AMD Ryzen (Summit Ridge) Benchmarks Thread (use new thread)

[sirmo]

Intel actually has problem with feeding the fabs. They are pushing the higher cost hardware (HEDT, mobile, BGA-socket) to keep revenue, to keep the money stream for feeding the fabs. ANYTHING that would help offload that will be extremely helpful. From that perspective there is no difference if they would be making hardware for themselves, or for anyone else. The wafers are sold. The fabs are kept fed with work. Simple as it can be.

Intel knows this.

True. Also with the ever increasing costs of die shrinks, they can use the extra funding to maintain the lead.

 

[krumme]

See my edit. GPUs are much easier to do die harvesting, at least that's why common sense tells me, so the effective yields for GPUs could very well be quite larger (but the same trick can obviously also be done with CPUs).

Your best estimate for zen is 115 usd. For a die aprox the same size as sold in rx470 and 480. The 470 is the harvested part sold at 160 usd for an entire gfx. And the price is at retail level Give your numbers a little check. Better get some extra sleep man. Take a beer, relax and come back

 

[Glo.]

True. Also with the ever increasing costs of die shrinks, they can use the extra funding to maintain the lead.

Precisely. And the best effect of this: we get better hardware.

 

[CatMerc]

Okay, I see what's the problem.

If you look at my "very optimistic" calculation, it also has a 0.01 defect density, but the yield is less than 80%.

But if I do the calculation where I set N=1, I get:

1 / (1 + 0.01*2.2) = 97.8.

So this site takes N=1, while the book I used (https://www.amazon.com/Computer-Architecture-Fifth-Quantitative-Approach/dp/012383872X), which seems fairly reliable, says you have to do the following calculation:

1 / (1 + 0.01*2.2)^N

Where N is a number between 11.5 and 15.5. So there you go... but we don't know what N really is, but probably not very close to 1.

Seems like an odd arbitrary thing to add to the calculation. One would think the defect density would already account for the complexity of a node.

 

[Glo.]

Guys, if you want to know rumors about die cost for Zen, Fottemberg have provided us with a rumor:

http://semiaccurate.com/forums/showpost.php?p=277336&postcount=4361

According to him, Zen die costs AMD 18-19$ to make.

 

[Nothingness]

There's this: https://www.reddit.com/r/Amd/commen...ks/db8hpfe/?context=3&st=iwur5a96&sh=ad479dc3

Thanks, I'll give it a try, stay tuned

 

[witeken]

Your best estimate for zen is 115 usd. For a die aprox the same size as sold in rx470 and 480. The 470 is the harvested part sold at 160 usd for an entire gfx. And the price is at retail Give your numbers a little check. Better get some extra sleep man. Take a beer, relax and come back

Well, there a few possibilities.

1. My calculation gave yield as being 30%. Die harvesting might push it to maybe 40, 50, 60, or 70% (I have no idea), which would decrease the chip price substantially.
2. The defect density might be lower. Again no way to verify this. I took the range provided for a 40nm process, but given that 14nm is much more complex, I figured it's probably in the higher end of the range.
3. The mystical N parameter is different.
4. Global Foundries might take a lower cut. If they are happy with 40% margins, that would also reduce the cost AMD has to pay.
5. Low-end GPUs might have lower margins for AMD.

So my point, if you read my conclusion in that post, was that there is a wide range of costs that are possible based on some "reasonable" assumptions. My own feeling is that the cost is probably somewhere between my neutral and optimistic calculation, but that is a very wide range.

So that's why I did not make any strong conclusions..

 

[CatMerc]

Thanks, I'll give it a try, stay tuned

Here is the already clipped file: https://drive.google.com/file/d/0B0uKOxJkKhQkcmlnaWUzU05QRlk/view?usp=sharing

 

[CatMerc]

Well, there a few possibilities.

1. My calculation gave yield as being 30%. Die harvesting might push it to maybe 40, 50, 60, or 70% (I have no idea), which would decrease the chip price substantially.
2. The defect density might be lower. Again no way to verify this. I took the range provided for a 40nm process, but given that 14nm is much more complex, I figured it's probably in the higher end of the range.
3. The mystical N parameter is different.
4. Global Foundries might take a lower cut. If they are happy with 40% margins, that would also reduce the cost AMD has to pay.
5. Low-end GPUs might have lower margins for AMD.

So my point, if you read my conclusion in that post, was that there is a wide range of costs that are possible based on some "reasonable" assumptions. My own feeling is that the cost is probably somewhere between my neutral and optimistic calculation, but that is a very wide range.

