New Zen microarchitecture details

[DeletedMember377562]

Is there a link to this new 55%> excavator claim? ( or some sort of citation at least)

http://cpugrade.com/amd-ryzen-revisited/

Granted, the tests are mostly the ones AMD showcased. Which are to be best-case. And they also don't show single core that much. But what they do show is pretty interesting: 4% better than Skylake. I, personally, believe right below Skylake, or matching it.

That being said, the huge number of motherboard manufacturers ready at launch in the stores, shows a kind of optimism for AMD products I haven't seen for a long while. Not to mention that the recent rumour from Canard of Intel shaking in fear, points toward AMD getting pretty good IPC in their architecture. Too bad that GloFo's bad yields will hinder them...

So what AMD make do in IPC, Intel will make up by superior clock speeds. A reversal of roles, from how we traditionally know these two companies.

 

[CentroX]

Will ryzen be delayed now that glofo has yield problems?

 

[Doom2pro]

Will ryzen be delayed now that glofo has yield problems?

No confirmation of yield issues, only assumptions based on lower clocked 4 core parts... Ryzen is no doubt in full ramp at Malta in NY (and perhaps in Samsung's fabs as well).

It was supposed to be released at the end of 2016, but apparently there were problems with the board makers and BIOS debugging issues (nothing wrong with the CPUs themselves).

So basically this has given AMD time to possibly create another spin (unconfirmed rumor) that clocked higher than expected while they were already mass producing the older stepping, and what this means is a few things, one there will be more stock at launch now due to the delay (they didn't just stop production because motherboards weren't ready) and perhaps (if the respin rumor is true) there will be two steppings at launch, one of them will be clocking higher and the older lower clocking stepping is possibly one of the reasons for the lower clocked 4 core SKUs.

 

[DeletedMember377562]

I hoper you are right. If AMD can get their chips to around 4,2 GHz, they will launch one hell of a product.

 

[majord]

http://cpugrade.com/amd-ryzen-revisited/

Granted, the tests are mostly the ones AMD showcased. Which are to be best-case. And they also don't show single core that much. But what they do show is pretty interesting: 4% better than Skylake. I, personally, believe right below Skylake, or matching it.

That being said, the huge number of motherboard manufacturers ready at launch in the stores, shows a kind of optimism for AMD products I haven't seen for a long while. Not to mention that the recent rumour from Canard of Intel shaking in fear, points toward AMD getting pretty good IPC in their architecture. Too bad that GloFo's bad yields will hinder them...

So what AMD make do in IPC, Intel will make up by superior clock speeds. A reversal of roles, from how we traditionally know these two companies.

Thank you. It's nice work but Id go one further though and say there's zero single thread / per core ipc calculations. It's all guesswork, related to throughput .

Have AMD actually made any such adjustment to there 40% claim? , I know Lisa said they met or exceeded, but that is all. This 55% business probably needs to die if that's the only source tbh. as the likelyhood of exceeding a design goal by that much is, well very small.

I'm personally keep standing by the notion that SMT yield (%) will be higher than broadwell, pushing its throughput /clock very close (sub 10%) on average (and not exceed on average) . I know there's strong disagreement on this, but we'll see soon enough. If i'm wrong, well and good, but i've seen zero evidence of the contrary, and in the meantime i'll at least do my 5c worth in trying to keep the hype train from running away, and eventually derailing.

 

[richaron]

Yeah the 55% improvement talk is hopeful thinking by some, and ammunition being prepared by others waiting for another "OMG FailZen!!@1" moment.

As we've known for years: Baseline expectations are around Sandy IPC, and Lisa said they met and exceeded these expectations (the previous official 40% claims). The other info we have, likely best case scenarios, gives us IPC around Broadwell levels (the source of our new 55%).

Knowing IPC is dependent on workload, I'm sure people understand it could have gone up more or less in different situations. And these percentages people are talking about are some sort of average. But expecting Zen to have 55% improvement over Excavator across the board is silly.

 

[Dannotech]

Yeah the 55% improvement talk is hopeful thinking by some, and ammunition being prepared by others waiting for another "OMG FailZen!!@1" moment.

As we've known for years: Baseline expectations are around Sandy IPC, and Lisa said they met and exceeded these expectations (the previous official 40% claims). The other info we have, likely best case scenarios, gives us IPC around Broadwell levels (the source of our new 55%).

Knowing IPC is dependent on workload, I'm sure people understand it could have gone up more or less in different situations. And these percentages people are talking about are some sort of average. But expecting Zen to have 55% improvement over Excavator across the board is silly.

