New Zen microarchitecture details

[JDG1980]

The problem with good 256bit AVX/AVX2/FMA performance is you need 256bit paths. And that's a TDP killer. Without these paths my 6700K could maybe have been sold as a 55W TDP CPU.

Same reason why AVX512 is server only.

If Zen gets these 256bit paths, then it will be clocked at 3Ghz or below for 8 core parts.

Even Sandy Bridge shared some of the AVX hardware path with SSE. From Anandtech's architecture review:

Sandy Bridge allows 256-bit AVX instructions to borrow 128-bits of the integer SIMD datapath. This minimizes the impact of AVX on the execution die area while enabling twice the FP throughput, you get two 256-bit AVX operations per clock (+ one 256-bit AVX load).

Haswell adjusted this to enable higher peak FLOPs, but it did indeed cause power utilization issues, to the point where Intel had to use a separate, lower "AVX base clock" in addition to the standard base clock rate.

 

[frozentundra123456]

I am not really interested in the deeply technical discussion going on in this thread. All i know is my next cpu upgrade will be sometime next year. So between i3-6100 and Zen quad core APU whichever has better performance will get my money. If the $120 Zen APU has igpu that's atleast as fast as an r7 250x and cpu performance atleast as good as Haswell i3 then that should be enough for me to consider Zen.

Zen APUs are not coming until 2017 at best, and unless they solve the bandwidth problem, it is highly unlikely the gpu will equal a 250x either. And if it does equal the performance of a 250x, I seriously doubt they will sell it for 120.00.

 

[itsmydamnation]

This seems overly cynical. The reason "FP-heavy benches" are ubiquitous isn't because the reviewers were biased against AMD, but because FPU-focused loads are extremely common in ordinary software. You have to stretch to find a handful of relevant modern programs that don't rely heavily on FPU operations (e.g. 7-Zip LZMA and x264).

What does FPU heavy load mean? Most software (99+% ) is not SIMD FPU heavy. there is a simple reason why, Automatic vectorization doesn't work. You are far more likely to find scalar FPU heavy workloads then you are SIMD heavy. scalar will likely have lower ILP and benefit more from "IPC" uplift then SIMD code.

FPU-focused loads are extremely common in ordinary software

This hear if your referring to SIMD FPU (SSE,AVX) is particularly false.

Even Sandy Bridge shared some of the AVX hardware path with SSE. From Anandtech's architecture review:

of course it does, it would make zero sense for it not to , the big difference between AVX and the later SSE's is that its 3 operand and that you can do 256bit ops, but you cna still do 128bit ops.

 

[Glo.]

Anyone seen this:
https://arch.b4k.co/g/thread/52492812/#52493743

If this is true, Zen will be very competitive with Intel.

 

[LTC8K6]

Anyone seen this:
https://arch.b4k.co/g/thread/52492812/#52493743

If this is true, Zen will be very competitive with Intel.

I didn't see any such info there.

 

[jpiniero]

I have serious doubts anything will be fabbed at Samsung. Just GloFo, if only because of the WSA. Although the big GPUs might be fabbed at TSMC.

 

[el etro]

Anyone seen this:
https://arch.b4k.co/g/thread/52492812/#52493743

If this is true, Zen will be very competitive with Intel.

Really have new info(debugging phase started in January(Fottemberg said it too, dual sourcing rumor coming for the second time). But Broadwell IPC is Fottemberg(SemiAccurate forums) rumor.

 

[Silverforce11]

Zen APUs are not coming until 2017 at best, and unless they solve the bandwidth problem, it is highly unlikely the gpu will equal a 250x either. And if it does equal the performance of a 250x, I seriously doubt they will sell it for 120.00.

Bandwidth problem already solved. HBM has proven itself already. HBM2 is now in mass production.

The only question will be, will desktop consumers pay a premium for a powerful HBM2 APU?

Servers may for their work-loads. But similar to Intel's Iris Pro+ Crystalwell, the premium makes it a poor choice compared to CPU + dGPU combo.

 

[csbin]

AMD Zen 8 Core 95W Summit Ridge CPUs Launching In October – Already Going Through Validation

Read more: http://wccftech.com/amd-zen-cpu-8-core-summit-ridge-launching-october/#ixzz42puYk3B4

 

[Silverforce11]

Anyone seen this:
https://arch.b4k.co/g/thread/52492812/#52493743

If this is true, Zen will be very competitive with Intel.

You don't need to see random rumors as a sign Zen will be competitive.

Lisa Su's statement to investors, "on track for greater than 40% IPC gains" should already be enough if you jump to the logical conclusion.

