Fudzilla: New AMD Zen APU boasts up to 16 cores (plus Greenland GPU with HBM)

[Haserath]

4 CPU cores(8 thread) and 12 GPU cores.

16 cores would be too big for it to also have a decent GPU unless the cores themselves weren't very powerful.

 

[SlowSpyder]

There's an AMD patent on this subject...

Didn't know they were working on something like that, thanks.

 

[ShintaiDK]

That was just to illustrate... I was talking originally about this: http://www.google.com/patents/US20130346695

Thats just their specific cache implementation. An alternative to Intels eDRAM and HMC on KNL.

The fastforward project is basicly just to deal with hinderence for implementation of new technology.

 

[Fjodor2001]

Ok not really reserved but like what others said, i meant dedicated to graphics only.

Sorry, I still do not follow. Can you please clarify what tasks you envision those CPU cores would be dedicated to performing?

And why would it differ on this APU compared to current APUs, where no CPU cores are reserved or dedicated to the GPU in any way?

 

[AtenRa]

Think we could make some rough estimates on die size, if this is true?

If Zen is a "big core", some very rough back-of-the-envelope math suggests that 4 Zen cores may be in the range of 100-140mm², based on past designs without CMT, which would put 16 cores at 400-600mm². I can't imagine 400-600mm² + iGPU, which suggests to me that Zen will probably be more like cat cores.

I base this off very little, though, so take it with a grain of salt. I'd like to see what others come up with.

EDIT: For reference, current cat cores are just about half as fast per clock as Haswell.

A Bulldozer Module including 2MB L2 Cache is close to ~31mm^2 at 32nm. At 14nm FF those Modules would be less than half the size making a 4 Module 8 Core Bulldozer less than 60mm^2 (for the Modules alone).

 

[Yuriman]

A Bulldozer Module including 2MB L2 Cache is close to ~31mm^2 at 32nm. At 14nm FF those Modules would be less than half the size making a 4 Module 8 Core Bulldozer less than 60mm^2 (for the Modules alone).

Yes, but there's a lot more in a CPU than just modules/cores, even without integrated graphics. Plus, AMD is (I've read) ditching CMT, which was worth pretty significant area savings, if they are to be believed (which I do).

 

[AtenRa]

4 CPU cores(8 thread) and 12 GPU cores.

16 cores would be too big for it to also have a decent GPU unless the cores themselves weren't very powerful.

12 GPU cores for 14nm FF are very low, try at least 16-20 GPU Cores .

 

[turtile]

A Bulldozer Module including 2MB L2 Cache is close to ~31mm^2 at 32nm. At 14nm FF those Modules would be less than half the size making a 4 Module 8 Core Bulldozer less than 60mm^2 (for the Modules alone).

And HDL would make them even smaller...

 

[AtenRa]

Yes, but there's a lot more in a CPU than just modules/cores, even without integrated graphics. Plus, AMD is (I've read) ditching CMT, which was worth pretty significant area savings, if they are to be believed (which I do).

I thought you were talking about the modules alone.

If you want to see a complete CPU die then take the 4 Module Bulldozer which is 300mm^2 at 32nm.

At 14nm FF it would be close to 80-100mm2. Add another one for a 8 module 16 Cores CPU with everything double (memory channels, L3 etc) and we get at 160-200mm^2.
So nowhere near what you are suggesting.

 

[Yuriman]

I thought you were talking about the modules alone.

If you want to see a complete CPU die then take the 4 Module Bulldozer which is 300mm^2 at 32nm.

At 14nm FF it would be close to 80-100mm2. Add another one for a 8 module 16 Cores CPU with everything double (memory channels, L3 etc) and we get at 160-200mm^2.
So nowhere near what you are suggesting.

Gotcha. What numbers did you use for your scaling? I was under the impression the area scaling of GF/SS's 14nm wasn't going to be as big coming down from 20nm as 28 -> 20 was.

 

[flash-gordon]

Thats just their specific cache implementation. An alternative to Intels eDRAM and HMC on KNL.

The fastforward project is basicly just to deal with hinderence for implementation of new technology.

Of course is their specific implementation, but using stacked dram, as described in the text and the images. It's the same solution of the image I posted...

 

[AtenRa]

Gotcha. What numbers did you use for your scaling? I was under the impression the area scaling of GF/SS's 14nm wasn't going to be as big coming down from 20nm as 28 -> 20 was.

Roughly,

28nm has 10% higher density over 32nm
20nm LPM has 2x density over 28nm.
14nm should be another 10-15% higher density over 20nm.

Total = ~120-125% higher density over 32nm.

 

[Rickyyy369]

Not that I wouldn't love to see this come to fruition - but that sounds massive and expensive. What cooler is this monster going to require? AMD branded dual-120mm rad water coolers?

 

[Headfoot]

Four channel DDR4 and HBM? Seems unlikely.

I can see it, this has "Server Chip" written all over it. HBM could be a fast L4 cache in the gigabytes range, with DDR4 being for further yet capacity -- similar to what Intel announced for Knight's Landing when it operates as its own processor and not as a coprocessor.

 

[AtenRa]

I can see it, this has "Server Chip" written all over it. HBM could be a fast L4 cache in the gigabytes range, with DDR4 being for further yet capacity -- similar to what Intel announced for Knight's Landing when it operates as its own processor and not as a coprocessor.

I was thinking the same.

 

[Fjodor2001]

Not that I wouldn't love to see this come to fruition - but that sounds massive and expensive. What cooler is this monster going to require? AMD branded dual-120mm rad water coolers?

I think it should be compared to a high end CPU + discrete GFX card. In the light of that the TDP sounds reasonable, and within the range of such combos. Likely it will give more bang for the buck than buying separate discrete components.

