Could the next nMP have AMD Zen?

[johnnyturbouk]

I was researching about upgrading to a nMP, when google took me onto upcoming intel processors, and then came across this article:

AMD Zen

and was intrigued with apple moving to AMD GPU, could they use AMD for their core CPUs?

Just a thought,


I was previously thinking of getting a RiMAC and then a nMP later, however, with target display mode not feasible on the 5k imac, looks like a refurbished nMP and a decent 4K IPS screen.

I think i will go for:
Refurbished Mac Pro 3.5GHz 6-Core Intel Xeon E5
16GB memory
256GB PCIe-based flash storage
Dual AMD FirePro D500 graphics processors with 3GB of GDDR5 VRAM each

 

[JesterJJZ]

No.

 

[deconstruct60]

I was researching about upgrading to a nMP, when google took me onto upcoming intel processors, and then came across this article:
....
and was intrigued with apple moving to AMD GPU, could they use AMD for their core CPUs?

Impossible? No. The odds of it happening are very slim. AMD Zen might get AMD back into the zone where Intel doesn't look at AMD's higher end offerings and simply chuckle. They might pay more attention to what they are doing, but it is doubtful that AMD will get to parity. They are being increasingly left behind in the rear view mirror now. What AMD needs is to stop the bleeding of the gap getting bigger over time. Not parity.

The fact that AMD is also chasing 64 bit ARM at the same time is indicative that they are not putting all of their wood-behind-one-arrow. That it is going to take them 4 years to unwind from Bulldozer's blunders is actually indicative of how poorly they are set up to do major leaps forward. In the same 4 years, Intel hasn't spent the R&D pipeline time unwinding significant blunders at the high end space. It is going to be easier for Intel to take stuff that works and move it to the next process tech than it will be for AMD to get things working and move to process tech at the same time.

This x86 and ARM in parallel smells like AMD is going to throw leaky power at the x86 solution ( 200W hot rod, funny car ) server package and try to cover server system folks who are on power budgets with ARM offering. That is a looser strategy for trying to get placed in a Mac Pro. It is questionable whether it will even work in the server space. That may get much deeper traction in the hot rodder workstation crowd.


AMD is getting Apple GPU wins in part because Nvidia has decided that throwing OpenCL under the bus is a good strategy for Nvidia. Perhaps it is a good strategy for Nvidia overall, but it is a looser strategy for trying to win Apple design bake-offs. It doesn't really lend huge leverage to AMD's x86 work though.

 

[bennibeef]

No. nMP without thunderbolt? nope.

 

[johnnyturbouk]

Impossible? No. The odds of it happening are very slim. AMD Zen might get AMD back into the zone where Intel doesn't look at AMD's higher end offerings and simply chuckle. They might pay more attention to what they are doing, but it is doubtful that AMD will get to parity. They are being increasingly left behind in the rear view mirror now. What AMD needs is to stop the bleeding of the gap getting bigger over time. Not parity.

The fact that AMD is also chasing 64 bit ARM at the same time is indicative that they are not putting all of their wood-behind-one-arrow. That it is going to take them 4 years to unwind from Bulldozer's blunders is actually indicative of how poorly they are set up to do major leaps forward. In the same 4 years, Intel hasn't spent the R&D pipeline time unwinding significant blunders at the high end space. It is going to be easier for Intel to take stuff that works and move it to the next process tech than it will be for AMD to get things working and move to process tech at the same time.

This x86 and ARM in parallel smells like AMD is going to throw leaky power at the x86 solution ( 200W hot rod, funny car ) server package and try to cover server system folks who are on power budgets with ARM offering. That is a looser strategy for trying to get placed in a Mac Pro. It is questionable whether it will even work in the server space. That may get much deeper traction in the hot rodder workstation crowd.


AMD is getting Apple GPU wins in part because Nvidia has decided that throwing OpenCL under the bus is a good strategy for Nvidia. Perhaps it is a good strategy for Nvidia overall, but it is a looser strategy for trying to win Apple design bake-offs. It doesn't really lend huge leverage to AMD's x86 work though.

Interesting answer, thanks!

 

[deconstruct60]

No. nMP without thunderbolt? nope.

Thunderbolt is present in more non Intel systems than in Intel CPU ones (i.e., peripherals usually don't ). AMD + TB isn't necessarily impossible. However, AMD is certainly no big fan of Thunderbolt. Probably even less with the arrival of USB 3.0+ and Type C (alternate mode) connectors. There is firmware , boot , and integration work that AMD hasn't done nothing on to date (but Zen isn't arriving relatively soon either). The relatively low run-rate for the Mac Pro is unlikely going to change AMD's mind. A $55M revenue sale probably wouldn't be enough for them to change their mind. $500-900M might but I doubt Apple would buy that much stuff without shifting major portions of the entire Mac line up.