So that's why I did not make any strong conclusions..

2. That defect defnsity number was apparently several months old by the time I found it, and it has already been a month or so since. So it could very well be lower by this point.
3. I still have no idea what's the point of this one.
4. The 6000$ price was for AMD, so after GloFo cut

EDIT: I looked around a bit more and found this: http://semiengineering.com/tsmc-tech-tour-de-force/

Considering how mature 28nm was by this point, and how new 16nm was, 14nm's defect density is probably around 0.08-0.1 or something of the sort.

 

[witeken]

Seems like an odd arbitrary thing to add to the calculation. One would think the defect density would already account for the complexity of a node.

This is what the book says:

A simple model of IC yield, which assumes that defects are randomly distributed over the wafer and that yield is inversely proportional to the fabrication process, leads to the following:

*Formula*

This Bose-Einstein formula is an empirical model developed by looking at the yield of many manufacturing lines [Sydow 2006]. (...) Defects per unit area is a measure of the random manufacturing defects that occur. In 2010, the value was typically 0.1 to 0.3 defects per square inch, or 0.016 to 0.057 defects per square centimeter, for a 40nm process, as it depends on the maturity of the process (recall the learning curve, mentioned earlier). Finally, N is a parameter called the process-complexity factor, a measure of the manufacturing difficulty. For 40nm processes in 2010, N ranged from 11.5 to 15.5.

A paragraph further they talk about redundancy to improve the yield.

So I'd say this N value is there because it's an empirical model. It's not like an equation that you can derive from somewhere.

 

[witeken]

2. That defect defnsity number was apparently several months old by the time I found it, and it has already been a month or so since. So it could very well be lower by this point.
3. I still have no idea what's the point of this one.
4. The 6000$ price was for AMD, so after GloFo cut

EDIT: I looked around a bit more and found this: http://semiengineering.com/tsmc-tech-tour-de-force/

Considering how mature 28nm was by this point, and how new 16nm was, 14nm's defect density is probably around 0.08-0.1 or something of the sort.

Great find. Then the defect densities are probably closer to 0.01. Interesting, but of course we have no way to know how the N parameter has evolved.

4. I took $4800 as wafer price, but not as price AMD pays. $6000 as the price that AMD pays seems too optimistic to me for a FinFET node. So source please.

 

[CatMerc]

Great find. Then the defect densities are probably closer to 0.01. Interesting, but of course we have no way to know how the N parameter has evolved.

4. I took $4800 as wafer price, but not as price AMD pays. $6000 as the price that AMD pays seems too optimistic to me for a FinFET node. So source please.

Sadly the source is my head. I have a ton of tidbits of information like that that I gathered from various forums from people I deem to have good sources (Fottemberg, The Stilt, etc'), but I don't keep track of where I got where.
If I had to guess the 6000$ price tag was probably from Fottemberg, but I'll have to go and look around for it.

If I find anything I'll let you know.

Edit: Well, while I continue my search, here's something to chew on. Fott mentinoned a Zen die costing about 19$ for AMD. So assuming 100% yield (yeah yeah), it's around 5K dollars. 4K with 80% yield. 6k wafer price being an old number from Polaris era doesn't sound too absurd with this in mind.

 

[prtskg]

If turbo isn't ready, it's hard to see zen (I like this name better) would be available in January. What do you people think?

 

[AtenRa]

Really guys, when AMD can sell a Polaris 10 (232mm2) dGPU with PCB + 4GB of Memory + larger Cooler at $199, even if ZEN die size is the same as Polaris 10, then selling a 4C 8T at $199 will have higher Margins and make more profit than any RX 480.

Edit: And no GPUs are not better at harvesting, AMD sells CPU SKUs using half dies (dual cores vs 4 Cores or 4 Threads vs 8 Thread Bulldozer die etc). Neither AMD nor NVIDIA have ever sold a half die GPU as far as i can remember.

 

[Atari2600]

That is 100% what I am suggesting.

and I think you are completely wrong in your suggestion.

AMD need Zen to:
1. Allow them to reduce their leverage which has hung over them since 2006.
2. Allow them to have 2 or 3 solid iterations for continued profit and CPU growth over the next 3-5 years.
3. Allow them to have a solid x86 base for forays into HPC APUs and custom silicon/FPGAs based on x86 baselines.


If Zen is as good as is hoped, then it demonstrates small teams well led by technically competent folks rather than the powerpoint mafia with their MBA can compete with behemoths that are inefficient due to their size.

