[AMD Processors] The future CPUs

[AtenRa]

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

Yes i believe it will have 3D HBM2 on die in 2017. An yes it may compete against Cannonlake.

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.

AMD have said they will release a new GPU architecture every two years. When was the last one ?? 2014 with Tonga, we are expecting a new one in 2016 and not just a 14nm shrink.

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

The thing is we already know how a 16x CU 1024 Shaders GCN 1 GPU performs today, but we have no real performance for Intel GT4e Gen 9 so anything about Gen 10-11 is only speculation.

 

[ShintaiDK]

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.

And you base that on the 6700K that scores 2308 vs the 4790K that scores 2532?

Or that the 7850K scores 1571 while the 7870K scores 1461?

Passmark is as useless as ever.

 

[Dresdenboy]

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

IPC?

"We expect across a wide variety of workloads for Barcelona to outperform Clovertown by 40 percent," Allen said. The quad-core chip also will outperform AMD's current dual-core Opterons on "floating point" mathematical calculations by a factor of 3.6 at the same clock rate, he said.

And then:

AMD is moving its manufacturing from a 90-nanometer process to a 65-nanometer process, permitting more circuitry to fit in a given amount of chip real estate. Even with that change, the quad-core chips will run more slowly, Allen said. He argued that it's worth the tradeoff, though, since the additional cores can run more jobs simultaneously, even if an individual job isn't completed as swiftly.

http://www.cnet.com/news/amd-go-to-barcelona-over-clovertown/
That sounds like some first signs of throughput computing. IIRC and IMO in the end the process brought the frequency expectations to a halt, costing already a lot of the promised performance.

BTW, just found this too:

"(On) average, we are 40 percent-plus better than the competition in performance and a little bit better in power, and the combination is 1.5X in terms of power/performance," said Pat Gelsinger

http://www.cnet.com/news/a-dazed-intel-shifts-into-comeback-mode/
He was right about power:

it outperforms Opteron by as much as 60% in Performance/Watt

but not about performance in any case:

Woodcrest led the way by anywhere from 18 to 35% at higher system loads

http://www.anandtech.com/print/2043/

Unfortunately there are no communication standars.

 

[sm625]

And you base that on the 6700K that scores 2308 vs the 4790K that scores 2532?

Or that the 7850K scores 1571 while the 7870K scores 1461?

Passmark is as useless as ever.

That is because a 6700k only has 10 samples, while a 4790k has many thousands of samples. So the 4790k numbers are probably more accurate. Over time, the discrepancy for the 6700k will smooth out. Same goes with the 7870k, it only has 29 samples. Passmark isnt the problem. Your ability to weigh the preponderance of the data is what is useless.

 

[ShintaiDK]

Yes i believe it will have 3D HBM2 on die in 2017. An yes it may compete against Cannonlake.

Using HBM2 will mean fixed and limited memory and skyhigh cost. Not just the HBM2 but also the dreaded interposer. I would say its cost prohibitive until 2020 or so.

AMD have said they will release a new GPU architecture every two years. When was the last one ?? 2014 with Tonga, we are expecting a new one in 2016 and not just a 14nm shrink.

AMD also said it would release a complete lineup of new GPUs, what did we get? When they talk to investors, they dont seem to share that idea. But rather a direct shrink or close to it.

The thing is we already know how a 16x CU 1024 Shaders GCN 1 GPU performs today, but we have no real performance for Intel GT4e Gen 9 so anything about Gen 10-11 is only speculation.

So is your Zen APU specs.

 

[rainy]

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

It is really not that difficult to find this news, however I'm hardly surprised remembering how "objective" you are: http://www.overclock3d.net/articles..._taped_out_their_first_14nm_finfet_products/1

 

[ShintaiDK]

That is because a 6700k only has 10 samples, while a 4790k has many thousands of samples. So the 4790k numbers are probably more accurate. Over time, the discrepancy for the 6700k will smooth out. Same goes with the 7870k, it only has 29 samples. Passmark isnt the problem. Your ability to weigh the preponderance of the data is what is useless.

So why do you compare a 7850K with 6700K?

 

[ShintaiDK]

It is really not that difficult to find this news, however I'm hardly surprised remembering how "objective" you are: http://www.overclock3d.net/articles..._taped_out_their_first_14nm_finfet_products/1

Thats IPC projections, not silicon performance.

 

[AtenRa]

Using HBM2 will mean fixed and limited memory and skyhigh cost. Not just the HBM2 but also the dreaded interposer. I would say its cost prohibitive until 2020 or so.

