New Zen microarchitecture details

[majord]

Intel uses a mobile designed CPU for mainstream desktop & laptop, not their server design that is offered for HEDT.

Ryzen is AMD's server designed CPU, unless they too have done a mobile specific design, it is very unlikely it will match Intel in laptops.

Hang on, this makes no sense.

Firstly Broadwell-E and Broadwell Core M are exactly the same core.

Secondly, in regards to future HEDT cores, the main factor that separates mainstream and server/HEDT Skylake is beefier AVX capabilty (AVX512 units/ datapaths) on the later. Zen has exactly the opposite next to mainstream Skylake, so following that pattern, that makes Zen more suitable for mobile..

 

[The Stilt]

Keep in mind that AMD has vastly been behind Intel in the manufacturing process side as well. Because of that, they had to develop alternative ways to compensate the disadvantage(s) as much as they could. The power management on AMD's most recent designs is extremely advance and effective. I dare to say more advanced and efficient than Intel's. The power management is basically the only thing, I saw vast improvements in during the whole 15h ERA. If AMD can turn the originally server-targeted, flawed 15h family into something as efficient as Excavator you can be certain that there won't be any issues in doing the same for Zen, regardless if it was originally server targeted or not. Unlike the 15h family, Zen was designed for versatility and efficiency from the beginning. The 15h family, even in form of Carrizo is a relic compared to it.

 

[imported_jjj]

how can you guys say any of this without knowing the power consumption...
I don't care about igpu perf I just need good enough. So if intels solution is lower power consumption than AMD then I'm happy. So we need the full picture.
You can't just state as a matter of fact that AMD will have same perf as intel, more GPU perf than intel and lower power consumption.....
Seriously?

Nobody can state that yet but there are some assumptions one can make.
The mobile process might help AMD scale down well and Intel has a large core, AMD could have done better as it is a new core designed to scale from the start.
On the GPU side, Vega seems a large step forward in power consumption. They seem to trade area for much lower power. The memory bandwidth required by Vega should be substantially lower than their previous GPUs ( remains to be seen if they have a large GPU cache to save even more) and that would be a substantial power saving. The work with game developers should help too.
It's also possible that there will be different winners at different TDPs. Maybe one does better at 15W and the other at 5W.
Ofc nobody knows much for sure but there is optimism, especially after the few Vega details we got last week.

Cost matters too. No chipset required reduces costs, board complexity, time to market for OEMs. Intel's prices are rather high, especially in low power. Intel might not try to match prices for the most part as the average buyer is not informed and businesses are comfortable with Intel so AMD would gain share at a relatively slow pace and no point for Intel to kill its margins and ASPs by matching AMD's pricing.

Edit - Not that small differences in SoC power consumption will matter much since everything else in the system uses power and when you add it all up, the difference becomes irrelevant,if you are talking 5-15W SoCs at low load.

 

[krumme]

Can the cpu use the same hbm2 cache as the gpu?

 

[CatMerc]

Can the cpu use the same hbm2 cache as the gpu?

Why would it though? CPU's need low latency cache for improvements, bandwidth doesn't tend to be their problem.

 

[beginner99]

I am not sure that you have considered the math at all and AMD's current revenue.
This CPU only Ryzen thingy is not for OEM, i have never claimed it is.It is for us,for the DIY market mostly and it is a huge market if you think about it.
If you look at JPR stats for discrete GPUs in desktop, last year will end up at likely above 45million units. PCs with CF and SLI are a small percentage. There is a huge market in units for people that don't need integrated GPUs.

dGPU doesn't mean it's DIY. HP, Asus, Acer, Dell and so forth sell plenty of crappy prebuilts with dGPU.

If the octa cores clock to 4GHz or better ,the quads will do much better.

Maybe. Maybe not. If the process can't to much more than 4 Ghz, then even a single-core with 200 W TDP can't break that limit. It's a hard limit. We don' t know at this point but it's a naive assumption.

Gaming doesn't scale with clocks ,some do but some scale with cores and the direction is towards more cores.

