[Deustche Bank Conference] AMD's New x86 Core is Zen, WIll Launch WIth K12

[PPB]

Guys, dont kid yourselves. Perf/watt pd is only barely competitive to sandy bridge, both on a similar process, and only on high paralellizable workloads. Not even talking abot IB/HW or single threaded performance. The problem with PD perf/watt is the ancient chipset but mostly lazy binning, most people will be able to at least reduce 0.1v on their vishera and still be stable. You can tinker your way to reach some crazy perf/watt numbers with PD via uv/uc.

I think AMD should have focused more on strong cpu perf for their apus until both HBM and HSA were ready. Then go full igp focus to leverage their advantage on. 512gcn cores is way too much for the little bandwith the igp is given to play with

 

[Enigmoid]

Actually 32nm BD CPUs have higher throughput than 32nm Intel CPUs.

A few samples, take a look at FX8350 vs Core i7 3820 and 2600K

http://www.guru3d.com/articles_pages/amd_fx_8350_processor_review,1.html

<SNIP>

Faster than Core i7 3820, higher power usage, understandable.

8350 = 4.0/4.2 (4 ghz on 8 cores)

2600k = 3.4/3.8 (3.5 ghz 4 cores)

There is no technical reason the 2600k doesn't run those clockspeeds, intel could have released an SKU like devil's canyon then.

So either add 14% to the 2600k or take off 12.5% from the 8350 if you want clock for clock.

They end up doing similarily but the 2600k has much greater ST performance and die area scales exponentially with ST performance (for a given arch).

Look at AtenRa's core size picture. The A15 is 6x larger than the A7 but 6 A7s would demolish an A15 in MT. Same with Apple's Cyclone which has 2x IPC compared to A15 but is more than 2x larger.

 

[Essence_of_War]

Actually 32nm BD CPUs have higher throughput than 32nm Intel CPUs.

A few samples
...

Of highly parallel, integer-only, workloads. You can't claim better throughput based solely on the best-possible case workloads.

What happens when you include the cinebench single-thread?

What happens when you use other sources that don't give such rosy results in cinebench multi (10 or 11!)?

What happens in something like 3D particle movement? That is 20W Atom, and PIIx6, both ahead of the 8350. The 8350 is keeping excellent pace with the i5s, but that's a FAR cry from favorable comparisons with the i7-3820...

 

[monstercameron]

Of highly parallel, integer-only, workloads. You can't claim better throughput based solely on the best-possible case workloads.

What happens when you include the cinebench single-thread?

What happens when you use other sources that don't give such rosy results in cinebench multi (10 or 11!)?

What happens in something like 3D particle movement? That is 20W Atom, and PIIx6, both ahead of the 8350. The 8350 is keeping excellent pace with the i5s, but that's a FAR cry from favorable comparisons with the i7-3820...

those results show exactly what the chips designers were after, multicore integer workloads. That is where it does well.

Also 3dpm is perfect example of why they want fusion. The author did intend to write a c++ amp version[I contacted him but it seems that port wont come to fruition]. Some tasks are just better sending to off to dedicated accelerators [fpgas, gpus etc] and that makes sense in the server world.

 

[Idontcare]

those results show exactly what the chips designers were after, multicore integer workloads. That is where it does well.

Also 3dpm is perfect example of why they want fusion. The author did intend to write a c++ amp version[I contacted him but it seems that port wont come to fruition]. Some tasks are just better sending to off to dedicated accelerators [fpgas, gpus etc] and that makes sense in the server world.

All of that is the tail wagging the dog type of an approach to the market and the customer though.

AMD may have decided for themselves that they were going to target a minor segment of the server TAM by selectively going after only those customers who were interested in multicore integer workloads, or those customers who were interested in rewriting (or waiting for some other business to do it for them) code to take advantage of fusion, or those customers who didn't mind paying AMD for integer capability while paying some other business for dedicated accelerators to handle the rest of the processing...but that doesn't mean the majority of the market was equally sold on the idea of handing over money to AMD for that privilege.

You don't get to wag the dog if you are the tail, AMD decided they knew what was best for the customer and they were wrong. The customers (vast majority anyways) went with the competitor's solution as it did not drive the requirement that the customer settle for only having decent performance with multithreaded integer workloads while needing to spend more money still on newly recoded software and dedicated hardware accelerators.

In other words, AMD's solution was (and continues to be) only viable in the event that the customer has no other less costly or less complex/risky alternative. And it doesn't matter what business you are in - be it selling CPUs or selling tires - that is a fail strategy when you approach the customer with that mentality.

 

[Abwx]

What happens when you include the cinebench single-thread?

