[SemiAccurate] AMD to differentiate cores

[jpiniero]

Cannonlake and Icelake Client do support AVX-512, it just doesn't have the extra AVX-512 unit.

Little Zen might also be useful for big.LITTLE type products.

 

[maddie]

What's the word? Node increases have and will continue to slow.

The expectation of traditional fab advances to lift all boats has died and so the need for specialized tuning for particular markets take the lead. Yes, Zen can go to ~4W if I remember correctly, but at the extremes of a range one would normally expect optimizations to have less effect. I maintain that a core optimized for 4W or less, will be better than Zen at the same wattage. Extremely long battery lifetime and a laptop with those properties is a big marketing advantage. AMD needs to take mobile to keep gaining marketshare, as I assume that their plan is not to be the 2nd player but to go for the win.

A simple thing like using different libraries will necessitate a new design and if you have to do a full redesign of masks, you should go the full step and completely reoptimize the core. Not saying that there will not be shared 'DNA' so to speak.

 

[beginner99]

Different cores means different designs and dies means additional 9-figure investment with current cost of nodes and compared to current situation would double the initial costs. I don't see how that in anyway is more profitable.

This goes against the clear advantages of the chiplet design and the amount of use-cases for AVX-512 stuff is rather niche and limited and can be covered by GPUs or dedicated accelerators.

 

[naukkis]

Differentiate like K12 vs Zen?

 

[maddie]

Different cores means different designs and dies means additional 9-figure investment with current cost of nodes and compared to current situation would double the initial costs. I don't see how that in anyway is more profitable.

This goes against the clear advantages of the chiplet design and the amount of use-cases for AVX-512 stuff is rather niche and limited and can be covered by GPUs or dedicated accelerators.

We have to differentiate strategies as relative to resources, meaning that which is optimal when resource limited might not be the best possible in all situations. A general purpose design, almost by definition, is good at all, but not great at any. This is not to take away credit from AMD, as I have stated many times that the 1 die strategy was absolutely brilliant.

Now however, with an increased cashflow and profitability, the opportunity exists, for the first time in many years to have optimal products in a wider range of computing.

I'm not saying to have different designs just for the sake of optimization, but that each design must have a large enough sales volume to support it.

 

[gorobei]

Different cores means different designs and dies means additional 9-figure investment with current cost of nodes and compared to current situation would double the initial costs. I don't see how that in anyway is more profitable.

This goes against the clear advantages of the chiplet design and the amount of use-cases for AVX-512 stuff is rather niche and limited and can be covered by GPUs or dedicated accelerators.

unfortunately this sort of vagueness from charlie isnt very helpful. most likely this is just a tease for amd's plans for heterogeneous chips on active interposer ie the same thing as foveros execept that amd has been working on it longer and will probably be first to get a real product out.

Forrest Norrod's talk at Rice seems to imply that they do have a cpu+gpu chiplet product in pipeline with lego block customization for the number of chiplets for server. active interposer would be the best way to get the io die out of the way to reduce substrate size so it can fit on atx motherboard socket sizes, but i dont hold any hope for zen3 on am4.

 

[moinmoin]

Yeaa. "However it also won't be the last; while the latest amendment sets purchase targets through 2021, the overall WSA itself will still run through March 1, 2024."

The current io dies and zen plus apu can probably fullfill that role fine in 19 to 21 but from there it seems to me its a bit unclear how amd should move gf capacity?

AMD's embedded products all have promised availability for 10 years.

 

[Topweasel]

Eventually AMD will probably get to a point where Semi custom will drive their biggest market breaks. So there are situations where they need a new die anyways. Like a true embedded sub 10w with video scenario. No amount of turning off functionality is going to really help them with that if you are still dealing with a 150-200mm die or using an IO die. On top of all that a true embedded CPU might need more functionality born into the chip (Sata, Nic and so on).

