[AMD Processors] The future CPUs

[mohit9206]

I have a few questions about AMD Zen APU's
1. What kind of performance increase can we expect from the iGPU of Zen compared to Kaveri?
2. Will Zen APUs support AMD dual graphics like current APUs?

 

[myocardia]

40% IPC increase, if true, will mean a heavy clock regression.

Haha, what's wrong with that? That's exactly what we've all been telling AMD they needed to do, for 3½ or 4 years now. Lower their clockspeeds, and raise their IPC. If done properly, this has the potential to be AMD's 'Conroe moment'. I mean sure, they're not also going to bypass their competition in one fell swoop like Intel did with Conroe, but who would expect them to at this point? They're monetarily well under half the company they were in 2006, when Intel released Conroe.

For one 40% is obviously an average, it will be less in Integer and likely more in FP,

Exactly. Stop sharing FP units, and you've magically regained 20% of your FPU IPC*. Hire (or promote) the guy who can raise the performance of the cache subsystem, and FX IPC will rise considerably, across the board. Spend some more time tweaking the cache hit ratio, and that will also raise performance in both FPU and integer.

for instance we can be sure that the FPU will be reused since its throughput is as good as the one of a Haswell core.

Umm, no. Even if we give the FX FP unit back the 20% that Dirk Meyer stole from it, it still has nowhere near the single-thread IPC of Haswell. Now, give it the improvements to the cache, and whatever other IPC improvements they can come up with, in addition to recovering the stolen 20%, and it may very well give Ivy Bridge a run for its money in FP performance, which would be awesome.

edit: I would buy a 16 thread Zen today, if it had roughly the IPC of Ivy Bridge, and had improved power usage, compared to Vishera.

Thus FP IPC can be increased by much more than those 40%,

Yeah, I think so as well. Well, maybe not a lot more, but definitely more than 40%, since just stopping the sharing of FPUs will immediately 'give back' 20%.


*Every reference in this thread to IPC is actually a reference to single-threaded performance of software, not to what CPU architects call IPC.

 

[AtenRa]

Just for fun (im not suggesting this will be ZEN performance),

Take the single thread performance in CB-15 below, keep the same clocks as Kaveri and add 40%.

 

[myocardia]

1. What kind of performance increase can we expect from the iGPU of Zen compared to Kaveri?
2. Will Zen APUs support AMD dual graphics like current APUs?

1. Nobody knows yet, its ~18 months away. That's what we've been speculating about for the past page or so. About the only guarantee on CPU performance is that it will not be lower performing than its predecessor, the way they did with Bulldozer. No tech company is that dumb.

Unless they're crazy, they'll give the iGPU the same memory tech that is in their current top of the line dedicated video card, the Fury X, known as HBM. Around here, we call it HBM1, since in the next year or 18 months, there is supposed to be an HBM2. It will of course have fewer stream processors, texture units, and maybe even ROPs, than the Fury X. It shares its TDP with the CPU, which means that it is forced to share its ability to be cooled properly. If they can give it the performance of today's 390, it will be a runaway bestseller.

2. I don't know why they wouldn't. They'd be shooting themselves in the foot, if they didn't. Besides, by the time the Zen is released, Windows 10 will be the operating system, and it is supposed to be supporting Crossfiring/SLIing across architectures, so I don't believe AMD could even stop it from happening, if they were dumb enough to want to stop it.

 

[mohit9206]

1. Nobody knows yet, its ~18 months away. That's what we've been speculating about for the past page or so. About the only guarantee on CPU performance is that it will not be lower performing than its predecessor, the way they did with Bulldozer. No tech company is that dumb.

Unless they're crazy, they'll give the iGPU the same memory tech that is in their current top of the line dedicated video card, the Fury X, known as HBM. Around here, we call it HBM1, since in the next year or 18 months, there is supposed to be an HBM2. It will of course have fewer stream processors, texture units, and maybe even ROPs, than the Fury X. It shares its TDP with the CPU, which means that it is forced to share its ability to be cooled properly. If they can give it the performance of today's 390, it will be a runaway bestseller.

2. I don't know why they wouldn't. They'd be shooting themselves in the foot, if they didn't. Besides, by the time the Zen is released, Windows 10 will be the operating system, and it is supposed to be supporting Crossfiring/SLIing across architectures, so I don't believe AMD could even stop it from happening, if they were dumb enough to want to stop it.

