is for AppleTV.

Could be its first use. Next Apple TV could be also a combination of Mac Mini and Apple TV. Supporting 4k first and VR later.

Maybe it becomes the VR Hub. Then the glasses need only Apple Watch level HW to transfer the screen data into the VR glasses... keeping them extra light.
[doublepost=1460151471][/doublepost]

P.S. HSA on Nvidia hardware is done through CUDA.

Yes. And it needs to use open standards before it ever finds itself to any Mac.

 

[Mago]

I don't see both AMD and NVidia cards on anything Apple also. Either one or the other, not both.

This is not an technical or economical issue, are only busines linings which prevented having two GPU vendors on the Mac Pro (apple used to sell different GPU across the line, as for a while having either nVidia or AMD on the macbooks and the other on the iMac.

Neither will raise the cost having to "develop"diffrent GPU cards, actually customize a logic board (either a motherboard or a GPU card) isnt nothing expensive, neither cost millions, actually each different GPU designs cost few thousand dollars to commision to an stablished GPU vendors, another question is to mass produce that GPU card, since Apple use to order Millions units those production lines are more complicated to modify, from logistics to tooling, but what apple ussualy does is to go on a single vendor to save few bucs on eacch supply contract, modifing a production line while not easy its an usual thing, and common from switchong GPU platform or model, to just an simple logic board revision due issues with thermal compound. if apple finds an economical reason to sell different GPU vendors, will do, ehst you unlikely wiill se are tho similar spectd' gpu offered to the same machine.

NVLink will not be coming anytime soon to the desktop parts, I would imagine. it will be HPC only, for a while at least I'd reckon. And I doubt it will be on the low and mid range cards as well, maybe on the higher end further in time.

I thing nvLink will come to all the next gen nVidia GPUs from desktop to workstations. a misconception we had until yesterday is that nvLink is specific for Power based systems, no we could have nvLink interconnecting GPU on x86 too (as the new DGX-1 based on Xeon E5-V3, what restricts nvLink on x86 is PCIe3 lower bandwidth will downgrade (32 to 40gbps) system ram interconnection, as power handles more that 40 this degradation is raised.

note xeon and nvlink on the image (DGX-1)

It would be a distinctive factor for Apple to have a workstation with "different" hardware from everyone else, and stop the HP/Dell/Lenovo comparisons of late. And that would be a Power based CPU (I still believe they will be ditching ARM and go OpenPower) with NVidia GPUs all networked with NVLink.
But that would be a long time from now, and it would surely set them apart from all the other workstation vendors. Thing is, that tech is server territory only for now. Time will tell.

What would be distinctive is updating it ASAP new tech is available, not at the next WWDC or next "launch event", this is a vacuum salesman mind policy not real tech vendors.

 

[koyoot]

GP104: 290-300 mm2.
Full chip: 3072 CUDA cores
Cut down: 2560 CUDA cores. Both with 8 GB GDDR5.

I am quoting myself to correct this information. It looks like full chip of GP104 will have 2560 CUDA cores. 4 GPCx10x64 = 2560. Full GP100 - 6 GPCx10x64 = 3840

 

[ManuelGomes]

Mago, NVLink has always been shown with even x86, the difference was that x86 would be connected to the GPUs by a PCIe switch, not directly by NVLink. Slides showing those configs have long been on the net, not only during the presentation.
But Power 9 will take full advantage of NVLink, since it interfaces with it. I believe ARM was also being worked on but I believe NVidia will be making their own Power CPU now.
[doublepost=1460152741][/doublepost]I believe we'll see the 3 major players with each their own solution (CPU+GPU), NVidia wants to brake away from Intel and they just might with IBM.

 

[Zarniwoop]

I believe NVidia will be making their own Power CPU now.

That could be very possible. And would explain why Intel had to renew some patents with AMD instead of Nvidia (Intel used Nvidia patents before..)

 

[koyoot]

Check this out: http://www.freepatentsonline.com/20160085551.pdf

Remember this post?

http://forums.anandtech.com/showpost.php?p=38154078&postcount=13 Read this. A lot is falling in line perfectly, so far.