So my guess is this is how it goes. Let's say someone wants a semi high performance low wattage APU for their embedded endeavor. They go to AMD, AMD say's "Sorry I don't have a design that works with that right now, but we can have one in a couple of years but we will have to split development costs and I have the right to sell the chip later and you agree to purchase X amount of them". Kind of like the development of the F-35. So AMD then takes out the L3 cache, only has 2-3 GPU bits and ends up with a 50mm 4 core APU that uses 4-7w.

But then someone else comes along and says I want EPYC but with native AVX-512. AMD tells them to stuff it because they would have to start from scratch. That's unless they were already going down that road with Zen 3 or Zen 4.

The main point being I think requests and split development cost like with Navi are going to drive AMD's offerings in the future. It won't be one size fits all like Intel attempts to do. But it probably won't be the wild west either. It might even be something like full package offerings like the Xbox and Playstation if the companies in question are willing to eat 90% of the cost and those products will be unique to them.

 

[moinmoin]

What you write there is happening since nearly a decade already, customers paying for semi custom design that are a mix and match of existing IP (GPUs were already modularized, CPUs were modularized with the construction and cat lines) as well as new R&D. It's what kept AMD afloat before Zen.

 

[Topweasel]

What you write there is happening since nearly a decade already, customers paying for semi custom design that are a mix and match of existing IP (GPUs were already modularized, CPUs were modularized with the construction and cat lines) as well as new R&D. It's what kept AMD afloat before Zen.

Well to a degree. But that's with the manufacturer owning the IP to some degree (and choosing when and how its built). I am thinking more along the lines of like Navi. A product that will be designed to serve both the console market and the retail market where they can use customer requested features to offer products to other customers. Like AMD will only offer a machine learning version of Zen if someone pays them to develop it, but then they now have a machine learning product stack. A perfect non-AMD example would be Kaby-G. Driven into development by Apple who chose not to use the product anyways, but still a product that only exists because Apple with their relationship wanted to see what the result would be so they could contemplate using it. It's a little different because Intel paid for most of that to make sure it looked like they were trying to develop stuff for Apple and in the end they really didn't change much at all (just a Kaby Lake with a Vega and HBM glued to it). But that I think is where AMD as they grow will want to create some seperation. They can't on their own create a die/cpu for every market. But if someone is willing to pony up the up front money, they will give it the college try.

 

[moinmoin]

I think people have and continue to give AMD's semi custom business too little credit. A lot of R&D from that area made it into AMD products. AMD's whole GPU development essentially progressed at the pace of the development for consoles, Navi is not special in that regard (though one could argue so far this coupling has been for the worse while Navi may finally be able to turn this around). The consoles used the 4+4 CPU cores structure before Zen 1 and now Zen 2 chiplets use them. I'm sure there are plenty other developments that originate from R&D as part of the semi custom business once one looks closer.

One has to keep in mind that the expense for new designs is at best the same for AMD and its customers. This obviously limits which customers can expend how much on new specialized designs and the necessary R&D for that. So it should come at no surprise that the biggest customers are Sony and Microsoft, both with multi-million console units sold, enough for both of them to warrant expending money for significant mid-gen console redesign refreshes. Kaby G appears to me as a smaller project that helped AMD create a mobile Vega and Intel to showcase its EMIB tech. In comparison Zhongshan's Subor Z+ is a much bigger deal imo, it being the biggest Zen based monolith APU existing. We don't know how much is happening for the cloud infrastructure, but at least the Vega cards for Google Stadia are reportedly customized as well.

 

[NostaSeronx]

I can get at least;
1. Zen2 w/ FP128 for consumers and Zen2 w/ FP256 for enterprise.
2. A ultra-small core for tablets, another core for laptops, another core for desktop, and another core for HPC.
3. Zen w/ SMT and another core w/ CMT. With Zen targeting premium(High Performance from IPC via increased Density on bleeding edge(FinFET/Nanosheet) at high Cost) and the other core targeting budget(High Performance from EPI via increased gate control w/ leakier LVT(FBB-RVT) at higher voltages on cost-effective edge(FDSOI) at low Cost).
4. One core for x86, one core for ARM, another core for RISC-V.