If that's compared to Piledriver, they would be on Broadwell IPC levels. Worse than Skylake, but not by much. That's a fail right? No.

Basic Zen for the high-end consumer is a 8C/16T SKU.

Think about that and compare that to Skylake 6700K, a 4C/8T product and eventually the hexacore variant, 6C/12T (and the prices!).

Zen can slot right in and price compete with 6700K and be a winner.

Think about Haswell vs Skylake, 4790K vs 6700K, there are differences in gaming, but we're talking about 100+ fps in modern games here, beyond most GPUs capability. So if 8C/16T Zen is able to match the 4790K for gaming or even better (if games are well threaded), but it offers twice the cores for programs, it will make for a very compelling purchase. Great for gaming, excellent for apps, for a 6700K price?

 

[frozentundra123456]

AMD Zen 8 Core 95W Summit Ridge CPUs Launching In October – Already Going Through Validation

Read more: http://wccftech.com/amd-zen-cpu-8-core-summit-ridge-launching-october/#ixzz42puYk3B4

"Allegedly" launching in October, to quote more accurately.

 

[frozentundra123456]

You don't need to see random rumors as a sign Zen will be competitive.

Lisa Su's statement to investors, "on track for greater than 40% IPC gains" should already be enough if you jump to the logical conclusion.

If that's compared to Piledriver, they would be on Broadwell IPC levels. Worse than Skylake, but not by much. That's a fail right? No.

Basic Zen for the high-end consumer is a 8C/16T SKU.

Think about that and compare that to Skylake 6700K, a 4C/8T product and eventually the hexacore variant, 6C/12T (and the prices!).

Zen can slot right in and price compete with 6700K and be a winner.

Think about Haswell vs Skylake, 4790K vs 6700K, there are differences in gaming, but we're talking about 100+ fps in modern games here, beyond most GPUs capability. So if 8C/16T Zen is able to match the 4790K for gaming or even better (if games are well threaded), but it offers twice the cores for programs, it will make for a very compelling purchase. Great for gaming, excellent for apps, for a 6700K price?

So basically what you are saying is that we are going to get 8 core Haswell E or better performance for 6700K price. Do you seriously believe that???

 

[Silverforce11]

Another thing that people tend to gloss over, Intel's Skylake has a massive portion of the chip as the iGPU.

Zen 8C/16T may well end up at something of an equivalent die-size as the 6700K.

Imagine Zen 8C/16T for $349 (obviously dream scenario, but a tiny chip can be very flexible with prices!) with Broadwell IPC, matching the i7-5960X 8C/16T on performance. Deal or no deal?

Really, anywhere close to Broadwell IPC with that many cores is an automatic winner unless they go full stupid on prices.

Edit: For added clarity, think about the huge margins/profit Intel makes on the 6700K being such a small chip and they sell it in bulk at $350. There's no reason AMD can't make a huge profit/margin selling a Zen 8C/16T chip that lacks the huge iGPU for a similar price.

 

[mysticjbyrd]

So basically what you are saying is that we are going to get 8 core Haswell E or better performance for 6700K price. Do you seriously believe that???

Well the 5960x is overpriced by $400-500 easily

 

[.vodka]

Exactly!

That's easily another 4 cores for what desktop Skylake actually is, in a tiny 120mm^2 chip.

Us dGPU users don't need the iGPU, apart from features like quicksync, for example... yet another 4 cores can more than make up for something like that. Yeah, there should be more L3 cache if we go by socket 2011 chips with increased core counts, yet it shouldn't be a much bigger chip with all the added cores and extra supporting stuff. The APUs will come later.

Bring it on AMD! Hope it ends up being something like this.

 

[frozentundra123456]

Hate to tell you but Haswell/Broadwell E has no igp.

 

[Arachnotronic]

Hate to tell you but Haswell/Broadwell E has no igp.

lol, don't bother. People have been whining about wanting CPUs with the die space dedicated totally to CPU/cache, and yet these are the same folks who thumb their noses at the "E" line. Of course, the "E" line doesn't have the newest cores and those people will complain, but at the same time those very same complainers will say that the IPC increases per generation aren't all that impressive & they would be cool with an AMD chip w/ more cores at IPC lower than Intel's latest mainstream CPUs.

The logic of Intel-haters truly and utterly escapes me

 

[Silverforce11]

Hate to tell you but Haswell/Broadwell E has no igp.

We already knew that, what is your point?

Do you have something to suggest AMD can't make an 8C/16T Zen on 14nm FF that is similar in the die size of Skylake? It's tiny, ~120mm2.

$349 for a chip that size is massive profits.

That's not to say they will price it as such, if the performance rivals 5960K, they can price it for the market, but the die size gives them flexibility on pricing if they need to. Get it?