R9 390 level performance is impossible even in the highest end Zen because then there will be too many challenges like
1. High combined TDP of CPU+GPU will be over 200 watts.
2. APU will be very expensive.
Don't forget AMD wants Zen to be affordable and power efficient . My much more realistic guess is r7 260x level performance in the top end APU.

 

[ShintaiDK]

Haha, what's wrong with that? That's exactly what we've all been telling AMD they needed to do, for 3½ or 4 years now. Lower their clockspeeds, and raise their IPC. If done properly, this has the potential to be AMD's 'Conroe moment'. I mean sure, they're not also going to bypass their competition in one fell swoop like Intel did with Conroe, but who would expect them to at this point? They're monetarily well under half the company they were in 2006, when Intel released Conroe.

But at what clocks and power consumption. It doesnt matter if they just exchange one problem for the other.

Also we dont even know yet what AMD actually specifies as IPC increase. It could for the sake of argument be SMT.

I dont even think AMD knows yet, since its unlikely they actually have any silicon.

 

[BigDaveX]

Haha, what's wrong with that? That's exactly what we've all been telling AMD they needed to do, for 3½ or 4 years now. Lower their clockspeeds, and raise their IPC. If done properly, this has the potential to be AMD's 'Conroe moment'. I mean sure, they're not also going to bypass their competition in one fell swoop like Intel did with Conroe, but who would expect them to at this point? They're monetarily well under half the company they were in 2006, when Intel released Conroe.

What AMD really needs is something akin to the K6-2 or Athlon XP; a chip that, while not quite able to compete with Intel at the higher-end, can undercut Intel's low-to-mid range chips while still being hugely profitable. Right now one of AMD's major problems is that their die sizes are gigantic in comparison to Intel's, so Zen addressing that alone would put them in a better position.

 

[myocardia]

R9 390 level performance is impossible even in the highest end Zen because then there will be too many challenges like

Don't forget AMD wants Zen to be affordable and power efficient . My much more realistic guess is r7 260x level performance in the top end APU.

Well, you're welcome to make that outright guess as to the performance of the Zen iGPU. Like I said already though, nobody on Earth knows today how it will be performing, which is why I didn't/won't guess. It will nevertheless be a runaway bestseller if it has 390ish performance, though!

 

[myocardia]

But at what clocks and power consumption. It doesnt matter if they just exchange one problem for the other.

Also we dont even know yet what AMD actually specifies as IPC increase. It could for the sake of argument be SMT.

I dont even think AMD knows yet, since its unlikely they actually have any silicon.

Haha, I agree with every single word you typed here. It's not like I'm convinced that all of those things I said earlier will be happening, and will be on time, under budget, etc. It's just that they're all still possibilities at this point. Since I would bet money that AMD has no functioning silicon at this time, we can only wait, and speculate.

 

[AtenRa]

Well lets see. If AMD wants to compete against Skylake/Kaby GT4e (Quad Core 8x Threads die size of 200mm2+) then they would need a 200mm2+ die as well.

With 14nm FF they will be able to install 2x the iGPU of Kaveri/Carrizo at close to 200-220mm2. That means close to R9 270 performance if they will use HBM2. That APU will be close to the PS4 GPU performance but at lower power and smaller die.

 

[ShintaiDK]

A GT4e looks to be around 160mm2 after looking at the Skylake quad GT2 die.

 

[Keysplayr]

40%? I'll take it. Beats the hell out of any improvements AMD has offered since being Conroe'd. I welcome the 40% IPC per core increase. Now, we have to factor in the hype quotient into the formula. So, on my chalkboard here, I am getting 10% actual increase in IPC per core without setting foot in the mobius continuum.

 

[AtenRa]

A GT4e looks to be around 160mm2 after looking at the Skylake quad GT2 die.

Broadwell GT3e (2x 24 EUs) Dual Core 4 threads is 133mm2.
Add two more cores (~20mm2) PLUs another 24 EUs (~40mm2) and you will have 200-220mm2 Skylake Quad Core GT4e.

 

[ShintaiDK]

Broadwell GT3e (2x 24 EUs) Dual Core 4 threads is 133mm2. Add two more cores (~20mm2) PLUs another 24 EUs (~40mm2) GT3 iGPU and you will have 200-220mm2 Skylake Quad Core GT4e.

6700K is ~120mm2 and the die looks like this.

Double the IGP and you got ~160mm2. And in case you wonder, eDRAM controller sits on the system agent and is present in all SKUs. eDRAM is purely optional.

