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

[DrMrLordX]

If you're going to do big power draw APU's properly it means custom motherboards, custom case designs etc etc. Will we get those? Maybe.

I'm sure it would have its own socket. So, in that sense, yeah, it'll be "custom", at least to the extent that no other CPU/APU would share the same socket. They might even just sell the things in BGA form factor, pre-soldered, though I'm not sure how well that would go over with critics . . .

Yes, cooling a 200 watt Video card and a 100 watt CPU would be a lot easier than a 300 watt APU.

Remember Video cards are BGA so they have direct heatsink to die contact. They also have fairly large die areas with heat that is not so concentrated in certain parts of the die. That and the highest powered ones have a blower design so they force the hot air outside the case.

Plenty of aftermarket, high-end cards actually vent hot air into the case. It's usually the reference designs that use the blower fans.

Direct die-to-HSF (which actually isn't direct, there's still TIM or solder involved) isn't that big of a deal either. A proper soldered nubshield isn't going to hurt cooling performance much.

Regardless, expect a HUGE die size which is what I was trying to say in the first place.

And if this large die APU is going to have conventional positioning within the case and TIM between the IHS and the die I think it will be more difficult to cool than a high TDP video card (which granted are known to have TDPs exceeding 250 watts)

They solder FX chips, they'll solder this thing. Unless they're just stupid.

According to these sources the CPU and GPU of PS4 and Xbox One are not MCM, they are integrated together on one die:

Actually I went out and got die shots of the APUs in question:

http://www.redgamingtech.com/xbox-one-die-shots-revealed-apu-info-comparison-against-ps4/

I was mistaken on the MCM part. Regardless, AMD makes MCM Opterons so they have recent experience with such package layouts.

 

[Fjodor2001]

After having read some articles on this, my guess is that the 16 core APU will be for the server market, and then there will be an 8 core variant for desktop.

Since 200-300 W was for the 16 core part, the desktop SKUs will likely be lower TDP. That is unless they compensate by adding more GPU cores on the 8 core SKUs.

 

[cytg111]

300 Watt APU on 14nm .. is that even doable? That is not alot of space^2 to dissipate 300 watts.. Look at haswell (moreso broadwell) having throttling issues/going over 100C just getting rid of sub 100 watts.

 

[el etro]

300 Watt APU on 14nm .. is that even doable? That is not alot of space^2 to dissipate 300 watts.. Look at haswell (moreso broadwell) having throttling issues/going over 100C just getting rid of sub 100 watts.

Lower clocks will allow it.

But first they need at least the 14nm LPP process being ready. At this time, only a lower power variant of it is ready.

 

[frozentundra123456]

Lower clocks will allow it.

But first they need at least the 14nm LPP process being ready. At this time, only a lower power variant of it is ready.

What do lower clocks have to do with it? Lower clocks would allow the TDP to be lower, but it would not help with heat dissipation. 300 watts is 300 watts.

But this whole thing is a lot of speculation at this point. Has AMD even officially announced it? Last I heard Zen was suppose to be 95 watts.

 

[Yuriman]

What do lower clocks have to do with it? Lower clocks would allow the TDP to be lower, but it would not help with heat dissipation. 300 watts is 300 watts.

But this whole thing is a lot of speculation at this point. Has AMD even officially announced it? Last I heard Zen was suppose to be 95 watts.

It would have to be a very massive die with very low heat output per area - low voltage, low clocks, large area.

 

[el etro]

What do lower clocks have to do with it? Lower clocks would allow the TDP to be lower, but it would not help with heat dissipation. 300 watts is 300 watts.

But this whole thing is a lot of speculation at this point. Has AMD even officially announced it? Last I heard Zen was suppose to be 95 watts.

300 Watts is easy to cool, just need bigger cooler. A Titan style blower already cools 250W or more. An air-cooler like a ZalmanCNPS12x can cool the 300W monster.


And yes, the 300W HPC APU was announced on a event at Japan.

 

[jpiniero]

If AMD is to have any shot with servers, K12 NEEDS to go small along with HDL. Now that I think about it Zen may simply be the design without HDL and the x86 tax. Still, it should be pretty small; 16 cores is actually pretty feasible.

 

[cbn]

Plenty of aftermarket, high-end cards actually vent hot air into the case. It's usually the reference designs that use the blower fans.

On those aftermarket cards, the extra GPU fans allow the video card to be quieter, but it needs to be installed in a case with a lot of fans.