 

[Mago]

Mago, NVLink has always been shown with even x86, the difference was that x86 would be connected to the GPUs by a PCIe switch, not directly by NVLink. Slides showing those configs have long been on the net, not only during the presentation.
But Power 9 will take full advantage of NVLink, since it interfaces with it. I believe ARM was also being worked on but I believe NVidia will be making their own Power CPU now.

Not right, power 9 will have it own gpu on board connected to nvLink as well pcie4.

nvlink don't replace pcie it's just an shortcut or express line to interconnect compute accelerators among them and they later the systems ram.

 

[goMac]

It's curious that CUDA 8 which requires special kernel access having support from Apple in this requirement (to enable Pascal's unified memory) on OSX. I think this it's an cue on nVidia comeback to Apple, at least the Mac Pro.

CUDA has always had kernel access. No change here. It uses a kernel extension. Doesn't really require anything special besides Apple providing a signing key, which Nvidia has had for a while. (Same is true of the web drivers.) I don't think anything about unified memory really requires Apple to do anything for Nvidia.

And again, it's worth pointing out that Nvidia is enabling unified memory for all their cards in CUDA, it's not the Pascal specific stuff. Nvidia has not said that the Pascal stuff is coming, they only said the older unified memory stuff from CUDA 6 is finally hitting OS X.

 

[tomvos]

And again, it's worth pointing out that Nvidia is enabling unified memory for all their cards in CUDA, it's not the Pascal specific stuff. Nvidia has not said that the Pascal stuff is coming, they only said the older unified memory stuff from CUDA 6 is finally hitting OS X.

Quoted from the NVidia blog post:
> In addition to Pascal support in CUDA 8,
> CUDA 8 platform support for Unified Memory expands to Mac OS X.

I added the emphasis. I would expect that it's not just the CUDA 6 stuff but all the latest unified memory features from CUDA 8. At least all the latest features supported on the officially available hardware.

 

[ManuelGomes]

Mago, I believe you're reading me wrong.
I didn't say it replaces PCIe. It's a high speed pipeline between GPUs and some CPUs.
Power9 interfaces directly with NVLink, other CPUs will have to use a PCIe switch, that's what I said.
I'll try and find the pic.
There you go:

Also, most don't like this but as usual take it with a grain of salt:
http://wccftech.com/nvidia-pascal-gp104-gpu-leaked/
[doublepost=1460225865][/doublepost]

 

[AidenShaw]

There you go:

Good job, thank you for taking the time to find and post this. There's been a lot of speculation and misinformation about NVlink lately.

An earlier comment about it being "SLI on steroids" is pretty much on the mark for x64 processors. Extremely high bandwidth between the GPUs, and PCIe bandwidth to the system RAM.

For future CPUs with NVlink support, the bandwidth to system RAM can approach the intra-GPU bandwidth.

For now, the only roadmaps to NVlink CPUs are on the Power side - but I'm sure that Intel is watching and will jump in with NVlink support if the supercomputers abandon Xeon.

But nothing is more telling than the fact that Nvidia's DGX-1 "supercomputer in a small box" is a dual Xeon machine. Clearly an 8-way GPU complex with 128 GiB of VRAM on NVlink and 80 lanes of PCIe 3.0 doesn't suffer too much from using PCIe for transfers to/from system RAM for many significant applications.

 

[jerwin]

But nothing is more telling than the fact that Nvidia's DGX-1 "supercomputer in a small box" is a dual Xeon machine. Clearly an 8-way GPU complex with 128 GiB of VRAM on NVlink and 80 lanes of PCIe 3.0 doesn't suffer too much from using PCIe for transfers to/from system RAM for many significant applications.