I say the exact opposite. All games scale with clocks, most also with more quad vs dual core basically none with more than 4 cores (and we are talking about 4 fast cores not consoles or crappy FX-cpus that do look a little less bad compared to i3 in most recent games).

If you want to pay 300+ for Kaby with integrated GPU for gaming, instead of 179$ or199$ for a 4.5GHz (or higher with good cooling) quad Zen, it's not very wise but what do i care.
The Pentium you mention is dual core and yes Ryzen doesn't try to compete with that, that's the entire point for this platform and that gives it such a huge potential.

And now to the main point. I quoted your first post because you where writing about quad-core zen for $99 and lower:

They could even go to 49$ for quads but it is likely that ,this year, most that buy bellow 99$ would make an effort and get the quad. Next year,when Zen+arrives,they can drop the prices for this gen to 49$.

And now suddenly you shift goal posts to a price figure at $179 and in line with what i said? Do you actually see you are contradicting yourself?

If they do offer a quad it will go against lowest i5, highest i3 pricing so around $150.

So you basically repeat what I said while contradicting yourself. Make up your mind and stick to it or admit you changed your mind and your previous logic was flawed.

You also seem to think that they won't have a native 4 cores Ryzen die but the hints so far point to such a die being launched from day one and as i said, any price band they don't address,is business left on the table and it just feeds Intel. It would be bad business.

Wrong again. If Zen actually performs and the process does not absolutely suck (2 assumption I'm making), then most dies should at least work as a 6-core part and there would actually be a quiet small number of dies that only have 4 working cores so AMD would need to take working 6-cores and disable 2 cores. However if Zen does perform well with low power usage, the server world will be all over it and AMD will have trouble keeping up with demand. So if you have supply issues and functional hexa-cores you can sell for $100-$200 more than a quad, why the hell would you sell them as quads? You would not. You would release either no quads or wait couple month and release them then.

 

[krumme]

Why would it though? CPU's need low latency cache for improvements, bandwidth doesn't tend to be their problem.

To simplify the memory interface.
As latency is also a result of bandwidth i asumed hbm2 was far faster than the standard ddr4 2400?

 

[The Stilt]

Can the cpu use the same hbm2 cache as the gpu?

Given how the current APUs are built, that would be mandatory. HBM would be useless as a CPU cache due it's latency, thou.

 

[SpaceBeer]

But isn't HBM latency lower than DDR4? It can't be as good as intel's eDRAM, but that one is limited (currently) at 256MB. There are always things that could use little bit faster memory

 

[imported_jjj]

@ beginner99

Plenty use a dGPUs is not quite plenty nowadays because of APUs.and those that do could do without integrated too.
We do know that Zen doesn't hit a wall at 4GHz and you know that too.
No, games overall don't scale with clocks when you have sufficient perf,only some, go check the best 20 reviews you can find for Broadwell-E.
If they would, Zen has the as good as the cooling turbo and single core clocks are likely to be very high.
And again, lets also remember that the quads will clock higher- must repeat it since you are bound to try and distort reality again and again.
Maybe you are doing it on purpose,maybe you are just making a silly mistake but the difference between the price points i mention at 99$ a quad Zen and another at 179$ is clearly about different SKUs(you really did not strike gold there), the cheapest SKU would be 99$ and the fastest 179-199$. Just like Intel prices its quad Kaby Lake between $182.00 for the 7400 and $339.00 - $350.00 for the 7700k.
Then you seem to not understand or pretend to not understand what a NATIVE quad core die means.

You really need to find some objectivity somewhere, these are just corporations, having feelings for one is irrational.

EDIT- Focusing on the top SKU is also utterly irrelevant,what matters is what each has to offer at every price point and with no GPU AMD can offer up to 2x the CPU perf.
And if you care about gaming perf in something that scales with clocks, save 150$ by buying a quad Zen that offers similar perf to what Intel has and buy a better GPU.Or save up to 250$ by buying the cheapest quad Zen,overclock it and buy a much better GPU. We'll all have that flexibility now.