For an OEM proposition, I expect they will not be so generous with the fans. To make this feasible at acceptable noise levels the TDP needs to be reasonable enough and the position of the processor within the case needs to be such that heat can easily be evacuated.

P.S. IntelUser2000 brought up some good point in posts #96 and #99 that GPUs run 15C to 20C hotter than CPUs. The TDP on Video cards also includes the PCB, memory and other components while the TDP of a CPU is only for the processor. I think that is a good explanation for why we see coolers on video cards (even rather high TDP ones) being fairly manageable size in comparison to what we see happening with CPU coolers.

 

[VirtualLarry]

I look forward to mo.mo.mo.more.cores!

 

[jhu]

I look forward to mo.mo.mo.more.cores!

Does AMD have anything that competes with Knights Landing? I'm thinking of getting that when it's available.

 

[NostaSeronx]

But first they need at least the 14nm LPP process being ready. At this time, only a lower power variant of it is ready.

Customer Tapeouts @ GlobalFoundries for 14nm LPe/LPP;
30-Sep-14 (High Density only)
31-Mar-15
30-Jun-15
30-Sep-15
7-Dec-15

LPP - High Performance -- 11.5 track
LPP - Fast High Density -- 9 track
LPP/LPe - High Density -- 9 track

fHD equals HP in perf and HD in areal density. Current in sources point to STMicro's 14nm High Performance FDSOI being cheaper and cooler than Samsung's 14nm Bulk FinFETs.

28nm SHP to either 14nm FDSOI or 14nm FinFETs is a ~4x shrink.
28nm A to either is a ~2.5x shrink.

 

[DrMrLordX]

300 Watts is easy to cool, just need bigger cooler. A Titan style blower already cools 250W or more. An air-cooler like a ZalmanCNPS12x can cool the 300W monster.

That's what I'm saying. If they have to MCM it to avoid power density issues then so be it. They've already got 220W monsters in socket AM3+, it's not like AMD has absolutely no recent experience with handling massive heat loads coming out of a single socket. I can't accurately predict exactly how they'll do it, but the signs are all there that it would be possible, even on air.

Heck an nh-d15 at least has enough fin surface area to do it (it's basically a 280mm rad all on its own, for crying out loud). Not sure if the heat pipes are enough for the job, but give it a bigger base to fit a larger IHS and larger heatpipes and I'm sure it would be fine.

Does AMD have anything that competes with Knights Landing? I'm thinking of getting that when it's available.

Competes in what way? Knight's Landing is an interesting product in that it splits the difference between HPC/GPGPU and general-purpose computing. I don't think AMD is ever quite going to have a product that's as versatile as that, since APUs still rely on two different kinds of compute resources. In terms of theoretical maximum performance, a 290x already has ~5.6 TFlops in 32-bit fp, isn't that faster than the projected 32-bit fp performance of Knight's Landing (~3-3.5 TFlops)?

 

[jhu]

In terms of theoretical maximum performance, a 290x already has ~5.6 TFlops in 32-bit fp, isn't that faster than the projected 32-bit fp performance of Knight's Landing (~3-3.5 TFlops)?

It is, although it's easier to run straight x86 code than to rewrite everything for OpenCL. Of course Knights Landing should also be able to do 64-bit FP at 1/2 the 32-bit rate vs. the 290x doing 64-bit at 1/8 the 32-bit rate.

 

[Nothingness]

Isn't Knight Landing theoretical perf 3 TFLOPS for DP and 6 TFLOPS for SP?

 

[cbn]

They've already got 220W monsters in socket AM3+, it's not like AMD has absolutely no recent experience with handling massive heat loads coming out of a single socket. I can't accurately predict exactly how they'll do it, but the signs are all there that it would be possible, even on air.

Heck an nh-d15 at least has enough fin surface area to do it (it's basically a 280mm rad all on its own, for crying out loud). Not sure if the heat pipes are enough for the job, but give it a bigger base to fit a larger IHS and larger heatpipes and I'm sure it would be fine.

For DIY something like the NH-d15 or that ZalmanCNPS12x might work, but for OEM these kind of coolers will likely be too heavy to ship in a completed system.

This leaves an AIO liquid cooler or some type of exotic air cooling arrangement that is not prone to damage as the most likely options for an OEM system.

 

[jpiniero]

Isn't Knight Landing theoretical perf 3 TFLOPS for DP and 6 TFLOPS for SP?

