AMD - Piledriver chitchat

[DanceswithUnix]

... just over £150 will grab you 768 shader's worth of dual graphics performance. Should be quite tasty

£108 for 896 streams: http://www.scan.co.uk/shop/computer-...-(896-streams)
£61 for quad core CPU: http://www.scan.co.uk/shop/computer-...ore-socket-fm2

Why wait

Seriously though, the shrink to 28nm gives them more space to play with, perhaps they can play with the cache sizes to mitigate the lack of memory bandwidth. Personally, I would prefer to see another 2 cpu cores in there though.

[CAT-THE-FIFTH]

The problem is February means Kaveri has been massively delayed and unless Intel 14NM has been massively delayed too it will be barely 6 to 8 months later,Intel releases Broadwell for laptop. Trinity was delayed by a quarter and Kaveri should have been out by summer 2013 or Q3 2013 at the latest. So its been delayed by 5 to 6 months now.
These repeated delays are killing AMD. Llano was delayed. Trinity was delayed and so was Kaveri.

I am not sure how they will be able to compete with a two node disadvantage in mobile. Its bad enough now with a one node disadvantage. The HD4600 is already getting close to performance of Trinity in mobile due to TDP limitations. The thing is 28NM is also only a half node shrink and appears to be bulk not SOI too,and neither does it use finfets either.

This is absolutely horendous execution by both AMD and GF,and the stupid WSA agreement is not helping either. I don't see 384 shaders as being enough to compete TBH in mobile for gaming purposes.

Look at the benchmarks of the HD7730 GDDR3:

http://www.tomshardware.com/reviews/...ew,3575-4.html

Thats 384 GCN shaders at 800MHZ. Its only 5% to 15% faster than an HD6670 GDDR3 at the same clockspeeds.

The A10 IGP is HD6670 GDDR3 level performance anyway. Even if AMD were to bump the clockspeed up and improve the performance of the shaders and say make it 25% to 35% faster than an HD6670 GDDR3,I am not sure if in TDP limited conditions whether it will beat Broadwell.

AMD really need to have got it out this year,otherwise Kaveri based laptops will be within 6 months pushed to bargain basement status. They better hope for their sakes Intel has big delays at 14NM or only ramps up some parts next year on mobile,ie,ultrabook parts and the like.

[Dooms]

I'm way behind on AMD tech... These new chips are they just laptop/HTPC chips with integrated graphics or are they also doing new high end?

Reason I ask is I've been looking at the 8350 as my next CPU and it would be a bit pointless if something else is coming out.

[DanceswithUnix]

The problem is February means Kaveri has been massively delayed and unless Intel 14NM has been massively delayed too it will be barely 6 to 7 months later,Intel releases Broadwell for laptop. I am not sure how they will be able to compete with a two node disadvantage in mobile. Its bad enough now with a one node disadvantage. The thing is 28NM is also only a half node shrink and appears to be bulk not SOI too.

Intel is bulk too, doesn't seem to have harmed their parts. It sounds like SOI it just a tech that has had its day. It adds to the mystery of the part, but I'm not going to get excited by that.

Edit to add: They don't make nodes like they used to either, so unless you are hoping AMD will catapult into the lead with this part (which would stagger me, and I think you too ) then I don't think 14nm is all that important.

I'm way behind on AMD tech... These new chips are they just laptop/HTPC chips with integrated graphics or are they also doing new high end?

Reason I ask is I've been looking at the 8350 as my next CPU and it would be a bit pointless if something else is coming out.

I really don't think we know what is happening on their enthusiast platform. They promised another release on AM3+, but they could claim that their recent 220W devices cover that (though I'm sure that isn't what they originally meant).

I suspect they will have to release the new core as an Opteron at some point, where people are going to want more than 4 cores and don't care for fancy graphics. That implies an AM3+ or similar (AM4?) socket part, but I haven't seen any clear roadmap.

That is why I just went for an 8350, couldn't be fussed waiting.