Big mistake trying to compare Gen8 with Gen9

https://software.intel.com/sites/de...ure-of-Intel-Processor-Graphics-Gen9-v1d0.pdf

 

[StrangerGuy]

40%? I'll take it. Beats the hell out of any improvements AMD has offered since being Conroe'd. I welcome the 40% IPC per core increase. Now, we have to factor in the hype quotient into the formula. So, on my chalkboard here, I am getting 10% actual increase in IPC per core without setting foot in the mobius continuum.

Remember how "Barcelona would have 40% better IPC than Q6600"? 7 years later, still can't beat an ancient Penryn in ST IPC.

 

[AtenRa]

No.

6700K is ~120mm2 and the die looks like this.

Double the IGP and you got ~160mm2. And in case you wonder, eDRAM controller sits on the system agent and is present in all SKUs. eDRAM is purely optional.

Big mistake trying to compare Gen8 with Gen9.

Skylake GT2 doesn't have eDRAM controller.

 

[ShintaiDK]

Skylake GT2 doesn't have eDRAM controller.

It does:
https://software.intel.com/sites/de...ure-of-Intel-Processor-Graphics-Gen9-v1d0.pdf

Its not used, but its there. Its quite clear that Intel is ready to enable all products with eDRAM almost instantly if needed.

In Gen9 EDRAM now acts as a memory-side cache between LLC and DRAM. Also, the EDRAM memory controller has moved into the system agent, adjacent to the display controller, to support power efficient and low latency display refresh.

 

[ShintaiDK]

Remember how "Barcelona would have 40% better IPC than Q6600"? 7 years later, still can't beat an ancient Penryn in ST IPC.

Thats really the problem with AMD marketing and reality. If we flat out was to believe it all. They be over twice as fast as Intel by now in IPC. AMD havent delivered to its promise for big cores since the K8. And thats 12 years ago since it first launched.

 

[AtenRa]

It does:
https://software.intel.com/sites/de...ure-of-Intel-Processor-Graphics-Gen9-v1d0.pdf

Its not used, but its there. Its quite clear that Intel is ready to enable all products with eDRAM almost instantly if needed.

What it says is, the eDRAM controller is optional and when available it has moved inside the system agent.
There is no eDRAM controller with the GT2 dies. The System agent in GT2 die is too small to also include the eDRAM controller and Intel would not increase the already expensive 14nm die with added die space if they wouldn't use it.

 

[ShintaiDK]

What it says is, the eDRAM controller is optional and when available it has moved inside the system agent.
There is no eDRAM controller with the GT2 dies. The System agent in GT2 die is too small to also include the eDRAM controller and Intel would not increase the already expensive 14nm die with added die space if they wouldn't use it.

The phrase "when available" isnt even in the PDF. Please link the exact phrase if it exist.

HT and 2MB cache are also in i5. Yet wasted on purpose.

eDRAM GT2 models could be very attractive for Xeons.

 

[Essence_of_War]

What it says is, the eDRAM controller is optional and when available it has moved inside the system agent.
There is no eDRAM controller with the GT2 dies. The System agent in GT2 die is too small to also include the eDRAM controller and Intel would not increase the already expensive 14nm die with added die space if they wouldn't use it.

Uhhh quote from the pdf please?

On page 4 it clearly indicated the eDRAM itself is optional, but it doesn't say that the controller is optional.

 

[AtenRa]

ok i had a better view of it and yes it seems the eDRAM controller is indeed present.

So the GT4e quad core die would be close to 180mm2 then, if the GT2 Quad Core is 120mm2.

 

[AtenRa]

So, a 1024 Shaders, 16x Compute units, 8x ACES GCN 2.0 Quad Core ZEN APU with HBM2 at 200mm2 will be way faster in GPU performance than GT4e Skylake/Kaby in 2017.

With 16CU (1024 Shaders) GCN2.0 + HBM2 could be faster than current R7 270 in DX-12 games.


Edit: Just a thought,
2016 APUs Socket AM4 Carizzo 8x CU iGPUs paired with 3200MHz DDR-4 will be equal if not faster then current Broadwell GT3e.

 

[ShintaiDK]

You really think it will have HBM2? And by the time it may launch in 2017 it could compete with Cannonlake instead.

And who knows if/when GCN 2.0 will ever exist. AMD more or less just hinted a shrinked GCN 1.2 for 14nm GPUs.

If we go that road we can just as well make up some crazy performance metric for Gen10 and 11 IGPs.

 

[sm625]

Passmark is one of the most well rounded measures of single threaded performance. And according to passmark, a 7850K would need a 46% increase in ST performance to match a 6700K.