New Zen microarchitecture details

[Dresdenboy]

Difference between hardware and software for smt?

Do you refer to "HSW"? That means "Haswell".

 

[The Stilt]

Where did you pull those ipc numbers?

Some tractor fp test from the k7 days?

Cinebench R10, R11.5, R15 and 3DPM (all used by Anandtech) for example result in 61.8% higher IPC for Broadwell-E. Add a single AVX2 workload to that and the average shoots beyond 70% (e.g. in HEVC BDW/SKL have > 100% higher IPC).

I used the results from Anandtech article (i7-6800K).
Excavator tested at 3.4GHz, score normalized to 3.8GHz with linear (1.117647x scale).

Cinebench R10 ST = 4426 vs. 7425 >> 67.75%
Cinebench R11.5 ST = 1.12 vs. 1.73 >> 54.46%
Cinebench R15 ST = 96 vs. 150 >> 56.25%
3DPM ST = 71.945 vs. 121.38 >> 68.71%

Average: 61.8%

Original scores for 3.4GHz (XV):

Cinebench R10 ST = 3960
Cinebench R11.5 ST = 1.00
Cinebench R15 ST = 86
3DPM ST = 64.3749

 

[The Stilt]

Is it poll time again? My take (what's at least somewhat likely):

• HSW IPC for integer, Ivy IPC for FP
• 3.0 GHz / 4.0 GHz :sneaky:
• SMT is at HSW level (they're working on it internally since Thunderbird according to my sources!)
• $999 for top model

You expect people to pay premium just because it is AMD?

 

[Hitman928]

You expect people to pay premium just because it is AMD?

How much is an 8 core intel?

Honest question, I don't know.

 

[The Stilt]

How much is an 8 core intel?

Honest question, I don't know.

That depends.

6900K BDW-E costs 1080€ in Europe. 5960X costs around 1030€. However since you didn't specify the segment: E5-2630 V3 sells for 630€.

In the very best case (where the 14nm LPP process exceeds my expectations), I expect the fastest Summit Ridge SKU to compete with Intel 6C/12T, rather than 8C/16T models. Regardless if they're Haswell or Broadwell.

6C/12T Broadwell-E's can be had for 445 - 640€ (6800K - 6850K). Meanwhile the older Haswell 5820K goes for ~ 377€.

 

[Dresdenboy]

You expect people to pay premium just because it is AMD?

The sustainability of high boost clocks might help a bit. But it's just speculation anyway.

 

[ozzy702]

Is it poll time again? My take (what's at least somewhat likely):

• HSW IPC for integer, Ivy IPC for FP
• 3.0 GHz / 4.0 GHz :sneaky:
• SMT is at HSW level (they're working on it internally since Thunderbird according to my sources!)
• $999 for top model

AMD can try to charge $999 for their top SKU all they want, it won't sell. Even with the specs above I see them not selling until $500 or less. People with deep pockets will want the best money can buy and that's still going to be Intel. AMD will do well in the sub $350 market but over that they won't sell well. Also remember that most people aren't enthusiasts and won't look at price/performance etc, they just don't. It's the same problem AMD has when competing against NVIDIA.

 

[krumme]

Cinebench R10, R11.5, R15 and 3DPM (all used by Anandtech) for example result in 61.8% higher IPC for Broadwell-E. Add a single AVX2 workload to that and the average shoots beyond 70% (e.g. in HEVC BDW/SKL have > 100% higher IPC).

I used the results from Anandtech article (i7-6800K).
Excavator tested at 3.4GHz, score normalized to 3.8GHz with linear (1.117647x scale).