Read again what i have said, I said 2017 ZEN APU with 3D HBM 2.0 on die.

So is your Zen APU specs.

Yes my ZEN APU specs are a speculation on my part, but the performance is extremely close to reality with such specs.

 

[rainy]

Thats IPC projections, not silicon performance.

Obviously, you have no idea what tape-out means: https://en.wikipedia.org/wiki/Tape-out

 

[ShintaiDK]

Obviously, you have no idea what tape-out means: https://en.wikipedia.org/wiki/Tape-out

The article doesnt state that Zen is taped out. And its unlikely it has.

Learn to read what you post:

While AMD have not confirmed what products they have taped out

14nm GPUs will be the first products.

 

[ShintaiDK]

Read again what i have said, I said 2017 ZEN APU with 3D HBM 2.0 on die.

Yes my ZEN APU specs are a speculation on my part, but the performance is extremely close to reality with such specs.

What is Hynix/GloFo/TSMC roadmap for 3D HBM 2.0 on die? Cost?

So you already decided your speculation will be reality. :whiste:

 

[rainy]

The article doesnt state that Zen is taped out. And its unlikely it has.

Learn to read what you post:
14nm GPUs will be the first products.

What sort of proof you have or maybe it's one of your "famous" predictions?

Advanced Micro Devices said on Thursday that it had taped out its first products, which will be made using a FinFET process technology. While AMD does not reveal which products it had taped out, it is highly-likely that one of them is a highly-anticipated microprocessor based on “Zen” micro-architecture.

http://www.kitguru.net/components/g...-we-have-taped-out-our-first-finfet-products/

 

[AtenRa]

What is Hynix/GloFo/TSMC roadmap for 3D HBM 2.0 on die? Cost?

So you already decided your speculation will be reality. :whiste:

No, i simple remind you what my speculation was. But even if you need an interposer, HBM 2 will be introduced in 2016 GPUs from both AMD and NVIDIA. By 2017 it will be cheaper than low volume production cost HBM 1 is today.

 

[ShintaiDK]

No, i simple remind you what my speculation was. But even if you need an interposer, HBM 2 will be introduced in 2016 GPUs from both AMD and NVIDIA. By 2017 it will be cheaper than low volume production cost HBM 1 is today.

The cost reference for APUs is DDR3/DDR4. Not HBM1.

And what GPUs? Entire lineups or 200$+ cards?

 

[AtenRa]

The cost reference for APUs is DDR3/DDR4. Not HBM1.

But you always take the cost of HBM 1 when anyone talks about HBM 2.0. HBM 2.0 cost is not comparable to HBM 1, especially in 2017.
edit: also, a 200mm2 14nm HBM 2.0 ZEN APU will only compete against the GT4e Intel SKUs, meaning high-end very expensive SKUs.

And what GPUs? Entire lineups or 200$+ cards?

$200+ cards. Thats a lot more volume against $500-650 AMD only GPUs in 2015.

 

[ShintaiDK]

But you always take the cost of HBM 1 when anyone talks about HBM 2.0. HBM 2.0 cost is not comparable to HBM 1, especially in 2017.
edit: also, a 200mm2 14nm HBM 2.0 ZEN APU will only compete against the GT4e Intel SKUs, meaning high-end very expensive SKUs.

$200+ cards. Thats a lot more volume against $500-650 AMD only GPUs in 2015.

Lets see what it will compete with and how it will perform. AMD not only have to deliver more than a miracle on CPU performance, but they also need to deliver on the IGP and be backed by HBM2.

By that logic current APUs should ship with GDDR5. Yet they didnt.

 

[AtenRa]

Lets see what it will compete with and how it will perform. AMD not only have to deliver more than a miracle on CPU performance, but they also need to deliver on the IGP and be backed by HBM2.

By that logic current APUs should ship with GDDR5. Yet they didnt.

Current APUs dont have the necessary CPU performance to compete in the very expensive high-end segment. ZEN APUs on the other hand will have both a better higher performing CPU and 14nm FF process. With HBM 2.0 those APUs in 2017 could be very competitive against Intel GT4e SKUs, especially in Mobile.

 

[ShintaiDK]

Current APUs dont have the necessary CPU performance to compete in the very expensive high-end segment. ZEN APUs on the other hand will have both a better higher performing CPU and 14nm FF process. With HBM 2.0 those APUs in 2017 could be very competitive against Intel GT4e SKUs, especially in Mobile.