AMD could stand up quite a few more small teams of the same ilk based on Zen profits without having to go near banks. Bigger is not always better and more money is often not the solution.

And you would be wrong.

Not really.
The amount of consumers shelling out for a HEDT compared to the overall market size is miniscule. The subset of those that don't know what they are looking for and don't compare across genuine competitors is vanishingly small.

And they will look at the price and go AMD in that case.

No. They will look at performance and go with whoever offers them best performance or best performance/watt.

This comment gives me insight into your understanding, or lack of, of the market.

You are free to think that. This is basic finance here. Its why companies are willing to sell long term debt for cash now. It allows you to take the cash now, and do even more. That results in a net gain over the long run. Not sure what else to say.

AMD are still paying up for the ATi decision. That monkey needs to be removed from their back.

I'm not suggesting going to zero-loans, or anything like it, but the bonds need to be brought under control. If Zen flops, the company probably dies due to the pending bond payment - that is not a healthy situation.

Never said it was. In my opinion, market share is very important. AMD should be able to make large profits and still undercut Intel to gain market share.

Which is what everyone else is suggesting. But not undercut them by half, which would lead to AMD not maximizing their profit - both in 2017 and over 2017-2020. That error could, and likely would, impact their long term prospects. Which is very bad for consumers and commercial alike.

 

[Nothingness]

Here is the already clipped file: https://drive.google.com/file/d/0B0uKOxJkKhQkcmlnaWUzU05QRlk/view?usp=sharing

Here we go:

Performance counter stats for 'ghb':

  1130485,461040  task-clock (msec)  #  3,393 CPUs utilized  
  483 439  context-switches  #  0,428 K/sec  
  48 996  cpu-migrations  #  0,043 K/sec  
  400 587  page-faults  #  0,354 K/sec  
2 599 208 890 203  cycles  #  2,299 GHz  
  <not supported>  stalled-cycles-frontend 
  <not supported>  stalled-cycles-backend  
3 523 266 638 003  instructions  #  1,36  insns per cycle  
  416 464 420 076  branches  #  368,394 M/sec  
  6 172 858 731  branch-misses  #  1,48% of all branches  

  333,149310230 seconds time elapsed

That's on a Haswell 4600U.

Branch density is low at 11.8%. Branch prediction is good. This means the branch MPKI is low at 1.75.

That's about the same MPKI as found on blender at RWT.

So both blender and handbrake as shown by AMD have low branch density and low branch misprediction per thousand instructions.

 

[SpaceBeer]

Really guys, when AMD can sell a Polaris 10 (232mm2) dGPU with PCB + 4GB of Memory + larger Cooler at $199, even if ZEN die size is the same as Polaris 10, then selling a 4C 8T at $199 will have higher Margins and make more profit than any RX 480.

Don't forget ~$10 cost for CPU cooler and packaging Another example is A8-7600 (245mm^2) sold for $75. Yes, it's much cheaper for production, but even if 14nm FF is 2x more expensive, AMD would probably earn more selling Ryzen for $150 than RX 470 for $180

 

[blublub]

Here we go:

Performance counter stats for 'ghb':

  1130485,461040  task-clock (msec)  #  3,393 CPUs utilized  
  483 439  context-switches  #  0,428 K/sec  
  48 996  cpu-migrations  #  0,043 K/sec  
  400 587  page-faults  #  0,354 K/sec  
2 599 208 890 203  cycles  #  2,299 GHz  
  <not supported>  stalled-cycles-frontend 
  <not supported>  stalled-cycles-backend  
3 523 266 638 003  instructions  #  1,36  insns per cycle  
  416 464 420 076  branches  #  368,394 M/sec  
  6 172 858 731  branch-misses  #  1,48% of all branches  

  333,149310230 seconds time elapsed

That's on a Haswell 4600U.

Branch density is low at 11.8%. Branch prediction is good. This means the branch MPKI is low at 1.75.

That's about the same MPKI as found on blender at RWT.

So both blender and handbrake as shown by AMD have low branch density and low branch misprediction per thousand instructions.

Nice

This leads to two assumptions:

1) AMD us hiding Ryzen's poor branch prediction

2) AMD isn't showing best case scenarios with blender and handbrake and is keeping branch prediction a secret because it is very good and chose low branch prediction benchmarks deliberately.