A couple of things; 1) The Zen core has 100% more IPC potential than the Excavator core, as can be seen in the block diagram of the Zen core which AMD released quite a long time ago. 2) I built an Excavator based PC and I ran the Blender benchmark with the Ryzen graphic that AMD provided on their website. If you recall, Zen rendered the image in 35.1 seconds. The Athlon X4 845 that I used rendered the image in 4 minutes and 2 seconds. So even if you quadrupled the number of cores in Excavator so that it was 16 Excavator cores vs. 16 Zen threads (only 8 Zen cores), It still would have taken Excavator over a minute to render the image.

So a 55% improvement in IPC seems like quite the understatement, if you ask me.

 

[Mechanical Man]

I'm pretty sure @The Stilt has ryzen and he has been talking about 55% earlier. I would expect it to be somewhere there. I'm also led to believe that will still be little under Broadwell ipc. So ryzen wont catch kabylake in any means. But will offer greatly competing offering (also in many games), specially by offering more cores than intel will for some time.

 

[Abwx]

In a separate paper, AMD said its upcoming Zen x86 core fits into a 10 percent smaller die area than Intel’s currently shipping second-generation 14nm processor. Analysts and even Intel engineers in the session said the Zen core is clearly competitive though many confidential variables will determine whether the die advantage translates into lower cost for AMD.

http://www.eetimes.com/document.asp?doc_id=1331317&page_number=2

 

[richaron]

A couple of things; 1) The Zen core has 100% more IPC potential than the Excavator core, as can be seen in the block diagram of the Zen core which AMD released quite a long time ago. 2) I built an Excavator based PC and I ran the Blender benchmark with the Ryzen graphic that AMD provided on their website. If you recall, Zen rendered the image in 35.1 seconds. The Athlon X4 845 that I used rendered the image in 4 minutes and 2 seconds. So even if you quadrupled the number of cores in Excavator so that it was 16 Excavator cores vs. 16 Zen threads (only 8 Zen cores), It still would have taken Excavator over a minute to render the image.

So a 55% improvement in IPC seems like quite the understatement, if you ask me.

I just ran that Blender file @ 150 samples on an a8-7600 (Kaveri 3.1/3.8) and got 2:05:97 - about 126 seconds. Granted I am running Ubuntu and I didn't check to match versions, but I think your results are way off.

Edit: or my results are

 

[tamz_msc]

Also from the EE Times report:

The company now has two eight-core designs running with simultaneous multithreading at 3.4 GHz.

How does that translate into actual number of retail 8c/16t SKUs?

 

[inf64]

Impressive indeed.

Also from the EE Times report:

How does that translate into actual number of retail 8c/16t SKUs?

Two 8C SMT designs manufactured by 2 different companies maybe?

 

[Nothingness]

Yep. He expects Sandy Bridge level int performance and not much on fp.

And the 40% referred to int_rate... A Throughput bench. No different than what I said they might have done, looking at historical data.

"AMD has disclosed little useful performance data, but it provided guidance that Zen’s IPC is 40% better on SPECint_rate2006 than Excavator’s,"

David Kanter is quite knowledgeable, but alas he also has a bias toward Intel. So I'd take any comment he makes about something that competes against Intel with a grain of salt.

 

[imported_jjj]

So 44mm2 for the quad cluster puts the 8 cores die at 180-186mm2 if the old leaked die shot is real.
Would be in line with expectations but only half the die for the cores with just 2 memory chans and 32PCIe is not that efficient.

 

[dfk7677]

I just ran that Blender file @ 150 samples on an a8-7600 (Kaveri 3.1/3.8) and got 2:05:97 - about 126 seconds. Granted I am running Ubuntu and I didn't check to match versions, but I think your results are way off.

Edit: or my results are

Yours are. I had the same 'problem'. Ubuntu version is much faster than Windows version.

 

[The Stilt]

I'm pretty sure @The Stilt has ryzen and he has been talking about 55% earlier. I would expect it to be somewhere there. I'm also led to believe that will still be little under Broadwell ipc. So ryzen wont catch kabylake in any means. But will offer greatly competing offering (also in many games), specially by offering more cores than intel will for some time.

Let's get this straight
I didn't say 55% based on any actual measurement (obviously), but based on the information released by AMD themselves and the leak from CPC.

Sorry, but I don't understand what you mean.

