If AMD were to release a new hexcore chip what should its specs and die layout be?

[DrMrLordX]

Based on what I've heard, architecturally, FM2+ will be limited to 3M for Steamroller. No idea what would happen if they went back to the drawing board and tried doing an Excavator chip for FM2+. But, it seems clear that they are not interested in releasing Excavator on FM2+.

The real question is: what would be the least expensive design option for them on FM2+? They already have Steamroller products in the channel for FM2+. Rationally speaking, it should be fairly simple for them to release a 2M Steamroller without iGPU (relatively speaking, there would still be costs involved), and it shouldn't be too terribly difficult for them to expand on that design and go for a 3M design. But, it would have no L3.

In my opinion, that is the cheapest/most realistic scenario. It is possible that HDL libraries could get the die size very small, making production per die cheap, but the clockspeed losses could be ugly, and you'd incur the cost of applying HDL libraries to what is basically an extrapolation of an existing design (Kaveri).

Having pushed and prodded a 7700k extensively, I can tell you that Steamroller is pretty nice, and a 3M Steamroller would definitely be a winner compared to, say, the 6300. I would much rather have 4.7 ghz Steamroller 3M than 5 ghz Piledriver 3M. Refinements to the planar 28nm process might get us up to the ~5ghz department for those willing to push the envelope on cooling. It would be a nice chip on a good platform, especially on a board like the Crossblade Ranger.

AMD just doesn't seem to think it's worth the trouble, which is funny, since Intel is so stagnant in the per-core performance department right now. 3M Steamroller (or Excavator, but seriously, dream on) would show up a bit like Thuban did for AMD back in the Nehalem days . . .

 

[cbn]

The real question is: what would be the least expensive design option for them on FM2+? They already have Steamroller products in the channel for FM2+. Rationally speaking, it should be fairly simple for them to release a 2M Steamroller without iGPU (relatively speaking, there would still be costs involved)

Well, the Athlon x 4 860K does exist.

Unfortunately for us as consumers, I think AMD is trying to balance price vs. limited # of die harvested chips. (Since there are probably not too many 245mm2 dies with the whole iGPU damaged, AMD is most likely being very careful on how many of their chips they down-bin to meet demand for Athlon x4 860K.)

Or are you meaning a unique 2M Steamroller die in which there is no iGPU?

and it shouldn't be too terribly difficult for them to expand on that design and go for a 3M design. But, it would have no L3.

Yes, I do think adding another module wouldn't be terribly expensive given the module size for Kaveri with L2 looks something like ~30mm2 with L2 (based on the very rough estimate I did a while back)

 

[cbn]

In my opinion, that is the cheapest/most realistic scenario. It is possible that HDL libraries could get the die size very small, making production per die cheap, but the clockspeed losses could be ugly, and you'd incur the cost of applying HDL libraries to what is basically an extrapolation of an existing design (Kaveri).

Regarding HDL libraries, from what I understand this is something that would be more useful for a mobile SKU.

Which brings up the question, "What will be the primary target of our hypothetical hexcore die: desktop or mobile?" (Assume the die would be used on both FM2+ and laptop)

How competitive would the hexcore chip with HDL be against Intel chips at the same price point for laptop with dGPU? Or should the primary target be desktop with laptop as the secondary target? (Both goals are very interesting to think about IMO)

 

[frozentundra123456]

I get the idea about iGPUs and Integrated graphics on chipset being effective cost reducers. However, with AMD being so far behind Intel with respect to manufacturing node I really question the idea of having so much iGPU. In some ways, I feel like it is really hurting AMD to try and do this.

In fact, I actually believe having too much iGPU could eventually be one of Intel's weaknesses (unless they make a return to having dedicated GT1 dies).

I agree, AMD apus are just in limbo. More gpu than the normal user needs, but not enough to be really powerful relative to a discrete card. So I agree, they need to make something like an FX but with just a minmal igp so the casual user does not have to add a discrete card or live with motherboard graphics. Either that or they need HSA to catch on big time and fast. But it is hard to drive software adoption with a small market share and limited resources like AMD has.

 

[Enigmoid]

A lot of people seem to be saying that AMD should go simply with a tiny (64 core igp) to reduce die size. It is important to remember that adding an igp, regardless of die size, will add a glut of other silicon dedicated to display, media, and decode/encode (H.265).

http://www.anandtech.com/show/8067/amd-am1-kabini-part-2-athlon-53505150-and-sempron-38502650-tested

Given that one core is 3.1 mm2, extrapolating out gives the size of the die at 31.4x the size of a single core, or 97.3 mm2. The GPU area is approximately 5.2x the size of a core, giving ~16.1 mm2 for 128 GCN cores, compared to 12.4 mm2 for CPU cores. The Video Codec Engine and Unified Video Decoder are not part of these totals, located on other parts of the APU. The memory controller clocks in at ~9.4 mm2 and the display/IO portion runs at ~7.3 mm2.