 

[Sweepr]

FX9590's launch price and this message from Lisa Su should give a hint about what an 8C/16T solution (with Broadwell IPC) would cost:

“The idea that AMD is a cheap solution has to be replaced with the idea that AMD is a very competitive solution,” Su said in a roundtable with reporters here at CES.

 

[Arachnotronic]

We already knew that, what is your point?

Do you have something to suggest AMD can't make an 8C/16T Zen on 14nm FF that is similar in the die size of Skylake? It's tiny, ~120mm2.

$349 for a chip that size is massive profits.

That's not to say they will price it as such, if the performance rivals 5960K, they can price it for the market, but the die size gives them flexibility on pricing if they need to. Get it?

Dude, you're the one who doesn't "get it." What makes you think that die size is the only thing that goes into determining the cost of a chip? Have you ever heard of a little thing called "binning"? Only the cream of the crop SKL 4+2 chips run at 4GHz base/4.2GHz turbo in a given power envelope, and since they are "rarer" they obviously cost a lot more.

 

[Silverforce11]

Dude, you're the one who doesn't "get it." What makes you think that die size is the only thing that goes into determining the cost of a chip? Have you ever heard of a little thing called "binning"? Only the cream of the crop SKL 4+2 chips run at 4GHz base/4.2GHz turbo in a given power envelope, and since they are "rarer" they obviously cost a lot more.

Please dude, if Intel can go to the bank with record margins on Skylake, why are you assuming Samsung's 14FF somehow can't be competitive at delivering good chips?

Where is this magic that you guys apply to only Intel's tech?

And regarding the price, I fully expect them to not be cheap if they have the performance. The point is a small chip gives them flexibility to price it well for the market. Without the iGPU, they can still make a small chip with 8 physical cores. Simple. -_-... such a simple concept.

 

[Arachnotronic]

Please dude, if Intel can go to the bank with record margins on Skylake, why are you assuming Samsung's 14FF somehow can't be competitive at delivering good chips?

Samsung 14FF so far has only been proven in low-power mobile chips (i.e. A9, SD820, Exynos 7/8). There is no evidence thus far that it can deliver >4GHz, high perf/clock chips in anything close to the kind of power envelope that Intel's can.

Intel's processes are fine tuned for the very limited range of fairly high-performance PC processors that it builds, while Samsung/TSMC are basically gunning for the mobile guys first and foremost.

Intel also builds its own chips & doesn't have to hand over any foundry margin to Samsung/GloFo/TSMC, which certainly helps on the cost structure front.

BTW, don't you think it's just a teensy-weensy bit suspicious that AMD is able to fit 8 whole super duper Zen cores into a 95W power envelope on a process that is probably inferior for high performance CPUs than Intel's 14nm? What do you think those CPU cores are going to come clocked at? 4GHz at Skylake-like IPC? Very doubtful.

 

[zentan]

Intel needs those margins,they run a much bigger business and have their hands in much more areas CPU,SSD,3dx-point,modems etc and a cutting edge IDM. To remain in front they need massive R&D and that's where most of the profit goes.The type of PC market demand that has been there since 1-2 years it is what Intel needs to carry on and take bigger challenges.The revenue from PC market is declining in the past.

~Broadwell IPC for Zen cores? Hmm,would need some actual proof of that. I just hope it doesn't turn out to be another hype-train till the launch.

"Simple concept"? Well many such things can be made but at the end what is economically more suitable.They already have a good option i7 5820k for those who require more cores.It's broadwell-e successor would come soon.

 

[Abwx]

Please dude, if Intel can go to the bank with record margins on Skylake, why are you assuming Samsung's 14FF somehow can't be competitive at delivering good chips?

Samsung s, or rather s GF s, 14nm is more efficient than Intel s, any competent enginer will state it, it s just that a given public is putting all his hopes in AMD not succeding, hence all the blank statements you can read here and there..

 

[Arachnotronic]

Samsung s, or rather s GF s, 14nm is more efficient than Intel s, any competent enginer will state it, it s just that a given public is putting all his hopes in AMD not succeding, hence all the blank statements you can read here and there..

Could you please provide proof that Samsung's 14nm is "more efficient" than Intel's? Thanks.

Additionally, if you could provide a statement from a "competent engineer" that'd be great, preferably one with experience with both processes.

EDIT: To be clear, Abwx, please provide a proof that makes very little assumptions on the part of the reader. If you have a legitimate proof, you should be able to explain it without saying things like "people without EE degrees don't understand" and "trust me, I'm a retired semiconductor professional." Take it from first principles, explain all terms along the way, and show your math. Try to make a valuable contribution.