Yeah. That being said, Hawaii is natively 1/2 DP. The W9100 is rated at 5.2/2.6 at 275 W.

No idea on how realistic, but I was trying to come up with a theoretical product where each cluster would have 4 K12/Zen cores alongside some amount of GCN cores for vector and parallel processing. Like a full cluster would have 1024 cores and that would be the basis of an AMD product to compete with Pascal Tesla and the Phi.

 

[monstercameron]

It is, although it's easier to run straight x86 code than to rewrite everything for OpenCL. Of course Knights Landing should also be able to do 64-bit FP at 1/2 the 32-bit rate vs. the 290x doing 64-bit at 1/8 the 32-bit rate.

I'm no programmer do forgive the ignorance but is knl really just straight x86 code? From what I gather it is possible to port code directly from a Xeon and it should run but that doesn't mean it will run well. Wouldn't one have to take advantage of a tonne of vectorization to get good performance out it? Which might be would seem to be non trivial to rework your code base for it.

 

[MiddleOfTheRoad]

What do lower clocks have to do with it? Lower clocks would allow the TDP to be lower, but it would not help with heat dissipation. 300 watts is 300 watts.

But this whole thing is a lot of speculation at this point. Has AMD even officially announced it? Last I heard Zen was suppose to be 95 watts.

This is a non-issue. I mean seriously.... Just look at the 9000 series of AMD FX processors -- if a cooler can keep those things going, they should have no problem with these new APU's.

 

[Yuriman]

This is a non-issue. I mean seriously.... Just look at the 9000 series of AMD FX processors -- if a cooler can keep those things going, they should have no problem with these new APU's.

The 9000 series FX chips are 220w and first shipped with an AIO watercooler.

 

[ShintaiDK]

This is a non-issue. I mean seriously.... Just look at the 9000 series of AMD FX processors -- if a cooler can keep those things going, they should have no problem with these new APU's.

220W to 300W is quite a jump. And 220W already requires watercooling from AMDs perspective.

 

[NTMBK]

I'm no programmer do forgive the ignorance but is knl really just straight x86 code? From what I gather it is possible to port code directly from a Xeon and it should run but that doesn't mean it will run well. Wouldn't one have to take advantage of a tonne of vectorization to get good performance out it? Which might be would seem to be non trivial to rework your code base for it.

If your code is already structured to vectorise well on Haswell, moving up to KNL shouldn't be too hard. Even more so for SKX, which has AVX512.

 

[Gikaseixas]

Im cooling off around 220 watts with an air solution. It does exhausts lots of hot air when fully loaded thou

 

[artivix]

Its Q2 2015, where are these 14nm chips?

Mubadala Development Co., the Abu Dhabi-based investment and development company, said this week that GlobalFoundries, which is owned by Mubadala, has already begun to ramp up production of chips using 14nm fabrication process.

“GlobalFoundries announced a strategic collaboration with Samsung to deliver capacity at 14nm, one of the industry’s most advanced nodes, as Fab 8 in Malta, New York began ramping production for customers,” a statement by Mubadala reads.

http://www.kitguru.net/components/a...s-begins-to-ramp-up-production-of-14nm-chips/

Several business news sites are also reporting that Samsung and Globalfoundries have won contracts for Apples A9 chips.

They indicate that Apple purchases $25 billion worth of chips per year and winning this contract will be a enormous economic boost for the companies!

 

[MiddleOfTheRoad]

220W to 300W is quite a jump. And 220W already requires watercooling from AMDs perspective.

Exactly, AMD has already offered watercooling from the factory. Why would this APU be any different?
They can clock it accordingly to whatever TDP they desire.

Again, do I really need to mention that some people run 2 or 4 video cards with TDP ranges from 750 to 1000 watts? That's a lot of heat for a case/cooling to manage. A mere 300 doesn't sound like much of an issue.

BTW, Shintai -- not to burst your "I hate AMD because their chips are so hot" bubble..... but Intel already makes chips with 300 watt TDP's. They're called Xeon Phi's. So whoops.... 300 Watts must be quite a jump since there are already chips being made with that TDP. Notice that the Xeon Phi processors are now packaged more like Video Cards.

http://www.cpu-world.com/news_2013/2013062001_Intel_expands_Xeon_Phi_co-processor_lineup.html

I know the local college would have loved to use a Xeon Phi setup -- but the grant couldn't cover the cost.

http://www.intel.com/content/www/us/en/processors/xeon/xeon-phi-detail.html

"More cores! More performance!" -Intel