It is quite likely for Kabini that the actual gcn cores are only half of the actual area allocated to graphics; the rest is video engine, display, etc. It is also likely that if AMD releases any 6 core chip it will have greater multimedia capabilities than kabini. Very likely the associated igp silicon will outweigh the gcn cores making a larger gpu more capable in proportion to the total size (ie misc - 20 mm^2, 128 gcn = 15 mm^2). A 64 shader igp would be something like 27 mm^2 but would be half as capable as a 35 mm^2 igp. Thus a tiny igp is not really worth it, IMO the sweet spot would be 256 cores.

These numbers are a rough estimation only.

 

[cbn]

A lot of people seem to be saying that AMD should go simply with a tiny (64 core igp) to reduce die size. It is important to remember that adding an igp, regardless of die size, will add a glut of other silicon dedicated to display, media, and decode/encode (H.265).

http://www.anandtech.com/show/8067/amd-am1-kabini-part-2-athlon-53505150-and-sempron-38502650-tested



It is quite likely for Kabini that the actual gcn cores are only half of the actual area allocated to graphics; the rest is video engine, display, etc. It is also likely that if AMD releases any 6 core chip it will have greater multimedia capabilities than kabini. Very likely the associated igp silicon will outweigh the gcn cores making a larger gpu more capable in proportion to the total size (ie misc - 20 mm^2, 128 gcn = 15 mm^2). A 64 shader igp would be something like 27 mm^2 but would be half as capable as a 35 mm^2 igp. Thus a tiny igp is not really worth it, IMO the sweet spot would be 256 cores.

These numbers are a rough estimation only.

Thanks for posting that.

My only concern would be that going from 128 stream processors to 256 stream processors adds 15mm2 die size. I think even this is probably too much of a cost adder for a chip that is meant to be used primarily with discrete graphics on desktop and mobile.

In some cases maybe even saving the 7mm2 by going 64 stream processors vs. 128 stream processors is worth it. Why have 128sp if 64sp gets the job done for non-gamers?

With that mentioned, we still have to imagine where this hexcore would be positioned for desktop and mobile? If planning to use dGPU, how does a hexcore compare to a Core i3 in laptop? My guess is that the hexcore would need a higher TDP to be competitive with the i3 for a gamer notebook, but maybe a hexcore with HDL helps reduce the power consumption gap we are used to seeing?

So hexcore aimed primarily at desktop with mobile as the secondary target? Or hexcore aimed primarily at mobile with desktop as the secondary target?

 

[cbn]

I think in an ideal world I would rather have a stripped down hexcore on AM3+/ or hypothetical AM4 with integrated graphics on chipset as an option.

Ideally such a set-up (in the case of a streamlined/integrated one chip AM4 format) would be power-efficient enough to use in a dGPU gamer laptop too. (But something tells me that day is probably a long ways off for various good reasons I don't quite understand yet.)

 

[Enigmoid]

Thanks for posting that.

My only concern would be that going from 128 stream processors to 256 stream processors adds 15mm2 die size. I think even this is probably too much of a cost adder for a chip that is meant to be used primarily with discrete graphics on desktop and mobile.

In some cases maybe even saving the 7mm2 by going 64 stream processors vs. 128 stream processors is worth it. Why have 128sp if 64sp gets the job done for non-gamers?

With that mentioned, we still have to imagine where this hexcore would be positioned for desktop and mobile? If planning to use dGPU, how does a hexcore compare to a Core i3 in laptop? My guess is that the hexcore would need a higher TDP to be competitive with the i3 for a gamer notebook, but maybe a hexcore with HDL helps reduce the power consumption gap we are used to seeing?

So hexcore aimed primarily at desktop with mobile as the secondary target? Or hexcore aimed primarily at mobile with desktop as the secondary target?

AMD is using high density libraries so the actual size is probably quite a bit less than 15 mm^2 (I'm not sure that that is also the actual correct size - pitcarin is around 5 mm^2 per CU).

64 stream processors would certainly be cutting it close. No redundancy and the igp would perform at such mediocre levels that it may affect the basic user experience in heavier tasks (only if lower clocked). It would be around BT level which is adequate but not particularly good.

 

[cbn]

64 stream processors would certainly be cutting it close. No redundancy and the igp would perform at such mediocre levels that it may affect the basic user experience in heavier tasks (only if lower clocked). It would be around BT level which is adequate but not particularly good.