Cinebench R10 ST = 4426 vs. 7425 >> 67.75%
Cinebench R11.5 ST = 1.12 vs. 1.73 >> 54.46%
Cinebench R15 ST = 96 vs. 150 >> 56.25%
3DPM ST = 71.945 vs. 121.38 >> 68.71%

Average: 61.8%

Original scores for 3.4GHz (XV):

Cinebench R10 ST = 3960
Cinebench R11.5 ST = 1.00
Cinebench R15 ST = 86
3DPM ST = 64.3749

You naughty boy

Its not fair to assess a new member of the family by a fp heavy suite. Purely fp suite and more of the same.

Not everyone on this planet is a professor in physics and makes scientific calculations 24h. The rest of us - And by that i mean topplayers like google amazon facebook and me. We dont want to pay 500 euro for that fp tractor. It can go fp farming where its needed.

 

[The Stilt]

You naughty boy

Its not fair to assess a new member of the family by a fp heavy suite. Purely fp suite and more of the same.

Not everyone on this planet is a professor in physics and makes scientific calculations 24h. The rest of us - And by that i mean topplayers like google amazon facebook and me. We dont want to pay 500 euro for that fp tractor. It can go fp farming where its needed.

In that case you can be relatively happy with the existing 15h family products. Nothing too badly wrong with their integer performance. FP is where they suffer and that's what needs to be remedied. That's the very reason why I chose FP workloads. Unless AMD cures the abysmal FP performance with Zen, they are as good as dead.

 

[KTE]

Is it poll time again? My take (what's at least somewhat likely):

• HSW IPC for integer, Ivy IPC for FP
• 3.0 GHz / 4.0 GHz :sneaky:
• SMT is at HSW level (they're working on it internally since Thunderbird according to my sources!)
• $999 for top model

My take:

- IPC between IB and Haswell.
- 3.2Ghz base / 3.7Ghz TB for low thread workloads (on top 8C/16T part)
- SMT at HSW level or better
- intel BDW hex core pricing (while performing a bit better)

Between SB to HSW IPC...

Launch 2.8-3GHz base 8C top model.

Sent from my HTC 10 using Tapatalk

 

[krumme]

In that case you can be relatively happy with the existing 15h family products. Nothing too badly wrong with their integer performance. FP is where they suffer and that's what needs to be remedied. That's the very reason why I chose FP workloads. Unless AMD cures the abysmal FP performance with Zen, they are as good as dead.

Doubling of fp units - effectively like double the cores - should bring them a long way wouldnt you say ? plus a less confused frontend.

Considering how bd was laid out the integer perf was what was most disapointing. It was made to be an integer pushing server robot but nothing happened except it nearly catched fire just walking itself. For that it was rightfully slammed. It didnt do what it was intended to do.

A processor needs to be balanced for the market it adresses and assessed for what it tries to accomplish. Zen is not targeting fpu heavy situations as they are a minor part of the market.
Cb and more fpu really heavy stuff is imo not that relevant anymore for the mass market and as zen is not intended for the fpu heavy situation it should not be assessed for it.

As a parallel look where arm a73 went - straight for the big mobile segment slimming down core. Cost and efficiency is of utmost importance and fat fpu parts cost.

 

[Dresdenboy]

AMD can try to charge $999 for their top SKU all they want, it won't sell. Even with the specs above I see them not selling until $500 or less. People with deep pockets will want the best money can buy and that's still going to be Intel. AMD will do well in the sub $350 market but over that they won't sell well. Also remember that most people aren't enthusiasts and won't look at price/performance etc, they just don't. It's the same problem AMD has when competing against NVIDIA.

Let's wait and see how that little bet turns out. ^^

 

[krumme]

Dresdenboy. Tell more about the smt story

 

[el etro]

Is it poll time again? My take (what's at least somewhat likely):

• HSW IPC for integer, Ivy IPC for FP
• 3.0 GHz / 4.0 GHz :sneaky:
• SMT is at HSW level (they're working on it internally since Thunderbird according to my sources!)
• $999 for top model

My take:

Broadwell IPC on integer, IVB IPC on FP;
3.2Ghz base, 4GHz turbo 1C, 3.5Ghz turbo 8C;
SMT better than Nehalem/far better than P4, don't know;
$999 for max bin. 220W TDP version incoming in Q4-17.
Power consumption at 5960x levels.