 

[whm1974]

I can get at least;
1. Zen2 w/ FP128 for consumers and Zen2 w/ FP256 for enterprise.
2. A ultra-small core for tablets, another core for laptops, another core for desktop, and to another core for HPC.
3. Zen w/ SMT and another core w/ CMT. With Zen targeting premium(High Performance from IPC via increased Density on bleeding edge at high Cost) and the other core targeting budget(High Performance from EPI via increased gate control w/ leakier LVT(FBB-RVT) at higher Vdds on FDSOI at low Cost).
4. One core for x86, one core for ARM, another core for RISC-V*.

*Original source: https://www.muycomputer.com/2018/12/28/chips-open-source/
https://riscv.org/2018/12/muycomput...c-v-to-manufacture-open-source-chips-updated/

Well certainly a plus for RISC-V if AMD is going to design Cores using that ISA.

 

[Tuna-Fish]

1. Zen2 w/ FP128 for consumers and Zen2 w/ FP256 for enterprise.

This is backwards. Despite Intel really trying to market it that way, enterprise, excluding HPC, doesn't really use wide SIMD. Or really much of SIMD at all. The largest use case is gaming. If it was worth it to segment by SIMD width, enterprise would largely want the narrower core.

3. Zen w/ SMT and another core w/ CMT.

CMT is dead, AMD is never doing that again.

 

[whm1974]

.
CMT is dead, AMD is never doing that again.

Certainly not as CMT only crippled performance and put AMD in the low end market.

 

[darkswordsman17]

If AMD wanted to compete in low power they'd do an ARM design, as ARM is already dominating most of that, and the bits they aren't they're moving to take over (ultrathin or cheap laptops; they already own most tablets outside of the full tablet Surface products and cheap 2n1s, and I have a hunch its just a matter of time til it takes over that).

AMD would be foolish to chase low power other than to look to use ARM cores (or minor tweaked ones that won't require major optimization over the standard ones) paired with their other tech (InfinityFabric, GPU - if it can be efficient and powerful enough for that market).

I do think its smart for AMD to be considering maybe doing different CPU chiplets for different markets. As far as their CPU development goes though, they should be focused on HPC/enterprise because that's where they'll get their money 5-10 years from now. If they can keep offering workstation and client and consumer stuff from that, great. But they'd be foolish to try to drastically chase cheap and low power markets that they can't naturally derive from what they already have. ARM will eat them alive and they'd just waste resources trying to compete. If they can make 4 core chiplets (basically shrink the smallest chiplet size) with little to no drawback (letting them offer more granular options) that could let them expand into markets shorter term, but I still feel like they just aren't going to be able to really compete well in those markets and it'd be pointless and a waste of resources to really try. Now if someone else would want to make the platform (take for instance that development board kit that someone put Ryzen APU on), great, AMD should be very open to helping them utilize AMD's chips, but they should keep the path they're on and build on that as its working. If there's fairly easy things they can do to offer variety of products, definitely go for it, but don't chase markets that don't suit what has been working for you (for instance, AMD would sholdn't be trying to develop network related solutions, they'd be better off trying to work with a company that specializes in that and finding ways to work with them to integrate to offer better end products).

Which I think is what they're aiming for, where they basically show they have good base tech (CPU, GPU, interconnect) that they could adjust to your needs and even combine with other chips. I don't think its a coincidence that Intel and AMD seem to be making similar moves (chiplets and looking at how to implement that on single package; looking to tailor the chips to different markets but not too drastically), as I think its at the behest of industry demands. They'll go about things in different manners, but in many ways, its to the benefit of both to make similar moves (as it establishes it for the industry, helping to keep x86 relevant). Intel has some definite advantages (in that they have robust networking and storage stuff they developed in house), but as we saw with Intel ditching 5G, things can change quickly. AMD just needs to focus on iterating their Zen plans, get their GPU division operating similarly, and then get them working together. And from there, best option would be to partner with others (either via licensing IP such as ARM cores, memory, etc, or being able to slot in chips/chiplets from others).