Maybe for mobile, 128 sp Beema/Mullins (cat core chips, etc) is a better option. And for higher iGPU needs Kaveri/Carrizo fills in the need.

Then for mobile, this hexcore chip with 64 sp gets used primarily (or even exclusively) with dGPU.

For desktop, the hexcore with 64sp iGPU could be used by itself or with dGPU.

 

[DrMrLordX]

Or are you meaning a unique 2M Steamroller die in which there is no iGPU?

Yes. The first logical step towards a 3M Steamroller for FM2+ would be to tape out a new die (presumably 2M) with the iGPU + supporting structures removed. Then they could expand by adding another module, space permitting.

Regarding HDL libraries, from what I understand this is something that would be more useful for a mobile SKU.

Which brings up the question, "What will be the primary target of our hypothetical hexcore die: desktop or mobile?" (Assume the die would be used on both FM2+ and laptop)

How competitive would the hexcore chip with HDL be against Intel chips at the same price point for laptop with dGPU? Or should the primary target be desktop with laptop as the secondary target? (Both goals are very interesting to think about IMO)

I would honestly be surprised if AMD pushed for any iGPU-less products in the mobile sector. Yes, there are still mobile devices with dGPUs, but that is not the direction AMD is going on right now. They are trying to cram as much of the system under the IHS as they can to make it harder (though certainly not impossible) for OEMs to compromise the quality of AMD laptops with poor build quality. AMD can't count on OEMs to pair their chips with worthy dGPUs.

So, if they did push out a 3M Steamroller, I would expect to see it on the desktop. It would certainly be strange to see 3M Steamroller on FM2+ emerge in this fashion, since usually, such chips trickle down from server R&D. And AMD hasn't launched any Steamroller Opterons.

 

[AtenRa]

FM2+ Socket is able to operate with 100W TDP CPUs, so an 8-core SR or EX on the FM2+ is not a technical problem.
The only reason they dont make more than 2M 4C on FM2+ is because they need to keep alive the 32nm production line for another year. That is the moment an 8-core FM2+ would hit the market, AM3+ platform (motherboards + CPUs) would immediately be rendered dead.

 

[DrMrLordX]

Right, that's what it really boils down to right there. 3M or 4M Steamroller would undercut sales of Piledriver CPUs, and they have stock to dump and fab capacity to use up (yay WSA?). But we've gone over this before.

 

[Shehriazad]

Yea no one would mind seeing 6-8 core CPUs that are 2 "generations" ahead of FX...except for AMD.

Which is sad for the consumer. Maybe they'll throw a "going out of business" party and create sexy chips one last time? Oh wait...that's called Zen. P)

 

[sm625]

It takes a considerable engineering effort to go from 4 cores to 6. A lot of extra uncore needs to be added if you want to keep the latencies down. I'd rather AMD spend their resources profiling javascript and game engines and optimizing their execution units to provide the best possible IPC.

 

[cbn]

With Carrizo being mobile only, I would hope AMD at least has plans to increase volume and do more aggressive prices cuts on the Vishera Piledrivers.

But then I don't know what is going on with Kaveri? Will they stop production of Kaveri when Carrizo takes the big core APU lead? I would think stopping production on Kaveri could help if AMD needed to clear out Vishera chips.

P.S. I'm not up on the WSA, so I really don't how that factors in. How does shifting production from 32nm to 28nm affect that?

 

[cbn]

Right, that's what it really boils down to right there. 3M or 4M Steamroller would undercut sales of Piledriver CPUs

Maybe AMD could clear out the Piledrivers and then transition the FM2+ and AM3+ sockets to newer chips?

Then we could have our three module hexcore on FM2+ and something with more modules (maybe 5 or 6) on AM3+ at a later date?

 

[cbn]

Regarding the discussion of iGPU size for the hexcore, several options were brought up.

I liked 64 stream processors, but Enigmoid brought up some ideas in post #30 on how iGPU sized scaled due to the fixed hardware that also needed to be present (display output, decoders, etc).

For the misc fixed hardware (display I/O, decoders) a figure of 20mm2 was used with each GCN CU coming in at either 7.5mm2 or 5mm2.

In a nutshell increasing iGPU size from 64 stream processors to 128 stream processors only adds 5mm2 to 7.5mm2 silicon. Adding another 128 stream processors (boosting to 256 stream processors total) would add another 10mm2 to 15mm2.

However, one thing to keep in mind is CPU throttling under iGPU load. So even if the silicon area is fairly cheap, there is always decreasing clocks of cpu to consider. And I would imagine six cores would be even more sensitive to cpu throttling under iGPU load than four.