From one of those nvidia whitepapers

NVLink is a key technology in Summit’s and Sierra’s server node architecture, enabling IBM POWER CPUs and NVIDIA GPUs to access each other’s memory fast and seamlessly. From a programmer’s perspective, NVLink erases the visible distinctions of data separately attached to the CPU and the GPU by “merging” the memory systems of the CPU and the GPU with a high-speed interconnect. Because both CPU and GPU have their own memory controllers, the underlying memory systems can be optimized differently (the GPU’s for bandwidth, the CPU’s for latency) while still presenting as a unified memory system to both processors.

So, if one needs throughput, but latency is not as much of concern, the DGX-1 might be a good fit... If you need both, prepare to shell out the big bucks.

 

[goMac]

Quoted from the NVidia blog post:
> In addition to Pascal support in CUDA 8,
> CUDA 8 platform support for Unified Memory expands to Mac OS X.

I added the emphasis. I would expect that it's not just the CUDA 6 stuff but all the latest unified memory features from CUDA 8. At least all the latest features supported on the officially available hardware.

And it's possible the CUDA 8 stuff is there, but back to my original point, it doesn't really point to Apple adopting Nvidia chips again. I read it as a "we finally caught up to Metal's unified memory" bullet point. The unified memory support that was present in CUDA 6 for Windows definitely wasn't in OS X until now, so my impression is they're trying to make up loses against Metal, not really point to new Mac hardware.

 

[Stacc]

Good job, thank you for taking the time to find and post this. There's been a lot of speculation and misinformation about NVlink lately.

An earlier comment about it being "SLI on steroids" is pretty much on the mark for x64 processors. Extremely high bandwidth between the GPUs, and PCIe bandwidth to the system RAM.

For future CPUs with NVlink support, the bandwidth to system RAM can approach the intra-GPU bandwidth.

For now, the only roadmaps to NVlink CPUs are on the Power side - but I'm sure that Intel is watching and will jump in with NVlink support if the supercomputers abandon Xeon.

But nothing is more telling than the fact that Nvidia's DGX-1 "supercomputer in a small box" is a dual Xeon machine. Clearly an 8-way GPU complex with 128 GiB of VRAM on NVlink and 80 lanes of PCIe 3.0 doesn't suffer too much from using PCIe for transfers to/from system RAM for many significant applications.

NVLink is really useful when you are trying to get > 2 GPUs to talk to each other. For 2 GPUs like in the Mac Pro it has more limited benefits but by putting both GPUs behind one 16x PCIe connection it opens up bandwidth for other things like thunderbolt 3 controllers and PCIe SSD(s).

If we see Pascal in the Mac Pro, it will likely be in the form of GP104 so those are the rumors to pay attention to. My bet would still be on AMD's Polaris though.

 

[AidenShaw]

NVLink is really useful when you are trying to get > 2 GPUs to talk to each other. For 2 GPUs like in the Mac Pro it has more limited benefits but by putting both GPUs behind one 16x PCIe connection it opens up bandwidth for other things like thunderbolt 3 controllers and PCIe SSD(s).

If we see Pascal in the Mac Pro, it will likely be in the form of GP104 so those are the rumors to pay attention to. My bet would still be on AMD's Polaris though.

NVlink has nothing to do with the PCIe lanes - you can "open up bandwidth" by using a PCIe switch without NVlink.

NVlink helps a bit in that GPU <-> GPU traffic doesn't have to go over PCIe, but it's not essential. If you have two or fewer x16 GPUs per processor, it doesn't really matter for many applications. For four GPUs per CPU like the DGX-1, it's a big deal.

 

[blindpcguy]

il just say this i hope that the mac pro for this year has nvidia gnu support. I am still torn on the whole matter i eventually want to go the nap route a year or two down the road but i prefer nvidia cards and right now the only way to get nvidia cards is the 5,1 or lower. but on the topic it will be an interesting year to see were apple take this

 

[koyoot]

According to some rumors, some boards of GP104 have GDDR5X. And TWO 8 pin connectors.
Secondly, Pascal has Asynchronous Compute capabilities. Worse than GCN1.0 but it has Async capabilities.

Edit: http://forums.anandtech.com/showpost.php?p=38156966&postcount=73 I suggest reading this post.