 

[The Stilt]

3.2Ghz base, 4GHz turbo 1C, 3.5Ghz turbo 8C;

Why would one have separate base and 8C turbo frequencies? :sneaky:

 

[DrMrLordX]

My take:

Broadwell IPC on integer, IVB IPC on FP;

I'm curious as to why you think Zen/Summit Ridge will do so much better in int workloads than in fp? AMD increased fp execution resources per core versus XV modules, but did nothing to increase int execution resources.

The only things I can think of that will help Summit Ridge outperform XV in int workloads will be improvements to the cache subsystem and frontend.

 

[itsmydamnation]

I'm curious as to why you think Zen/Summit Ridge will do so much better in int workloads than in fp? AMD increased fp execution resources per core versus XV modules, but did nothing to increase int execution resources.

The only things I can think of that will help Summit Ridge outperform XV in int workloads will be improvements to the cache subsystem and frontend.

then your not understanding the design very well. i'll explain when i get around to addressing your reply to my last post

 

[JDG1980]

Lets have some betting here. Can some setup a combined freq and ipc test? Perhaps add smt efficiency of some sort and max consumer price?

I guess today for 8c part:
Hw ipc class
2.5 GHz boost 3.2
Smt sucks efficiency class
450 euro tops

https://m.reddit.com/r/Amd/comments/44hrj7/a_theoretical_calculation_about_of_zens_single/

OK, let me throw my hat in the ring. For some time I have predicted IPC close to Sandy Bridge, and that prediction remains unchanged. Why that level? First of all, it's about the minimum required to make a credible and competitive product at this time. Secondly, AMD is claiming 40% IPC improvement over Excavator. We know that the X4 845 (Excavator with 3.8 GHz max turbo speed) does 95 points in single-thread Cinebench R15. Conveniently, the Sandy Bridge i7-2600K also had a 3.8 GHz max turbo. Anandtech's benchmarks have that at 135 points in Cinebench R15 ST. 95 x 1.4 = 133, so that would bring IPC up to about Sandy Bridge levels.

So, what kind of overall performance would that mean? Summit Ridge is rumored to have double the performance of FX-8350. That's aggregate performance, so we're looking at multi-thread Cinebench R15 performance. FX-8350 gets 640 points in that benchmark, so double that would be 1280 points. That puts us right around the Ivy Bridge-based E5-2687W v2 (1297 points). This is an 8C/16T CPU, just like Summit Ridge will be, and if my prediction is right, IPC is also going to be similar. E5-2687W v2 has a base clock of 3.4 GHz, and a max turbo frequency of 4 GHz. Thus, it's likely that targets for Zen's top bin are similar.

At $299, this would be a very competitive chip. (And still profitable - look at $/mm^2 compared to Polaris 10...)

 

[cytg111]

* Cannonlake IPC
* 4.0 GHz / 5.0 GHz Turbo
* SMT is at CMT levels
* $99 for top model
* unicorns reclaim their caveland from north korea

Seriously, anything competetive would be nice.. not holding my breath for glofo.
(but they did just produce polaris, did they not? is things turning around for glofo? isnt that the real question in estimating zens success?)

 

[majord]

Updated test results to include the latest x265 bench which uses AVX much better and results in much better performance on Skylake ( Thanks Stilt for bringing it to my attn!)