Who is using CB in single thread set apart for benchemarks..?.
Seriously.?..

What happens when you use other sources that don't give such rosy results in cinebench multi (10 or 11!)?

Why not Povray rather than old versions of CB..??.
What happens suddenly..?..The 2700K struggling with the FX8150..?.?.

What happens in something like 3D particle movement? That is 20W Atom, and PIIx6, both ahead of the 8350. The 8350 is keeping excellent pace with the i5s, but that's a FAR cry from favorable comparisons with the i7-3820...

What happen in 3D particle movement is that it run on X87 on all AMD CPUs and use up to SSE3 on intel chips past the pentium 4 wich is also relegated to X87 since it doesnt support SSE3, nice exemple you have taken that this mock of a bench.

You didnt notice that Avoton has 10% better ST IPC than a Kabini in this so called bench while it s trounced by 30% or so on CB and Povray..???

 

[DrMrLordX]

I dont doubt HSA for 3D rendering is so superior its a gamechanger for the market if its used. It will just chew a cpu to pieces.

I just doubt blender is a big enough market with a profit that matters. And if say Blender adopts HSA, what is the chance professional solutions will adopt HSA? - its different segments and you dont move from a professional product to Blender because of HSA.

And then - if HSA is adopted - what is the benefit for AMD vs years and years of starving to get there? Ofcource there is solid benefit if they get there - but damn its a huge risk for the entire company.

Blender isn't a huge market in and of itself, but there are plenty of indies and college-level learners who cut their teeth with free software. When your budget is effectively zero, it's what you use. If people learn their 3D on software that just flat-out runs better with AMD CPUs, that bias will follow them later in life.

And really, even though Blender may not be as strong a package as professional-grade software, major changes in performance would raise some eyebrows.

You don't get to wag the dog if you are the tail, AMD decided they knew what was best for the customer and they were wrong.

I don't think that's entirely true. AMD decided that they could not compete with Intel by continuing to focus on general-purpose x86 CPUs. So, they tried a different path. Ideally, AMD's decision to pair integer-friendly general purpose CPU cores with FP-friendly GPU cores would not force consumers to make many difficult or expensive sacrifices. You buy AMD GPUs, you run HSA/OpenCL 2.0 compliant software, and you win. It's the software side that just hasn't held up yet. It may yet, but time is running short.

If/when HSA becomes a reality for the many, many workloads that could benefit from it, people should be saving money without experiencing many headaches by choosing AMD hardware. There should be lower costs on the hardware side to more than offset any possible expense involved with updating their software to be HSA compliant.

it is literally around the corner, read up on java sumatra and c++ amp. Fusion is the future that is why even intel is doing huma style interface and OCL 2.0 in their upcoming products.

The problem for AMD is that Intel is not asleep at the wheel. SkyLake in 2015 isn't so far away. AMD needed Sumatra and C++ amp in January of this year, not . . . whenever next year.

So, with all that being said . . . what role does Zen play in AMD's future? Is it solely a replacement for Opteron/FX, or will it find its way into APUs as well?

 

[mrmt]

AMD may have decided for themselves that they were going to target a minor segment of the server TAM by selectively going after only those customers who were interested in multicore integer workloads, or those customers who were interested in rewriting (or waiting for some other business to do it for them) code to take advantage of fusion

This. And don't forget that even if AMD had actual customers willing to use its solution, they wouldn't be able to use it because the development tools are MIA.

It seems an engineering failure of epic proportions. They missed the estimates of how complex CMT chips would be, how hard it would be to manufacture them and how good their foundry partner would be, and on top of that, they seem to have been through a similar failure chain with the software support. All we got is the hot steam.

 

[monstercameron]

All of that is the tail wagging the dog type of an approach to the market and the customer though.

AMD may have decided for themselves that they were going to target a minor segment of the server TAM by selectively going after only those customers who were interested in multicore integer workloads, or those customers who were interested in rewriting (or waiting for some other business to do it for them) code to take advantage of fusion, or those customers who didn't mind paying AMD for integer capability while paying some other business for dedicated accelerators to handle the rest of the processing...but that doesn't mean the majority of the market was equally sold on the idea of handing over money to AMD for that privilege.

You don't get to wag the dog if you are the tail, AMD decided they knew what was best for the customer and they were wrong. The customers (vast majority anyways) went with the competitor's solution as it did not drive the requirement that the customer settle for only having decent performance with multithreaded integer workloads while needing to spend more money still on newly recoded software and dedicated hardware accelerators.

In other words, AMD's solution was (and continues to be) only viable in the event that the customer has no other less costly or less complex/risky alternative. And it doesn't matter what business you are in - be it selling CPUs or selling tires - that is a fail strategy when you approach the customer with that mentality.