Personally, I don't think going over 128 stream processors would be a good idea for that reason.

Then there is also the intended target of the hexcore, If primarily meant to be used with dGPU on desktop and laptop then having extra CUs just adds more cost.

 

[DrMrLordX]

Maybe AMD could clear out the Piledrivers and then transition the FM2+ and AM3+ sockets to newer chips?

Then we could have our three module hexcore on FM2+ and something with more modules (maybe 5 or 6) on AM3+ at a later date?

It really depends on what wafers they get from GlobalFoundries. You know, WSA and all that. If GF keeps cranking out wafers in fabs tooled for the 32nm process, that sort of limits what AMD can do with them. Stop and think about their entire product lineup, and ask yourself: what could they move to 32nm without messing themselves up somewhere else? Chipset ICs or anything else. AMD can't just idle fabs 'cuz they feel like it, and they can't refuse delivery without paying penalties. So, they're going to have AM3+ Piledriver parts in the pipeline until GF changes the mix of wafers they force on AMD.

All speculation of what AMD could do within their budget aside, that plus the stock-dumping scenario dominate what AMD can and can not do with respect to their AM3+ product lineup.

 

[Erenhardt]

They could use 32nm for HBM dies and use those with 20nm APU for ultimate performance APU.

But HBM chips are fabbed by hynix, so it probably is not possible for amd to licence it from hynix, and manufacture HBM at GF

 

[ShintaiDK]

They could use 32nm for HBM dies and use those with 20nm APU for ultimate performance APU.

But HBM chips are fabbed by hynix, so it probably is not possible for amd to licence it from hynix, and manufacture HBM at GF

Memory is manufactored at an entirely different process node. SK Hynix already uses a 20nm process for it.

 

[DrMrLordX]

I agree that Hynix would probably not agree to such an arrangement, unless Hynix is having fab capacity problems.

If AMD punted on AM3+ and ceased all production (assuming they have not already), then they must figure out what to do with the 32nm wafers for Piledriver FX (and, by logical extension, Richland/Trinity, again assuming they're still fabbing those). They must also figure out what to do with the wafers currently in use to produce chipset ICs for AM3+ boards. For the sake of simplicity, let us assume that they can not merely increase production of 32nm Piledriver Opterons and associated G34 chipset ICs.

What they could do is move FM2+ chipset production onto their 32nm process (more $$$ there), along with AM1 chipset production. It is doubtful that their new Carrizo BGA systems will require many ICs since so much of the board logic is now on the CPU, though fwiw they could move that to 32nm as well. They could punt on the wafers currently used for AM3+, FM2+, and AM1 chipset ICs and pay a penalty. Then, after all that, they'll have to:

Tape out an iGPU-less 2M design for Steamroller on FM2+ using 28nm planar
Tape out an iGPU-less 3M design (with modified NB, as sm625 pointed out) for Steamroller on FM2+, again using 28nm planar

???

Profit!

And that's the cheapest, shortest-term path I could see for them to actually get 3M Steamroller on FM2+. That assumes no iGPU whatsoever. Add that in there, and things get more expensive.

 

[Yuriman]

Am I correct in assuming AMD's chipsets are built on 28nm right now?

 

[AtenRa]

Am I correct in assuming AMD's chipsets are built on 28nm right now?

If you mean motherboard Chipsets no, they are at 65nm.

Kaveri/Kabini/Beema and Mullins APUs are made at 28nm, FX series at 32nm

 

[Yuriman]

Seems reasonable to me for there to be a 32nm chipset revision, to help fill their wafer needs and bring down platform power consumption. FM2 is already competitive with Intel's latest sockets, makes me wonder what the economics are that have kept AMD from doing this already. Raw cost effectiveness is not the only reason one might move to a smaller process.

 

[DrMrLordX]

Seems reasonable to me for there to be a 32nm chipset revision, to help fill their wafer needs and bring down platform power consumption. FM2 is already competitive with Intel's latest sockets, makes me wonder what the economics are that have kept AMD from doing this already. Raw cost effectiveness is not the only reason one might move to a smaller process.

AMD uses fab capacity from older process nodes to produce ICs that don't necessarily need to be from a "cutting edge" (for AMD, anyway) node. Instead of retooling a fab producing 65nm chips, for example, they can just tape out their ICs for AM3+ on that node and produce them there. I suspect that a bunch of 65nm and 45nm ICs are going into AMD motherboards.

So, if AMD moves all their motherboard ICs to 32nm to use up the GF wafer supply from those fabs, then you've got a bunch of wafers from 65nm/45nm fabs that aren't being used for anything. Unless, of course, GF agrees to idle the fabs or retool them.