Skylake 3M is Still sitting at 60% faster than exv (the updated bench bumped it up from 58%)

So I'm sticking with that for a mixture of Int FP SSE, and AVX2

Even if you forget all the more obscure and legacy stuff , and just focus on commonly used, current benches, throw in 1 AVX2 bench in every 5 - not a lot changes:

62% faster avg for both Cinebench 11.5 and r15
44% faster in Pov ray
60% faster in x265 w/out AVX2
53% faster 3Dmark(13) Physics
60% faster in non AVX2 x265 encoding
107% faster AVX2 x265


63% faster avg

Take AVX2 out completely and it's 55%

and at the end of they that's going to equal

10-15% Syklake ->Zen
7 - 12% Broadwell -> Zen

As for frequencies , I think I said before.. Betting they won't launch below 3Ghz. They'll delay launch before going below. Will see if i'm wrong in time


Spec/ Performance prediction:

3.0Ghz base - 3.6 Turbo

MT performance smack bang in the middle of 6850K and 6900K
ST/ < 8threaded performance 10-15% below all of the BW-E

Price - $499-$599 US

 

[SarahKerrigan]

Why would one have separate base and 8C turbo frequencies? :sneaky:

It's not impossible. Power8 can turbo on all cores, for instance - it's purely thermally limited as far as I know.

 

[MajinCry]

I'm with JDG; minimum expectations is Sandybridge, which is perfect, really. Did something similar, in taking a benchmark (compared Piledriver & Excavator to Nehalem) and scaling up the scores. Worst case was Sandybridge-level IPC, if +40% over Piledriver. Best case was around Haswell, if +40% over Excavator.

Hell, an 8c/16t Sandybridge-level CPU with decent clocks would be a pretty damn fine processor.

 

[deasd]

Updated test results to include the latest x265 bench which uses AVX much better and results in much better performance on Skylake ( Thanks Stilt for bringing it to my attn!)

==img snip==

Skylake 3M is Still sitting at 60% faster than exv (the updated bench bumped it up from 58%)

So I'm sticking with that for a mixture of Int FP SSE, and AVX2

Even if you forget all the more obscure and legacy stuff , and just focus on commonly used, current benches, throw in 1 AVX2 bench in every 5 - not a lot changes:

62% faster avg for both Cinebench 11.5 and r15
44% faster in Pov ray
60% faster in x265 w/out AVX2
53% faster 3Dmark(13) Physics
60% faster in non AVX2 x265 encoding
107% faster AVX2 x265


63% faster avg

Take AVX2 out completely and it's 55%

and at the end of they that's going to equal

10-15% Syklake ->Zen
7 - 12% Broadwell -> Zen

As for frequencies , I think I said before.. Betting they won't launch below 3Ghz. They'll delay launch before going below. Will see if i'm wrong in time


Spec/ Performance prediction:

3.0Ghz base - 3.6 Turbo

MT performance smack bang in the middle of 6850K and 6900K
ST/ < 8threaded performance 10-15% below all of the BW-E

Price - $499-$599 US

When you go from x32 to SSE and even to AVX and resulting performance boost, I'm afraid there's no so-call IPC optimized factor here. If you have wider pipeline and load/store, you could often have higher performance even if your 'IPC' is the same. So your calculation might not accurate.

EDIT:[sarcasm]I would rather believe Zen to be 40% faster than Excavator in Superpi.[/sarcasm]

 

[Zstream]

Do you refer to "HSW"? That means "Haswell".

Yeah, I read that wrong lol. Isn't SMT performance going to be the same for Intel as AMD? I mean, it's just using leftover resources on that core, right? So, if a hardware core is at 70% utilization, SMT can only provide 30% on that logical OS core (considering all things equal and evenly split).

 

[frozentundra123456]

If the only can reach SB IPC, I would consider that a fail. They need at least something between IB and HSW ipc.

8 cores at Sandy Bridge IPC, assuming it can reach the same clocks as BW-E (which is *far* from a given, and highly doubtful in my mind) would be similar to 6 core BW-E, maybe slightly faster in workloads that can use all the cores, slower in everything else. So it is competitive only if the price is right. They really need better than that though to slot closer to 8 core intel, since there is such a huge price gap between intel 6 and 8 core chips. Of course there could be considerably savings on the motherboard compared to intel, so that could make the chip more price competitive.