All I am trying to do is separate the core design form the business. I am not trying to say that BD was a huge money maker for AMD, just that it did hit seemingly most of it's design goals. Whether the market accepted AMDs bet/risk is another story. Also we all know the invisible hand of the free market has a high amount of randomness, one trend at the right time could have given them a better deal[as in cards].

 

[Vesku]

All I am trying to do is separate the core design form the business. I am not trying to say that BD was a huge money maker for AMD, just that it did hit seemingly most of it's design goals. Whether the market accepted AMDs bet/risk is another story. Also we all know the invisible hand of the free market has a high amount of randomness, one trend at the right time could have given them a better deal[as in cards].

Only if one of the goals was to replicate P4s power consumption curve. That decision to design for high clockspeed with an uncertain fab future was a real blunder. Something that it's safe to say the whole semiconductor industry won't be repeating anytime soon.

I really get the feeling most of Bulldozer's key design decisions were made while Intel still just had P4 with no Core 2 on the horizon. For some reason the people in charge of those decisions at AMD didn't quite get that their own approach with the Athlon series was more in step with the chip fab areas of concern and seemingly decided they could show Intel how to make a better P4.

I'd think efficiency and HSA integration would be a top priority with K12 based on when the design was initiated. Even while AMD was downsizing they were advertising for SoC engineers so I'd also expect to see more crammed onto the APU interposer. Reducing platform complexity is a big deal for their potential customers. Eventually I'd expect the none APU part of the PCB to be almost entirely just routing the input output ports to the SoC package. Example: http://en.wikipedia.org/wiki/Cubieboard

 

[Phynaz]

All I am trying to do is separate the core design form the business. I am not trying to say that BD was a huge money maker for AMD, just that it did hit seemingly most of it's design goals. Whether the market accepted AMDs bet/risk is another story. Also we all know the invisible hand of the free market has a high amount of randomness, one trend at the right time could have given them a better deal[as in cards].

The BD core design sucked, resulting in the business position AMD is in now. The two are inseparable.

 

[mrmt]

All I am trying to do is separate the core design form the business. I am not trying to say that BD was a huge money maker for AMD, just that it did hit seemingly most of it's design goals. Whether the market accepted AMDs bet/risk is another story.

If that were the case then we would probably see AMD management release nokia-style statements, something along the lines that they had a very solid piece of engineering in their hands but that they had misread the direction the market would take.

But I don't think this revisionist approach to Bulldozer is defensible. AMD fired a lot of people in the wake of the failure, and you don't fire the engineers that delivered exactly what you wanted just because it was a market failure, you fire the executives giving the general direction. Did AMD fire just the executives? No, they fired both the executives and the engineers. Something happened at the engineering level that prompted the company to eject them.

Another point is that there are workloads in the very market Bulldozer was targeting, servers, that perform better on the old 45nm Opterons than on Bulldozer, especially the first one. To say that Bulldozer wasn't a failure would be to say that the company was ok with 7 years of R&D generating a solution that had worse performance and worse cost structure for the same market, even with a node advantage. I don't think someone can argue that in front of the most amateur of the executive boards, and certainly not in front of AMD board.

I read the "unmitigated failure" statement as an admission of failure at multiple levels, a flawed conception, a flawed engineering execution and a product that went right into the opposite direction of that the market was going. Only a failure of that size would prompt the company into a messy re-org process like AMD went through.

 

[krumme]

Look at the total diesize for BD vs the core size. The design put far more demands on the more uncore including L3 increasing total diesize to give some of the needed performance. There is a reason they move to a more smt approach.

And even the size of cores Altenra showed, understate the problem. Look at the fat 32nm Atom vs. the Jaguar at 28nm with 60% higher ipc and with its super fat FPU. Intel 32nm was not more dense than 40 TSMC, as could be shown if Bobcat was in the picture (originally posted by Hans D Vries at SA).

There is a solid point in that TSMC or GF process is cheaper than Intel (or GF ought to be). Denser but less performant. Its therefore expected that the BD should be relatively bigger, because of lack of some freq/watt. But the end result is nada presence in the server market. If BD could push those integer/watt there would be a very valuable market. And thats not the reality. And its not like AMD lacked presence in the market beforehand.

 

[krumme]

Another similar but less know failure is the current Atom. In 20 years i havnt seen a product wrapped in so many dollar bills. There is probably a so called "design win" for each customer. If its going to be a win, its purely because of a Monopoly situation, and not because of the products objective qualities. Its a major financial failure because the customer simply dont want to pay for it.