If Zen delivers 40% more IPC than Excavator, then it matches Sandy / Ivy Bridge. That hasn't changed and won't ever change.
However if it infact matches Broadwell-E's IPC, then the actual increase over Excavator is ~55% and not 40% like AMD stated.

Of course not, if the performance demonstrated in the leaks and the results AMD has demonstrated are actually constant and true.
However in this case, the performance hasn't increased by 40% but around 55%.

AMD has officially only stated 40% and nothing else.

 

[itsmydamnation]

So 44mm2 for the quad cluster puts the 8 cores die at 180-186mm2 if the old leaked die shot is real.
Would be in line with expectations but only half the die for the cores with just 2 memory chans and 32PCIe is not that efficient.

What you ignoring is the massive complex in the bottom left that is their for nothing other then connecting other Zepplins SOC's together. AMD has traded making 3 chips that intel do (D,E,EP) for 1 at the cost of that GMI interconnect space.

 

[Mechanical Man]

Let's get this straight
I didn't say 55% based on any actual measurement (obviously), but based on the information released by AMD themselves and the leak from CPC.

Yes ofc. I did not claim u did, maybe i was not clear about that thought.

But you have not changed your position either (on the 55%).

 

[KTE]

He also expects 3.2GHz which is at least 12.5% lower than what we know is true and Zen's IPC has been upgraded to 55%, which is 15% more than expected. You can do the maths on that.

Base, yes. All Turbo he lists 3.7GHz.

Clocks can improve when delays of months occur... IPC doesn't.

But IPC is what AMD has told him, not a random conjecture like us. And to be honest, no further figure has been confirmed.

+40% on throughput will not be Skylake/Broadwell competitive at the slightest, and will need high clocks to compete with BDe...

We will see shortly if it's int/fp or just int_rate this figure applies to. If Ryzen is any good, AMD will publish industry SPEC figures to show.

Sent from HTC 10
(Opinions are own)

 

[itsmydamnation]

But IPC is what AMD has told him, not a random conjecture like us. And to be honest, no further figure has been confirmed.

Now your just throwing conjecture to match your agenda, in no instance has David kanter said AMD have told him any IPC numbers specifically ..... given AMD have said much later on that they exceeded 40% and the CPC benchmarks align as well as the AMD shown benchmarks you need to rethink your position.

 

[imported_jjj]

What you ignoring is the massive complex in the bottom left that is their for nothing other then connecting other Zepplins SOC's together. AMD has traded making 3 chips that intel do (D,E,EP) for 1 at the cost of that GMI interconnect space.

I am not ignoring anything just not naming all the items in the SoC in the previous post and ,no matter what, using less than half the die for cores+ cache and less than 25% for cores+L1 would be less than ideal. Kinda negates the benefit of having a nice and compact quad cluster.

 

[itsmydamnation]

I am not ignoring anything just not naming all the items in the SoC in the previous post and ,no matter what, using less than half the die for cores+ cache and less than 25% for cores+L1 would be less than ideal. Kinda negates the benefit of having a nice and compact quad cluster.

Except all that other stuff combines into a massive amount of I/O in the Server SKU's and that is going to be very important for AMD and their overall GP numbers. But really we are taking about a few dollars at most per chip in terms of manufacturing. The fact the SOC is 160mm or 200mm isn't really a big deal especially given the price of P10 which is 232mm has A PCB, VRM's , 8GB of GDDR and AIB margin sell for $150 to 260.

If we assume that the product mix results in around 32-5% GP selling Zen at that price (no im not saying they will) will deliver maybe GP around 50%? Maybe even higher all depends on the mix.

All the AMD will price like inteler's are ignoring what AMD's GP will look like if they target 200-300 hard because its inconvenient.

 

[imported_jjj]

@KTE

That article is from August after Hot Chips, based on the info disclosed there. We have more info today than what he had back then.

@ itsmydamnation
Sure cost per mm2 might not be huge compared to retail ASPs but it's the principle while the few $ in OEM matter quite a bit. A 140mm2 8 cores die would have been more fun.
And then there are the 4 cores, in theory those could retail as low 50$ and a few $ make a big difference. Not saying they will sell at 50$, this year at least, but they could if they had a native die.
Hopefully they'll have more resources in the next years.

 

[The Stilt]

Yes ofc. I did not claim u did, maybe i was not clear about that thought.

But you have not changed your position either (on the 55%).

True.
However I wouldn't say even if the actual improvement I allegedly measured was 31% or 125%.

 

[itsmydamnation]

True.
However I wouldn't say even if the actual improvement I allegedly measured was 31% or 125%.

well aren't you the party pooper