 

[whm1974]

At this point I'm thinking that AMD should mostly focus on improving their current APUs/CPUs and dGPUs. Any small cores should be derived from those products. I'm not sure if AMD can afford to license ARM cores or pay for developing them.

 

[scannall]

At this point I'm thinking that AMD should mostly focus on improving their current APUs/CPUs and dGPUs. Any small cores should be derived from those products. I'm not sure if AMD can afford to license ARM cores or pay for developing them.

The Architectural License isn't that bad of a cost. The dev time is however. If you are going to make an ARM core, then you have to aim for high end. Slogging through the swamp in pursuit of $20 parts is a money pit. Apple aimed high on ARM and have the performance to back it up. Not to mention revenue. ARM is a modern, well designed ISA. What you get out of it depends on how many transistors you throw at it plus whatever secret sauce your engineers can tweak in.

 

[naukkis]

At this point I'm thinking that AMD should mostly focus on improving their current APUs/CPUs and dGPUs. Any small cores should be derived from those products. I'm not sure if AMD can afford to license ARM cores or pay for developing them.

You do know that Zen is a side-project for K12, which was AMD's high-performance ARM core. AMD shelved K12 and put more resources to Zen for quick profits which made it's chief architect to leave, but as soon as AMD has funding for K12 why they shouldn't continue with it?

 

[whm1974]

You do know that Zen is a side-project for K12, which was AMD's high-performance ARM core. AMD shelved K12 and put more resources to Zen for quick profits which made it's chief architect to leave, but as soon as AMD has funding for K12 why they shouldn't continue with it?

I highly doubt that K12 would have profitable enough for actual products anyway. As far as "PC" type systems go, ARM is limited to notebooks and subnotebooks. Other companies already designing and/or producing ARM based SOC/processors would have eaten AMD alive.

 

[naukkis]

I highly doubt that K12 would have profitable enough for actual products anyway. As far as "PC" type systems go, ARM is limited to notebooks and subnotebooks. Other companies already designing and/or producing ARM based SOC/processors would have eaten AMD alive.

K12 and Zen were primarily data-center oriented products. Yes, AMD probably was little bit in weak position to gain attention from Microsoft to start desktop transition to ARM but it's totally different situation now as Qualcomm has desktop plans with Microsoft. Both Intel and AMD will be buried with x86 if they don't start to began transition towards ARM, and where's AMD K12 architect now?

 

[whm1974]

K12 and Zen were primarily data-center oriented products. Yes, AMD probably was little bit in weak position to gain attention from Microsoft to start desktop transition to ARM but it's totally different situation now as Qualcomm has desktop plans with Microsoft. Both Intel and AMD will be buried with x86 if they don't start to began transition towards ARM, and where's AMD K12 architect now?

x86-64 isn't going away anytime soon as that ISA is well entrenched and customers will be buying such PCs for a long time to come.

 

[naukkis]

x86-64 isn't going away anytime soon as that ISA is well entrenched and customers will be buying such PCs for a long time to come.

Customers won't even notice that cpu isa change after X86-64 support layer is good enough and ARM designs are so much faster than native x86-64 devices that both perform similarly. Intel is , of course, doing everything to prevent that but what they can do against AMD using it's own ISA with emulation?

 

[whm1974]

Customers won't even notice that cpu isa change after X86-64 support layer is good enough and ARM designs are so much faster than native x86-64 devices that both perform similarly. Intel is , of course, doing everything to prevent that but what they can do against AMD using it's own ISA with emulation?

Show me the benchmarks please.

 

[whm1974]

@naukkis I will give you that ARM has better performance Intel Atom based cores, but that a rather low bar to reach.