Die Harvested TSMC Kabini for socket AM1?

[cbn]

The following interaction in posts #78 thru #82 of the "AMD Athlon 5150 Kabini - cat, Jaguar or more fun?" thread ---> http://forums.anandtech.com/showthread.php?t=2378163&page=4 got me thinking of some possibilities for the AM1 socket.

62W with CPU @ 2.7Ghz (+205mV offset, RAM @ 1.65V 2288Mhz)

AMD should be releasing something like that as a higher end option for 2nd/3rd world. For ultimate budget should also include S-video with RCA adaptor

That is an interesting idea. I like the concept behind that, but If 2.7 Ghz (or maybe more realistically 2.2 to 2.5 Ghz) I would definitely say the die could be dual core and with a smaller iGPU.

Same idea goes for Intel.

Smaller die size run at higher frequency, but keep TDP manageable at 25 watts.

(As I understand it this would basically be the opposite of mobile where the die area is larger but the voltage and fequency are both lower.)

I think that I see what you're saying. Why don't they optimize purely for cost, rather than absolute performance / watt. Keep die size small, and up the voltage to up the freq until they have a viable desktop-class part.

Yes that is what I am thinking.

Also assuming the Puma Core SOCs will be compatible with socket AM1:

1.) Maybe AMD could even save up some of the Beema/Mullins chips (built at TSMC) with too many defects for mobile and repurpose them for the AM1 socket (alongside the dedicated DT Puma core SOCs that I believe will be built at GF).

2.) I would even consider a single core Beema with 64 GCN stream processors (packaged as AM1 socket chip, of course) if AMD could boost clocks high enough to compensate. I'm thinking a single Puma core might need a pretty high CPU clock though, but at 2+ Ghz I'm sure it would get the job done for basic web browsing while still being able to come in at 25 watt TDP.

Actually the more I think about it, having the socketed cat cores (ie, AM1) really makes sense when the possibility of using severely die harvested SOCs arises.

Where I would never buy a single cat core SOC as BGA, I would definitely consider if the same single core were socketed (Then at a later time I could switch to dual core or quad core as the need arises).

With this mentioned, the price of the high clocked single core would have to be really good though.

But what if (instead of my speculatation about whether AMD could use die harvested TSMC Beema SOCs for socket AM1) AMD just goes ahead and uses severely die harvested Kabini (from TSMC) and packages them for AM1? (eg, single core with 64 GCN stream processors, dual core with 64 GCN stream processors, etc)

Surely, by this time, AMD probably has quite a bit of TSMC Kabini dies that didn't make the cut for mobile since the TSMC Kabini was launched so far ahead of the DT (Global Foundries) Kabini?

With increased clockspeed they could probably make for some good enough, but inexpensive desktop chips. This while easily being within the 25 watt spec of the AM1 socket.

 

[Abwx]

4C dies will hardly yield three faulty cores, more probably that dies with 3 functionnal cores getting a core disabled and sold as 2C is by far the higher occurence, it s better to throw such very rare single core to the trash bin and sell a 2C for a slightly higher price instead, lower GCN units is not efficient , better to keep the 128 and cut their frequency by a 2 ratio , you ll get better perfs/watt with such an implementation and the same for CPU cores, 2 cores at 1GHz with 80% MT scaling will have better perfs/watt than a single core clocked at 2GHz.

 

[cbn]

lower GCN units is not efficient , better to keep the 128 and cut their frequency by a 2 ratio , you ll get better perfs/watt

I do realize that 64 GCN cores at higher frequency will be less efficient than 128 GCN cores at a lower frequency, but this would be for desktop and not mobile.

 

[cbn]

If the performance per watt for a 25 watt desktop socket is not that bad, and I don't think it would be......

.....then maybe dual core and quad cores with 64 GCN stream processors?

 

[Abwx]

It would be better to allow the user to cut the GPU portion through the bios, factory disabled features are never a good thing and keeping some reserve will extend our items lifecycle, that is a fully configurable chip that you need actualy.

 

[cbn]

It would be better to allow the user to cut the GPU portion through the bios, factory disabled features are never a good thing and keeping some reserve will extend our items lifecycle, that is a fully configurable chip that you need actualy.

If I personally had a 128 stream processsor GCN socket AM1 APU, I would not disable 64 GCN stream processors in the BIOS.

If I paid for 128 SPs I would use 128 SPs.

What I am referring to is AMD selling harvested TSMC Kabinis with 64 GCN SPs at a discount (then packaging them as AM1 chips)

For example, If I had the choice of buying a quad core Kabini with 64 GCN iGPU or a Kabini dual core with 128 GCN iGPU.....I would probably take the quad core (depending on the clocks and price).

Likewise, if I had a choice of buying a quad core with 64 GCN SPs or a dual core with 64 GCN SPs....I would probably buy the dual core (depending on the clocks and price).