 

[bjt2]

Two considerations on the GF 14nm process:

1) Low power. The famous NEON test chip, has a nominal Vcore of 0.8V@2.41GHz. With the usual +-10% margins. Power measured at 0.9V and 125C Tj and minimum voltage to work at 2.41GHz of 0.7V. AFAIK SKL requires 0.9V for similar clocks (2-5-2.7, i don't remember). INTEL does not have AVFS and DVFS that allow AMD to drop the 10% margin. Moreover Zen should have lower FO4 than the NEON FPU, so almost certainly will require at most 0.8V (if not 0.7V) to clock to 2.4GHz... Another advantage and a hint on very low power usage. I posted a while ago a graph in which a 15W excavator battles in cinebench with a 15W SKL. A 28nm BULK vs 14nm FF CPU! With an awful architecture. Imagine Zen on the low power 14nm process...
2) High clock. Zen will not hit a wall at 4GHz. This nonsense is only the last hope of deluded INTEL fans. Excavator does not wall even at 4.9GHz. 14nm FF has higher transconductance, lower capacitance, lower Vth, lower leakage and so it's faster AND lower power. How can be Zen walled at an inferior clock? Actually we have rumors of at least 5GHz OC... Moreover Zen does not even have cold bug.

 

[AtenRa]

More than 50% faster than current generation only makes it about as fast as 940MX, rebranded low end first gen Maxwell card. That's hardly an achievement making an APU a compelling device for anything but accelerated video. And i would know, i have Bobcat netbook.

I dont know what more you are expecting from a 15W TDP cheap SoC. Even installing HBM2 in the 15W TDP SoC will not raise performance that much due to Thermal constrains, they will have to go to 7nm to gain significantly more performance from those 15W TDP SoCs.

I am betting my void signature you will not see HBM2 APU for mainstream market (read: a laptop or desktop SKU) until at least 2019, if ever. Are you willing to bet your own?

I said late 2018, it could be Q1 2019 i dont know. But im expecting HBM2 to be available in mainstream in the next gen ZEN APU. The product will be in high volume production at the time because both AMD and NVIDIA will use it in Server/HPC and Desktop products. Also the HBM2 technology allows to minimize BOM from a design and offset the increased HBM2 cost to create a higher valued product.
What we have to remember is that HBM2 was not created to be a low volume product, they have to use it in mainstream in order to really increase production volume and that will decrease the product cost. I cant find it now but there was a slide that had HBM2 in mainstream from 2018, a single stack 4GB 128GB/s mainstream APU could be viable for production in late 2018.

 

[imported_jjj]

But im expecting HBM2 to be available in mainstream in the next gen ZEN APU. The product will be in high volume production at the time because both AMD and NVIDIA will use it in Server/HPC and Desktop products. Also the HBM2 technology allows to minimize BOM from a design and offset the increased HBM2 cost to create a higher valued product.

Adoption is not sufficient from a cost perspective.I think Samsung was proposing a budget HBM at Hot Chips (will edit to add link) and they also need a cheaper interposer. I think AMD likes organic but even Intel's silicon bridge could work- limited but cheap.And ofc ideally they split the CPU and GPU to save more money and gain flexibility.

EDIT
Samsung budget HBM proposal, video at time https://youtu.be/61oZhHwBrh8?t=3280

 

[leoneazzurro]

But isn't HBM latency lower than DDR4? It can't be as good as intel's eDRAM, but that one is limited (currently) at 256MB. There are always things that could use little bit faster memory

The actual CPU cache latencies are much lower than those of all DRAM types. Please note that CAS latency is not actual latency when addressing DRAM -> https://en.wikipedia.org/wiki/Memory_timings .
This is quite normal, as full speed SRAM cache on-die has significant signal propagation advantages respect to a DRAM memory wired by means of an external bus requiring traces on a PCB.

 

[SpaceBeer]

Yes, I'm aware of that. But I'm not saying AMD should decrease/remove L2/L3 cache and place HBM instead. Just that HBM could be used as a "first memory choice", before DDR. Similar to GPUs - when it runs out of VRAM, it will use slower DDR3/4 if needed. It might be good thing if it is cheaper/better than investing in quad (or more) channel platform, and insisting in many RAM modules.