[CAT-THE-FIFTH]

Intel is bulk too, doesn't seem to have harmed their parts. It sounds like SOI it just a tech that has had its day. It adds to the mystery of the part, but I'm not going to get excited by that.

The problem is the continued delays are going to kill them. We have seen how long it takes AMD to translate launches into product wins in mobile. Look at Jaguar. By the time tablets have appeared,and more laptops have appeared,a whole load of BT parts will have swamped the market sadly.

If desktop parts are this delayed,how long for mobile parts then?? Desktop parts don't need to be as highly binned.

This is why the announcement is very worrying. It indicates they can barely get less stringently binned desktop parts out in volume after months of delays.

I smell another Llano style debacle.

By the time we start to see Kaveri based laptops,Broadwell will only appear a few months later. I really,really hope Broadwell is delayed in 2014.

28NM is not even a full node shrink FFS. Its half node. GF is pile of crap TBH. It took until the middle of this year for them to have a small ARM based chip made on their 28NM process. TSMC had 28NM out in late 2011 and with 300MM2+ GPUs.

The worst thing it is a bulk process too,not with the added complications of using SOI. So it means they are so incompetent that they cannot manage to do a half node shrink, without the complications of using SOI on time.

The 28NM refresh of Brazos was meant to be produced at GF IIRC,and it was cancelled as it was not ready.



That is the 2010 roadmap regarding process nodes. It has been delayed by at least 15 months for 28NM and the fact that the larger X86 parts have been delayed even longer probably means it probably is broken in some way. That makes it nearly two years since AMD was supposed to produce products on the GF 28NM process.

14NM against 28NM is going to be messy. Its been bad enough with 32NM against 22NM.

AMD will be more or less finished in the laptop arena at this rate,unless they just want to make bargain basement laptops. Their one main advantage,ie,graphics performance will be over IMHO.

Intel even with a worse GPU design will win out just via a process node advantage,due to the utter incompetence of GF.

Kaveri should have been out NOW not next year. Intel will literally have a two node advantage in mobile. Finfets seem to show more improvements at lower voltages typical of mobile devices. Intel used finfets instead of SOI BTW. Its mentioned that was the two choices they had and they considered per wafer costs with finfets were lower,but then they can afford to spend so much money in R and D due to volumes.

Intel is showing big improvements in IGP performance each generation and I am not sure Kaveri will have much of an impact as it has just been too delayed too long now.

At this rate Broadwell will match or exceed the top end 384 shader AMD IGPs in mobile chips.

AMD needs to become more relevant in the mobile arena.

Regarding SOI,the current 32NM process uses PD SOI and from what I gather FD SOI still has advantages. The problem is cost,and I am uncertain GF wants to bare it anymore for AMD sadly,although their delays are also a big problem. They seem to be more interested in low power processes for mobile chips and a number of ARM based SOCs have already been produced using the GF 28NM process,but then even with all the financial backing they seem to be unable to do anything on time and they are dragging down AMD with them.

Edit!!

It gets better:

http://semiaccurate.com/2010/10/13/g...20nm-gtc-asia/



http://www.brightsideofnews.com/prin...s-by-2015.aspx



Bahahahaha!!

[Noxvayl]

It is no wonder AMD bit the bullet and got rid of Globalfoundries, the sooner they are able to get away from them the better it seems.

[watercooled]

The release time-scale is close to what's been implied for a while; 'release' late this year but retail availability 2014. Although I was more expecting January TBH.

AFAIK SOI is still a useful technology but the wafers cost more than bulk CMOS, however the combination of a half-node shrink and lack of SOI/FinFET makes me wonder if 28nm will really be much of an improvement over very mature 32nm SOI? Although, the node 'sizes' now don't really tell us much about feature size, so it's possible it's more comparable to 22nm - it wouldn't be the first time 'larger' half-node sizes have been denser than Intel's nodes IIRC. (Edit: also, the fact GloFo 28nm has taken this long past 32nm, and the fact AMD still want to use it, suggest it must be some improvement at least).