 

[frozentundra123456]

Not only is it die size, it is tdp. Supposedly the tdp will be 95 watts. 14 nm or not, there is no way they are going to get 16 cpu cores plus an igp for 95 watts, unless it is small core or a big core server chip clocked at 2 ghz or something.

 

[MiddleOfTheRoad]

4 CPU cores(8 thread) and 12 GPU cores.

16 cores would be too big for it to also have a decent GPU unless the cores themselves weren't very powerful.

Not from what I read.
The Zen platform is rumored to be strong enough to support APU's with 200 to 300 watt TDP (to launch for consumers in early 2017).

I don't see anyone launching a chip with only 4 cores and 12 GPU cores for a 2017 design (at least for their consumer flagship). Those would be okay specs for an entry level chip, sure.

If they are talking about that much thermal headroom -- they have no problem giving us 8 actual CPU cores (with SMT for 16 threads).

Not sure how accurate these slides / figures are, but here is the source:
http://wccftech.com/amd-gpu-apu-roadmaps-2015-2020-emerge/

 

[mikk]

16 cores for the server variant and 8 vor consumer makes more sense to me.

 

[Enigmoid]

Roughly,

28nm has 10% higher density over 32nm
20nm LPM has 2x density over 28nm.
14nm should be another 10-15% higher density over 20nm.

Total = ~120-125% higher density over 32nm.

Makes sense too that they could also cut out some of the cache if they had HBM right there.

On a high end CPU though the actual cores are less than half the die with the rest cache, interconnect, and memory/IO.

Whatever the calculated die size is (without cache) I would add about 75% as uncore. Obviously with HBM this may be reduced as you could get away with a smaller L3 and then I would go with adding about 50% uncore.

 

[MiddleOfTheRoad]

16 cores for the server variant and 8 vor consumer makes more sense to me.

If they already have the die done for servers (16 cores) -- Why wouldn't they want to sell some at high margins to enthusiasts (assuming the performance is worth the price)?

With HBM, this could potentially be the first APU that actually replaces dedicated video cards for mainstream PC gamers.

 

[cbn]

If this is really 16 cores, it will not be for gamers.

Outside of HPC usage, Maybe video editors or other types of professionals?

P.S. Regarding CPU to GPU balance (a the high end), I did ask that in this post:

http://forums.anandtech.com/showpost.php?p=37289331&postcount=73

EDIT: The HP Workstation I linked from Videomaker.com has two Haswell E5 Xeons for a total of 16C/32T with 2.6 Ghz base clock and 3.4 Ghz turbo. The Kepler based Nvidia Quadro has 2304 CUDA cores. Total TDP for the three processors is 330 watts (90 watts for each Xeon and 150 watts for the Quadro).

 

[Shehriazad]

I also think that 16 cores means compute cores...and not CPU cores.


Kaveri has 12 of those.

4 CPU cores (or 2 "dual core modules")
8 GPU cores.


Since AMD will be ditching their old design and use something new in Zen it will very likely look like this:

4 CPU cores /8 threads
12 GPU cores
Some HBM (probably somewhere in the GB area for the high end versions)



I don't see how 12 GPU cores would be "low". Kaveris 8 cores are actually already very solid in performance....now use a new/improved architecture, improve GPU count by 50(!)% and add HBM to alleviate bandwidth issues stemming from DDR3.

Those 12 GPU cores would be enough to best 1080P gaming...and while likely not strong enough for serious 4K...we have yet to see what this future card would crossfire with...at 12GPU cores it would be a more grown card...thus possibly crossfiring with an X70 or even an X80.


This of course is all just speculation...but it's the only thing that really makes sense.

 

[Fjodor2001]

I also think that 16 cores means compute cores...and not CPU cores.

Then it would not add up to the ~300 W TDP that has been indicated in other articles. Also, the article in the OP specifically mentions 16 Zen cores.

Remember that this is on 14 nm, not 32/28 nm, so they have a lot more transistors to play with.

 

[Shehriazad]

Then it would not add up to ~300 W TDP as indicated in other articles.

Eh....wouldn't be too sure about that.

You go from 2 modules to 4 actual cores (most likely with extra threads).

You go from 8 GPUs to 12.

You add HBM.


Since AMD also wants to walk more toward SoC style...that is possible.

Also the reports say somewhere from 200W - 300W. AMD wouldn't possibly be able to already give us any sort of real value since I'm pretty sure that there won't even be an actual working sample out, yet. (Because someone would've leaked the heck out of that, already xD)


Let's say they reach 250W...perfectly adds up. I can see the new cores eating up 100ish Watt on their own as they would be more close to a FX8XXX series (4 ACTUAL cores, 8 threads) instead of what current APUs have. Then AMDs GPUs...if it's anywhere near a X70 or X80 model like I suggested...that can easily sneak in another 100W+ Watts...we know that AMD isn't too great with their GPU wattage.
Then you have another 50ish for whatever else they may add + some HBM. (which of course won't be sucking that much).

And we don't even know what those cores clock at...AMD likes to impress with high clocks rather than high IPC. Even with the new architecture I can already see AMD forcing their top end chips to the 4+ Ghz area by default.


But yea...SOMEWHERE BETWEEN 200-300 is what that "news article" said...not 300. For all we know it could end up being 210w.


Edit:I don't see how my prediction is unrealistic...see what happens to Intels new chip when pushing it just to 4 Ghz. AMD loves running their stuff at 4 Ghz base on their top chips.

This is not meant to be a flame post...I'm just saying...people seem to over/underestimate wattage all the time. And mind you...that Intel Chip is not stuffed with a fat GPU.