Intel can hide it, talk hot air and give us the cost. AMD doesnt have such a luxury. The faster the product can be named a failure loud and publicly, the faster you can clean up and get on. Yet it took to long for AMD to kill their darling.

 

[Abwx]

Some updated old news for the skeptic, courtesy of Boom at SA...

https://semiaccurate.com/forums/showthread.php?t=8244&page=9

 

[ClockHound]

Are you kidding me? With all the cross-marketing potential next year, AMD doesn't have the 'Millennium Falcon' SOC ready to fly?

Oh, man. What a waste of naming schemes.

 

[TurtleCrusher]

The shills-pretending-to-be-enthusiasts are out in full force these days. We know hardly anything about this entirely new architecture but sure, lets thread-crap the hell out of it anyway!

 

[nvgpu]

http://www.kitguru.net/components/g...psys-to-co-design-14nm-10nm-apu-gpu-products/

In exchange, AMD transfers certain IP and engineering resources to Synopsys.

In addition, Synopsys hires approximately 150 AMD IP R&D engineers and gains access to AMD&#8217;s leading interface and foundation IP.

While the move clearly saves AMD money, it makes it weaker in terms of resources, whereas Synopsys becomes stronger.

KitGuru Says: It looks like AMD has just transferred its fundamental IP and 150 R&D engineers to Synopsys in exchange for IP that it is going to use in the next four or five years. While the company did save a lot of money, it lost a lot of engineers and ability to develop certain technologies going forward. Does such business approach make sense? Maybe. But it looks like another form of asset-light strategy announced many years ago.

 

[TurtleCrusher]

http://www.kitguru.net/components/g...psys-to-co-design-14nm-10nm-apu-gpu-products/

^^Case in point. Guy gets an obscure press "release."

 

[Idontcare]

The shills-pretending-to-be-enthusiasts are out in full force these days. We know hardly anything about this entirely new architecture but sure, lets thread-crap the hell out of it anyway!

^^Case in point. Guy gets an obscure press "release."

Sorry, normally I totally get your posts and see where you are coming from but in this instance I don't get it. What are you saying here? (not a rhetorical question, am serious, you probably have a great point, I just can't figure it out based on the way you worded your posts)

http://www.kitguru.net/components/g...psys-to-co-design-14nm-10nm-apu-gpu-products/

This is an interesting development if true.

AMD has no choice but to do things different or they will be squeezed out of the business for simple lack of having enough volume to justify the mask set costs of shrinking their chips as their competitors march towards 10nm and 7nm.

 

[TurtleCrusher]

The guy has no posting history and that article has only been documented on one other forum. Literally, you'd have to be scouring the internet like a mad man or be referred to it by a marketing group. Heck, he made a pit stop between posts on 4chan by saying "IT'S OVER, AMD IS BANKRUPT & FINISHED"

And that's not necessarily a bad move by AMD. They might believe their most talented engineers needed the the know how on 16-10nm and the engineers lost were expendable. To make a sports analogy, this is no different than trading a developing player for a sure-thing role player. It gives AMD a big leg up/head start on practical designs to be made on smaller processes. It's a trade and we'll find out in the coming years if it was the right one.

 

[nvgpu]

I don't post on that site and wow, the fact you're so mad about something that is a fact that was posted by Kitguru just shows what kind of person you are.

 

[krumme]

AMD has no choice but to do things different or they will be squeezed out of the business for simple lack of having enough volume to justify the mask set costs of shrinking their chips as their competitors march towards 10nm and 7nm.

Agree. Its nothing but positive AMD does attack things differently sharing ressources and IP. With their size and lacking profit, they should have done far more of that years ago. Hopefully we will see more of that.

 

[Idontcare]

I don't post on that site and wow, the fact you're so mad about something that is a fact that was posted by Kitguru just shows what kind of person you are.

I don't think its like that, to be honest. Yes you have every right to take it personal but I think this is less personal and more of a general observation than anything else. I could be wrong though.

The guy has no posting history and that article has only been documented on one other forum. Literally, you'd have to be scouring the internet like a mad man or be referred to it by a marketing group. Heck, he made a pit stop between posts on 4chan by saying "IT'S OVER, AMD IS BANKRUPT & FINISHED"

And that's not necessarily a bad move by AMD. They might believe their most talented engineers needed the the know how on 16-10nm and the engineers lost were expendable. To make a sports analogy, this is no different than trading a developing player for a sure-thing role player. It gives AMD a big leg up/head start on practical designs to be made on smaller processes. It's a trade and we'll find out in the coming years if it was the right one.

I only see personal opinion being posted, no agenda or ax to grind. Personally I think folks are doing the blind feeling out an elephant thing here. We each perceive a different outcome and necessity here.

 

[USER8000]

NVM.