Though I don't know why I'm talking about this, since I'm not so familiar with this stuff and it might just be some stupid/unreal solution

 

[imported_jjj]

Yes, I'm aware of that. But I'm not saying AMD should decrease/remove L2/L3 cache and place HBM instead. Just that HBM could be used as a "first memory choice", before DDR. Similar to GPUs - when it runs out of VRAM, it will use slower DDR3/4 if needed. It might be good thing if it is cheaper/better than investing in quad (or more) channel platform, and insisting in many RAM modules.
Though I don't know why I'm talking about this, since I'm not so familiar with this stuff and it might just be some stupid/unreal solution

The way to think about this is that ,in consumer,it's about costs not perf. Spending a lot for small gains in perf doesn't work, just like quad channel is not worth it today in consumer.
If they invest in a solution ,they need to at least generate sufficient revenue to avoid loses.Revenue would be about units and price.Low volumes would require a very high price.
High perf is low volumes so how do you make a product that can be sold at a high price by being better than alternatives.
A very large APU would have to compete with discrete GPUs and ,in this case,beat their own solutions and Nvidia's solutions to sell well enough.The math might not add up.
A small APU with HBM would be too costly vs normal APUs while offering small perf gains.
The memory could be HBM like so some 3D stacked DRAM and much cheaper interposer or even just a DRAM die very close to the CPU with a very high bandwidth link. And sure if the CPU can use this memory ,maybe they save on DRAM costs. However, it's likely not quite enough to justify spending on it.

The best chance to see an APU with HBM anytime soon would be if they make it for server and decide to also sell it in consumer.Server pays for it and consumer just widens the market a bit.
Longer term, with cheaper advanced packaging solutions, they could do it to save money. They develop a CPU and a separate GPU die and they pair them- 1 CPU no GPU,CPU and 1GPU, CPU and 2GPUs, 1GPU no CPU and so on and use those combinations across the board. Development on the most advanced process is very costly and getting worse. A small die also has better yields and in consumer, some extra latency between CPU and GPU doesn't quite matter if it enables substantially lower costs - ofc compared to discrete ,the latency is nothing. They would lose a bit in perf but gain a lot in costs and flexibility.
If they do this, they are already paying for advanced packaging and developing every die that goes with it and since the GPU needs bandwidth, they might as well use some HBM like memory in the SKUs that require it. Cheaper advanced packaging and stacked memory solutions are crucial to enable this and those are arriving soon.

They might be already doing this in server to save on dev costs although it is a bit weird to lose perf in a segment where costs matter a lot less but they have limited resources today and at least the ASPs in server are high enough to cover the current costs for advanced packaging. It's a bit upside down for now and maybe they give up on that in server to chase every bit of perf and power they can, but adopt it in consumer when cheaper packaging and stacked memory solutions will be available.

 

[SpaceBeer]

I agree we won't see anything like this in consumer segment anytime soon, and when/if they come it will be in server/HPC market. I'm just curoious what is the best strategy for the future. Will HSA live?
https://www.amd.com/en-us/innovations/software-technologies/25x20
Kaveri was released 3 years ago, and I don't know any program except Libreoffice Calc that can take advantage of its architecture. And my A10-7850K + LO Calc was faster than i5-4440 + MS Excel in some tests I've done. And Kaveri even wasn't the the true HSA chip

 

[Abwx]

Two considerations on the GF 14nm process:

1) Low power. The famous NEON test chip, has a nominal Vcore of 0.8V@2.41GHz. With the usual +-10% margins. Power measured at 0.9V and 125C Tj and minimum voltage to work at 2.41GHz of 0.7V. AFAIK SKL requires 0.9V for similar clocks (2-5-2.7, i don't remember)..

2) High clock. Zen will not hit a wall at 4GHz.

0.89V@2.41GHz for Intel, to compare to said 0.8V@2.41GHz for GF.