As for GloFo, I wonder if AMD are still tied to them in some way? I know there are potential capacity problems, especially for AMD's relatively large dies (vs ARM), but why not jump to TSMC (who seem fine with huge GPU dies) or Samsung for example?

Looking for positives, it seems Kaveri will use SteamrollerB - maybe those strange die shots we've seen recently, which looked to have Excavator features, are legit? Not saying I think the delay will be worthwhile, but maybe it will help? And, just because there's been a stall on this node doesn't mean it will carry on. Especially if GloFo/TSMC are offering a decent 20/14nm process as planned, and with AMD supposedly making their designs more portable, we might see them used by AMD more quickly.

[DanceswithUnix]

I think a lot of those old process node predictions were based on EUV lithography being "just around the corner". Well it turns out to have been one hell of a long corner, and we still aren't around it. So instead people are doing all sorts of exotic multi-pattern steps just to expose a mask. So yes Intel are late with 14nm, we don't know just how late yet.

My understanding with SOI is that it doesn't work down below 32nm, so you end up having to use finfets anyway. At that point, you may as well ditch the insulator as it is just extra cost.

Putting all their eggs into the TSMC basket may not be the best way forwards in the long term.

In the mean time, it seems AMD are sweetening the deal and offering games with their CPUs.

http://www.brightsideofnews.com/news...-for-2410.aspx

[CAT-THE-FIFTH]

Its part of an ongoing promotion as SC4 was available from some retailers even with the A4-4000.

It seems they have expanded it to more CPUs now and included another game. It does make the FX6300 even better value though!

[watercooled]

I agree, it would be best to avoid relying entirely upon TSMC long-term, but AMD need competitive CPUs out short-term to stay competitive. I know TSMC aren't part of CPA but I wonder if AMD would consider multi-sourcing in future to avoid delays?

[CAT-THE-FIFTH]

The problem is that a few years ago AMD signed a WSA agreement(which lasts like a decade IIRC) with GF,and this has been the main source of its losses,and delays. The people before the current lot in charge of AMD really made some very short sighted decisions indeed.

Kaveri was demoed in May this year,so it shows you how long AMD has had working samples.

[DanceswithUnix]

Sadly multi-sourcing really doesn't work as a short term fix. AMD have been moving their designs to ones that are more synthesized and less hand layout designed, but there are still major efforts involved in getting a design ported over. People keep saying figures like "$1M" for setting up a mask for a big chip, and the whole process would probably take 6 months. By that time, the design would be outdated and never make the money back.

If they did something like a 6 core/128 shader FM2 Phenom III chip at GF, and a 4 core 896 shader part at TSMC then that might make more sense. They would have to plan it from day 1 though.

I remember the Sim City offer a while back, though that was long gone though. Far Cry would interest me more.

[DanceswithUnix]

Charlie talks about the new GPU architectures:

http://semiaccurate.com/2013/10/23/l...-architecture/

Interesting comment in there that caches are tweaked so that tessellated stuff stays on chip to avoid using the memory interface. That would help a lot in an APU too.

[CAT-THE-FIFTH]

Some more info about Mantle:

http://www.dsogaming.com/news/pcars-...ng-draw-calls/

[watercooled]

Just spotted something in a new Anandtech article - BF4 will apparently utilise the Mantle API. I've probably just missed news of it, but it's nice to see it's being used this soon, especially in such a big game. I guess it will have both D3D and Mantle 'modes' so we can compare performance - although there could still be room for further optimisation over first-gen titles (especially with such a low-level API)?

AMD again confirmed Q4-13 shipping and Q1-14 public availability for desktop Kaveri, which OFC does allow Feb. Still, we might get some reviews this year.

[scaryjim]

From what I've read BF4 will be released in DX11 first, then have a Mantle update for PC a few weeks later. Mantle will work on any GCN card, so 7-series from the 7750+, R7/R9, and Temash/Kabini/Kaveri APUs. While I'm sure most of the review focus will be on how it affects performance on the R9 290X, I'd be more interested to see it tested on an R7 250: surely the place that a low-level API will help out most is no lower end hardware?