Since Su was proud about Ryzen's neural-net branch prediction option 2 isn't at all unlikely

 

[CatMerc]

Here we go:

Performance counter stats for 'ghb':

  1130485,461040  task-clock (msec)  #  3,393 CPUs utilized 
  483 439  context-switches  #  0,428 K/sec 
  48 996  cpu-migrations  #  0,043 K/sec 
  400 587  page-faults  #  0,354 K/sec 
2 599 208 890 203  cycles  #  2,299 GHz 
  <not supported>  stalled-cycles-frontend
  <not supported>  stalled-cycles-backend 
3 523 266 638 003  instructions  #  1,36  insns per cycle 
  416 464 420 076  branches  #  368,394 M/sec 
  6 172 858 731  branch-misses  #  1,48% of all branches 

  333,149310230 seconds time elapsed

That's on a Haswell 4600U.

Branch density is low at 11.8%. Branch prediction is good. This means the branch MPKI is low at 1.75.

That's about the same MPKI as found on blender at RWT.

So both blender and handbrake as shown by AMD have low branch density and low branch misprediction per thousand instructions.

Interesting stuff, two things:
A. What did you use to get these numbers?
B. If I understand you correctly, this means that both the Blender and Handbrake scenarios are situations where an inferior branch predictor could still do good vs Intel?

 

[AtenRa]

AMD would probably earn more selling Ryzen for $150 than RX 470 for $180

Most definitely, lets assume same dies between ZEN and Polaris 10 at the same GloFo 14nm LPP process and its immediately apparent that selling a 4C 8T ZEN SKU at the same price as Polaris 10 RX 470 will earn AMD more profit (Higher Margins) simple because BOM is higher for RX 470.

 

[Nothingness]

Interesting stuff, two things:
A. What did you use to get these numbers?

Linux perf command. That uses performance counters as Intel VTune and AMD CodeAnalyst.

B. If I understand you correctly, this means that both the Blender and Handbrake scenarios are situations where an inferior branch predictor could still do good vs Intel?

1) AMD us hiding Ryzen's poor branch prediction

2) AMD isn't showing best case scenarios with blender and handbrake and is keeping branch prediction a secret because it is very good and chose low branch prediction benchmarks deliberately.

Since Su was proud about Ryzen's neural-net branch prediction option 2 isn't at all unlikely

It's hard to draw any definitive conclusion except for the fact both benchmarks have low branch density and MPKI, so it's likely that even a moderately good branch predictor would perform well. The rest is speculation

 

[BigDaveX]

Edit: And no GPUs are not better at harvesting, AMD sells CPU SKUs using half dies (dual cores vs 4 Cores or 4 Threads vs 8 Thread Bulldozer die etc). Neither AMD nor NVIDIA have ever sold a half die GPU as far as i can remember.

Technically they have, but you'd have to go back the better part of a decade to find them. The reason GPUs tend to harvest better right now is because they're organized in such a way that you can disable something like 1/16th of the chip and still have the rest functional, whereas with most desktop CPUs your choices boil down to quad-core, tri-core and dual-core.

 

[CatMerc]

Linux perf command. That uses performance counters as Intel VTune and AMD CodeAnalyst.



It's hard to draw any definitive conclusion except for the fact both benchmarks have low branch density and MPKI, so it's likely that even a moderately good branch predictor would perform well. The rest is speculation

Well, I didn't draw a conclusion, I just asked if it COULD hide a poor branch predictor. But thanks for answering

 

[AtenRa]

Technically they have, but you'd have to go back the better part of a decade to find them. The reason GPUs tend to harvest better right now is because they're organized in such a way that you can disable something like 1/16th of the chip and still have the rest functional, whereas with most desktop CPUs your choices boil down to quad-core, tri-core and dual-core.

Well the thing is that today AMD can use half the ZEN die and make a huge profit by selling 4C 8T SKU at $199, they cannot do the same with Polaris 10 GPU because BOM is way higher than selling a CPU alone.

For example, if you use half the Polaris 10 die you end up close to Polaris 11 performance territory and because of the higher BOM (memory + PCB) you cannot sell it at the same price as RX 460. You cannot sell it higher because your competition has higher perf product at same price.

Also, the same 200-220mm2 ZEN die will scale from $199 for the 4C 8T SKU and up to $400-600 for the 8C 16T, something no GPU die at the same size can do.

 

[Nothingness]

Well, I didn't draw a conclusion, I just asked if it COULD hide a poor branch predictor. But thanks for answering

Oh I didn't mean you drew a conclusion, and in fact I basically agree with what you wrote