Since conductance of a fet transistor increase as a square law of voltage GFs transistor have (0.89/0.8)^2 = 1.237x better transconductance and are of course faster by the same ratio..

As for maximum frequency that s another matter, because the faster device is not automatically the one that has the higher possible operating frequency, there s other paramaters at play here, most importantly the current drive capability at high voltage/currents (= at high frequencies).

That said, and given GF s LVT transistors parameters, 4GHz should require 1.03V if we are to extrapolate from the test chip voltage numbers, and surely a little more to take account of the bigger voltage drops that occur with a 95W CPU.

 

[KTE]

0.89V@2.41GHz for Intel, to compare to said 0.8V@2.41GHz for GF.

Since conductance of a fet transistor increase as a square law of voltage GFs transistor have (0.89/0.8)^2 = 1.237x better transconductance and are of course faster by the same ratio..

As for maximum frequency that s another matter, because the faster device is not automatically the one that has the higher possible operating frequency, there s other paramaters at play here, most importantly the current drive capability at high voltage/currents (= at high frequencies).

That said, and given GF s LVT transistors parameters, 4GHz should require 1.03V if we are to extrapolate from the test chip voltage numbers, and surely a little more to take account of the bigger voltage drops that occur with a 95W CPU.

There is NO high performance AMD mass production chip that runs at 0.8V@2.41GHz. Like Skylake.

Complete biased nonsense.

Sent from HTC 10
(Opinions are own)

 

[TheELF]

I agree we won't see anything like this in consumer segment anytime soon, and when/if they come it will be in server/HPC market.

Server market is more concerned about getting huge amounts of data from the disk to the CPU HBM won't help much there.
Optane dimms is what the servers will go after so that the CPU won't have to stall while waiting for data from the disk.

 

[bjt2]

0.89V@2.41GHz for Intel, to compare to said 0.8V@2.41GHz for GF.

Since conductance of a fet transistor increase as a square law of voltage GFs transistor have (0.89/0.8)^2 = 1.237x better transconductance and are of course faster by the same ratio..

As for maximum frequency that s another matter, because the faster device is not automatically the one that has the higher possible operating frequency, there s other paramaters at play here, most importantly the current drive capability at high voltage/currents (= at high frequencies).

That said, and given GF s LVT transistors parameters, 4GHz should require 1.03V if we are to extrapolate from the test chip voltage numbers, and surely a little more to take account of the bigger voltage drops that occur with a 95W CPU.

Could you give a link were you found that informations on GF transistor transconductance, etc... I am eager to read it...

 

[bjt2]

There is NO high performance AMD mass production chip that runs at 0.8V@2.41GHz. Like Skylake.

Complete biased nonsense.

Sent from HTC 10
(Opinions are own)

There is a test chip of TWO years ago that runs at that frequency. And it's a synthetized ASIC design, with an high FO4 (>30). Zen is a low FO4 (<30) custom design. And it has AVFS and DFVS that let you avoid the 10% voltage margin. I think that at 0.8V Zen will go at least at 2.4GHz...

 

[Dresdenboy]

Adoption is not sufficient from a cost perspective.I think Samsung was proposing a budget HBM at Hot Chips (will edit to add link) and they also need a cheaper interposer. I think AMD likes organic but even Intel's silicon bridge could work- limited but cheap.And ofc ideally they split the CPU and GPU to save more money and gain flexibility.

EDIT
Samsung budget HBM proposal, video at time https://youtu.be/61oZhHwBrh8?t=3280

Here is the slide from the video:

And this could be an organic interposer in the lower right:

 

[imported_jjj]

Here is the slide from the video:

And this could be an organic interposer in the lower right:

In the same session , Xilinx also talks a bit about advanced organic

 

[Atari2600]

Well if AMD pulls of the wonder and get's cpu parity then a 4 core zen APU is going to be massively bottlenecked by it's graphics...

Except probably >90% of the world doesn't give 2 craps about 3D graphics performance but needs a power efficient processor (i.e. desktops in professional environments).