However, based on your analysis in post #2, I am guessing the quad core with 64 GCN SPs will be a lot more common. So even if all AMD could release was a quad core with 64 GCN SPs I would probably still be happy enough (but I really think the Kabini dual core with 64 GCN SPs would be the sweet spot for me and hopefully that could be released as well.)

64 GCN stream processors should still make for a good desktop processor for everyday web browsing and other light tasks. I am basing this on the fact 64 GCN SPs appropriately clocked should at least be equal to the 80 VLIW5 stream processors @ 500 Mhz found in the old Zacate E350's HD6310 iGPU.

 

[podspi]

No way I'd go back to a single-core anything, though. I am very interested in seeing how AM1 evolves, though.

One of the big disappointments for me is their sticking with single-channel memory. Kind of rules out a high-performance APU.

 

[cbn]

One of the big disappointments for me is their sticking with single-channel memory. Kind of rules out a high-performance APU.

The single channel doesn't actually bother me. (re: these cat core APUs are just meant to be low cost basic use chips):

[In a perfect world] When AMD does go DDR4, I'd hope they would improve things more on the CPU side than the iGPU side. Then make the DDR4 cat core chips backward compatible with pre-existing AM1 if at all possible. (Eg, including DDR3 and DDR4 memory controllers on die of future chips so it will still work on a DDR3 AM1 board.)

If the future "cat core" SOC includes enough CPU and GPU improvements then AMD can just sell the appropriate die harvested version for DDR3 AM1 boards so the processors are not bandwidth starved. (Eg, Future cat core SOCs with 256 GCN SPs preferntially go to DDR4 AM1 boards while the future die harvested SOCs with 128 GCN SPs will work nicelly with pre-existing DDR3 AM1 boards.)

With all that mentioned, I am still concerned about the direction both AMD and Intel are going with their iGPU plans. In many ways I feel they are focusing too much on GPU and not CPU (This is in sharp contrast to my previous position where I was very Pro OPEN CL).

 

[cbn]

No way I'd go back to a single-core anything, though. I am very interested in seeing how AM1 evolves, though.

I would buy it if it were clocked high enough and were cheap enough. But apparently yields would not be high enough to justify its existence anyway. (see post #2)

Plus another question arises:

How much should a very low end chip sell at retail and does AMD even bother spend money on expensive box packaging (seen below) if the cpu is going to be really cheap?

Maybe AMD should offer the lowest end die harvested chips (like a dual core with 64 GCN SP iGPU) only as a bundle coupled to the most entry level budget motherboards?

An example of a bundle would be the Biostar and Sempron 130 combo kit currently being sold in Newegg embedded solutions:

http://www.newegg.com/Product/Produc...82E16813138377

If AMD could hit $50 for an AM1 Motherboard/ cpu combo I would be more than interested myself.

 

[seitur]

Single core for normal use (office work, programming, workstation, gaming, etc ) nowadays is just bad idea.

I would never buy single core it for this, even if it would be very cheap.

I think most people who remember how it is to work on single thread would agree with me.

No way I'd go back to a single-core anything, though. I am very interested in seeing how AM1 evolves, though.

One of the big disappointments for me is their sticking with single-channel memory. Kind of rules out a high-performance APU.

+1

Yeah. AM1 or it's succesor can be potentially nice & interesting thing.

It really need turbo and dual channel though.

Yes it is cheap, but remember what competition it faces. Intel Celeron&Pentium Core and beggining to show D-Baytrails while costlier are preety close price wise, so I don't think AMD can justify single-channel and no turbo for long.

 

[cbn]

Single core for normal use (office work, programming, workstation, gaming, etc ) nowadays is just bad idea.

I would never buy single core it for this, even if it would be very cheap.

I think most people who remember how it is to work on single thread would agree with me.

Based on my experience with the E1-2100 (1 GHz Jaguar dual core with no turbo, 300 Mhz 128 GCN SP iGPU) I'm sure a 2+ Ghz single Jaguar core would be fine for light web browsing. I am sure it would play old games like CS:S fine too.

For NAS and router, a single core would also work.

+1

Yeah. AM1 or it's succesor can be potentially nice & interesting thing.

It really need turbo and dual channel though.

Yes it is cheap, but remember what competition it faces. Intel Celeron&Pentium Core and beggining to show D-Baytrails while costlier are preety close price wise, so I don't think AMD can justify single-channel and no turbo for long.

I agree turbo is definitely needed, but I actually hope we don't see dual channel on cat core APUs for quite some time (if ever) and here is why:

If AMD decides to put dual channel on cat core APUs they would probably also bloat the die with even more iGPU. A bloated die will just drive up the cost and force the processor to compete with Intel Core based Celeron processors too much. (In fact, this is already happening with the Athlon 5350 and Athlon 5150).

Instead, let the construction core APUs (or whatever big core design succeeds them) have the dual channel memory controllers.