In contrast, Intel invests deeply in greasing-the-wheels for the use of TB system ( including reference system designs). It is much less work for Apple to just ride the bow wave Intel creates. But it isn't really a threat that Intel can dangle over Apple's head. For the immediate future, without Apple TB is dead. Until some other major system builder(s) starts buying in at least as large numbers, if Apple walks away there is no momentum.


The notion that Apple has no choice but only buy Intel CPU I think is incorrect.


If TB merges with the USB Type C connector it will be even easier for Apple to dump TB if Intel chooses to primarily using it as a 'stick' (lock-in tool) as opposed to a 'carrot' ( a useful tool in whatever system design want to do).

 

[TechGod]

It is certainly plausible if AMD offers Apple what they need. The HBM on die should prove to be very tempting for Apple.

 

[deconstruct60]

It is certainly plausible if AMD offers Apple what they need. The HBM on die should prove to be very tempting for Apple.

It is not very likely that Zen is getting direct access to HBM any time soon. HBM isn't an "on die" technology. Same package but not the same die.

Even if AMD build one of these rumored Exascale Heterogeneous Processor (EHP) ( http://www.fudzilla.com/news/processors/38402-amd-s-coherent-data-fabric-enables-100-gb-s ) those are not the bulk of the Zen line up. Those more so a modules highly targeted at supercomputer builders not particularly where the Mac Pro is positioned. Can put the CPU and a HBM GPU onto the same logical board with a large MCM package but then have a cooling problem ( higher heat in a denser volume) and a scaling issue with the DDR4 for the bulk of the memory ( for workloads that need to scale past what HBM capacities can store local). For supercomputers you can just throw in another board and add bigger fans (and/or make more noise. )

This MCM approach of coupling CPU with GPU means if want to go 2 GPUs also have to drag along another CPU. If trying to build a exascale ( exaFLOP ) supercomputer that is useful. For a "grow as you go" render farm... it is a decent match (makes for decent blades/nodes can probably just easly add over time). For a workstation not quite as much.

AMD might make their money back on these EHP products but they aren't going to being enough to keept the lights on. Zen has to competitive against the mainstream Core i line up for AMD to be healthy again. Some of Intel's business at the top end of the supercomputer spectrum would be nice but that isn't where the bulk of the Zen effort should be spent.

Looks like this EHP will be an Opteron flavor offering where the x86 cores will perhaps get access to a "Last Level Cache" similar to the eDRAM designs that Intel is using ( only may hop through the GPU cores to get to it).

 

[TechGod]

It is not very likely that Zen is getting direct access to HBM any time soon. HBM isn't an "on die" technology. Same package but not the same die.

Even if AMD build one of these rumored Exascale Heterogeneous Processor (EHP) ( http://www.fudzilla.com/news/processors/38402-amd-s-coherent-data-fabric-enables-100-gb-s ) those are not the bulk of the Zen line up. Those more so a modules highly targeted at supercomputer builders not particularly where the Mac Pro is positioned. Can put the CPU and a HBM GPU onto the same logical board with a large MCM package but then have a cooling problem ( higher heat in a denser volume) and a scaling issue with the DDR4 for the bulk of the memory ( for workloads that need to scale past what HBM capacities can store local). For supercomputers you can just throw in another board and add bigger fans (and/or make more noise. )

This MCM approach of coupling CPU with GPU means if want to go 2 GPUs also have to drag along another CPU. If trying to build a exascale ( exaFLOP ) supercomputer that is useful. For a "grow as you go" render farm... it is a decent match (makes for decent blades/nodes can probably just easly add over time). For a workstation not quite as much.

AMD might make their money back on these EHP products but they aren't going to being enough to keept the lights on. Zen has to competitive against the mainstream Core i line up for AMD to be healthy again. Some of Intel's business at the top end of the supercomputer spectrum would be nice but that isn't where the bulk of the Zen effort should be spent.

Looks like this EHP will be an Opteron flavor offering where the x86 cores will perhaps get access to a "Last Level Cache" similar to the eDRAM designs that Intel is using ( only may hop through the GPU cores to get to it).

http://www.fudzilla.com/news/processors/37494-amd-x86-16-core-zen-apu-detailed

Just an FYI for others in the thread. I'll respond appropriately when I have the time.

 

[goMac]

Zen is interesting, but I don't think Apple would be that into an APU GPU in the Mac Pro.

 

[TechGod]

Zen is interesting, but I don't think Apple would be that into an APU GPU in the Mac Pro.

Nono. The GPU bit of the APU would work WITH a Fury Nano type GPU.