What do you think the 14FF/16FF process will bring? Currently its nothing but a huge disappointment on the ARM side with A9 series and Exynos. A9X being 13-16% faster than A8X for example.

It seems to me you just keep building the tower higher. Carrizo was also to be amazing, that one failed flat.

 

[AtenRa]

What do you think the 14FF/16FF process will bring? Currently its nothing but a huge disappointment on the ARM side with A9 series and Exynos. A9X being 13-16% faster than A8X for example.

14nm LPE is not the same as 14 LPP and going from 28nm gate first planar to 14nm Gate Last FF is a tremendous change for density, performance and power reduction.

It seems to me you just keep building the tower higher. Carrizo was also to be amazing, that one failed flat.

Carizzo performance is exceptional for the architecture and the process used.

 

[Fjodor2001]

What do you think the 14FF/16FF process will bring? Currently its nothing but a huge disappointment on the ARM side with A9 series and Exynos. A9X being 13-16% faster than A8X for example.

40% extra IPC + 16% on top of that for the process tech is quite good. Also, A8X was on 20 nm. But AMD's current CPUs are on 28 nm, so the jump to 14/16 nm will be greater for AMD.

Also, do you think Intel's 14 nm is a huge disappointment too? Because it has not added much more performance either over 22 nm.

 

[Yuriman]

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%.

I agree, it doesn't look good. Kaveri overclocked has a 24% clockspeed advantage and still needs another 33% on top of that to compete. This works out to Haswell having 65% higher IPC.

Comparing stock Kaveri to the 4330, Kaveri still has a 14% clockspeed advantage, and still needs another 45%. This works out to Haswell having 66% higher IPC.

To get another point of reference, I clocked my Ivy to 4ghz (same as single-core turbo on Kaveri) and ran CB ST. My result was 142, which suggests that Ivy Bridge has ~51% better IPC than Kaveri.

A 2500K scores ~127 at stock. Stock Kaveri has an 8% clockspeed advantage, but Sandy Bridge is still 35% ahead. This works out to Sandy Bridge being 46% faster per clock.

EDIT: Let's throw in Skylake too. A stock i5 6600K scores about 164, with a 2.5% clockspeed disadvantage. That puts Skylake at 79% faster perclock than Kaveri.

~

If we trust AMD's figures that Excavator has 4-15% better IPC than Kaveri and do some basic math, optimistically (using 15% and 40%) Zen will be just shy of Haswell (within 5%). Pessimistically (using 4% and 40%), it will tie Sandy Bridge.

EDIT: Optimistically (for AMD), Skylake will still have an 11% IPC advantage. This would be game-changing, if AMD can hit this without major clockspeed losses. Pessimistically, Skylake will still have a 23% IPC advantage, which is potentially problematic. Another generation of Intel's tick-tock and you're nearing the disparity between Piledriver and Ivy Bridge, again.

So, at best, AMD is looking to match Intel's 2013 uarch in IPC.

 

[Fjodor2001]

So, at best, AMD is looking to match Intel's 2013 uarch in IPC.

That sounds much worse than it is, because Intel's desktop CPU performance has not increased very much since 2013 (Haswell).

If AMD can provide an 8 core CPU at Haswell (or even IB) level of performance next year that for sure will be competitive (if priced sensibly of course). Heck they have been selling decent amounts of desktop and server CPUs when the gap to Intel was far greater than that.

Honestly, most people will not notice the performance difference between Haswell and Skylake. So if you can get an AMD CPU at Haswell level of performance cheaper than Skylake, that will sell.

The problem is that the gap to Intel in ST performance has been much greater than this the last few years, and thus AMD has not sold very well. But that may change with Zen.

 

[ShintaiDK]

They better deliver more than 40% and same clocks if this DX12 game is to be taken serious. Moar cores flop again in a game that is supposed to utilize up to 16 cores.

 

[stateofmind]

They better deliver more than 40% and same clocks if this DX12 game is to be taken serious. Moar cores flop again in a game that is supposed to utilize up to 16 cores.

This benchmark mostly shows that Ashes needs optimization badly or that the benchmark is a long way from home, measuring what?
If you get almost the same FPSs with an I3 over very different graphics settings, then something is wrong.

Considerably lower settings -> no change => too much is thrown at the CPU while there is a lot of GPU room left.

It seems more like a very specific benchmark which makes no discounts for slower CPUs, saying "I need at least X CPU power, below that I'm not playing with you".