 

[deconstruct60]

Nono. The GPU bit of the APU would work WITH a Fury Nano type GPU.

the 16 x86 cores sharing die space with the APU-GPU means something has to give. It is a fixed amount of space (dies are only so big for lithography constraints). So more x86 cores means smaller GPU. Or vice versa. Perhaps Zen cores punt SSE/Altivec/AVX work into APU-GPU to save space, but likely sacrificing on other aspects too ( e.g., L3 size , etc. )

In the exascale context need the Greenland GPU added to the MCM because have tossed parallel horsepower from the APU chip because trying to juggle the same set of functionality as the overall MCM is trying to address.

Apple has problems squeezing max performance out of just two dGPUs or a dGPU and a Intel iGPU. This "layered, works with Greenland on a MCM" approach likely is more complexity added on top of what Apple already doesn't particularly handle all that exceptionally well. It isn't one of their strengths. So yes. they probably would back away from it.

 

[deconstruct60]

http://www.fudzilla.com/news/processors/37494-amd-x86-16-core-zen-apu-detailed

While these are linked to the article I posted from the same source covering the Zepplin coherent fabric (just keep following fudzilla "previous covered" links ) they don't don't match up.

This APU design doesn't have any coherent link connectivity functionality function block. ( the DDR4 controller is built in whereas in the Zepplin article the coherent chip is interfacing with DDR4 ). I suspect there are multiple products that will come out at multiple stages here.

One is a next gen console chip. That's kind of what this chips looks like HBM and DDR4 mixed into one SoC. The console game logic mostly in DDR4 (and some prefetched data if the DDR4 modules are cheap enough) and the graphics in the HBM. All rolled up into one $200/box package. I doubt those come in 2016. Likely a 2017 product ( if not also a Zen v2 or v1.1 design ) .


Those tweaked with Zepplin interface ( prehaps chop out HBM and DDR4 interface and add coherent one ) and used as x86 subcomponents to a MCM module? Perhaps. I suspect it will have substantial problems when get to supercomputer workloads with substantive scalar code sections in them.


If try to link the MCM module concept on the Zepplin article to this logic layout floor plan on this article then the x86 grouping could be its own seperate module. The Greenland and HBM are grouped like on the Zepplin article diagram.

Not a everything all on one die fashion. The "DDR4 Memory controllers below" are the same thing labeled Zeppelin above.

the I/O chipset ( platform processsor ) , PCIe and 1GbE could all be grouped in the x86 module.


I suspect there are multiple Zen derivative designs in flight and folks are cross matching the different floorplans in ways that don't make sense. .... or AMD is pitching them in ways that don't make sense ( throwing ideas against the wall in NDA meeting hoping to see what sticks. ) .

 

[TechGod]

While these are linked to the article I posted from the same source covering the Zepplin coherent fabric (just keep following fudzilla "previous covered" links ) they don't don't match up.

This APU design doesn't have any coherent link connectivity functionality function block. ( the DDR4 controller is built in whereas in the Zepplin article the coherent chip is interfacing with DDR4 ). I suspect there are multiple products that will come out at multiple stages here.

One is a next gen console chip. That's kind of what this chips looks like HBM and DDR4 mixed into one SoC. The console game logic mostly in DDR4 (and some prefetched data if the DDR4 modules are cheap enough) and the graphics in the HBM. All rolled up into one $200/box package. I doubt those come in 2016. Likely a 2017 product ( if not also a Zen v2 or v1.1 design ) .


Those tweaked with Zepplin interface ( prehaps chop out HBM and DDR4 interface and add coherent one ) and used as x86 subcomponents to a MCM module? Perhaps. I suspect it will have substantial problems when get to supercomputer workloads with substantive scalar code sections in them.


If try to link the MCM module concept on the Zepplin article to this logic layout floor plan on this article then the x86 grouping could be its own seperate module. The Greenland and HBM are grouped like on the Zepplin article diagram.

Not a everything all on one die fashion. The "DDR4 Memory controllers below" are the same thing labeled Zeppelin above.

the I/O chipset ( platform processsor ) , PCIe and 1GbE could all be grouped in the x86 module.


I suspect there are multiple Zen derivative designs in flight and folks are cross matching the different floorplans in ways that don't make sense. .... or AMD is pitching them in ways that don't make sense ( throwing ideas against the wall in NDA meeting hoping to see what sticks. ) .

I'm afraid I can't refute these points as I don't know enough about CPU architecture as you do. But I think its kinda obvious they'll have multiple Zen variants from dual core to 16 cores. It'd be stupid to sell only a 16 core variant.

 

[goMac]

Nono. The GPU bit of the APU would work WITH a Fury Nano type GPU.

As others have said, there really isn't a point to give up space on the CPU die and sacrifice CPU performance when 2/3s of the machine is build to hold GPU dies.