AMD "Piledriver" refresh of Zambezi - info, speculations, test, fans

[FlanK3r]

FX in first generation is behind us, now is coming speculation about refresh of BD1 - Piledriver core. Piledriver will be in one version for notebooks and mainstream products (Trinity) and second as Vishera for "highend" (this info was at xsnews maybe one or two months ago).



http://www.donanimhaber.com/islemci/...-detaylari.htm


But now Im more pesimistic, I will be happy, if AMD show us really 15% increase of performance clock to clock....

[informal]

20% x86 uplift versus Llano now?? We first saw 10-15% uplift figure vs BDv1 and now we see this other number which doesn't make sense IMO. Even with higher clocks versus Llano this "20% more x86 performance" would mean it has more IPC than Llano which is hard to believe. Or they just picked up x264 benchmark with XOP support...

[FlanK3r]

maybe, at slide is the note: "based on AMD projection using digital media workload"

[Olivon]

More slides :




http://www.computerbase.de/news/2011...uer-notebooks/

[ilkkahy]

Im afraid that Llano will wipe the floors with Trinity. Hmm.. I dont remember any socket that had shorter lifespan than FM1? Im sure there has been some rare "special" sockets though but i would imagine that FM1 is in quite many computers by now already.

[informal]

Im afraid that Llano will wipe the floors with Trinity. Hmm.. I dont remember any socket that had shorter lifespan than FM1? Im sure there has been some rare "special" sockets though but i would imagine that FM1 is in quite many computers by now already.

If desktop PD based Trinity can bring 5-6% more IPC,on average, versus BDver1 and if it can Turbo up to 3.8Ghz while staying in 100W TDP range then for sure it can be 20% faster than Llano:
-llano is 15% faster than BD per clock => 2.9x1.15=3.3GHz QC BD needed to match it or 3.3/1.05=3.17Ghz PD.
- to be 20% faster than 2.9Ghz Llano, PD would have to Turbo up to around 3.8Ghz. For 2M part this Turbo core is very achievable IMO. Especially if you think about the TDP of 3850 Llano part which is 100W. Today's 95W 8120 with 4M/8T can Turbo to 4Ghz with 2 modules active. Pretty simple math .

So bottom line: 20% faster than both desktop and mobile Llano is very much possible. Clock+IPC uplift will do the trick easily. 30% uplift in GPU department will be a bit harder to achieve given the memory BW limitations of Llano,but BD and PD have new revamped IMC which is far better than what K10 has (around 35-50% higher mem BW). Trinity will support DDR2133 so mission accomplished.

[Smartidiot89]

If desktop PD based Trinity can bring 5-6% more IPC,on average, versus BDver1 and if it can Turbo up to 3.8Ghz while staying in 100W TDP range then for sure it can be 20% faster than Llano:
-llano is 15% faster than BD per clock => 2.9x1.15=3.3GHz QC BD needed to match it or 3.3/1.05=3.17Ghz PD.
- to be 20% faster than 2.9Ghz Llano, PD would have to Turbo up to around 3.8Ghz. For 2M part this Turbo core is very achievable IMO. Especially if you think about the TDP of 3850 Llano part which is 100W. Today's 95W 8120 with 4M/8T can Turbo to 4Ghz with 2 modules active. Pretty simple math .

So bottom line: 20% faster than both desktop and mobile Llano is very much possible. Clock+IPC uplift will do the trick easily. 30% uplift in GPU department will be a bit harder to achieve given the memory BW limitations of Llano,but BD and PD have new revamped IMC which is far better than what K10 has (around 35-50% higher mem BW). Trinity will support DDR2133 so mission accomplished.

10-15% IPC every new generation starting with Piledriver was what AMD said specifically during our brief of Bulldozer, which you can also see in the first picture provided by Flank3r Much problem lies in the manufacturingprocess from GlobalFoundries (overclocking Bulldozer takes power consumption to new extremes so it's leaky).

http://www.nordichardware.com/news/6...bulldozer.html

Since the first Piledriver-product on the menu is Trinity I'll leave this here:




http://www.nordichardware.com/news/6...confirmed.html

[cesariuth]

More slides :




http://www.computerbase.de/news/2011...uer-notebooks/

what happened with "hondo"?
will be Replace with optimizations of brazos only?
I'm confused

[Smartidiot89]

what happened with "hondo"?
will be Replace with optimizations of brazos only?
I'm confused

Hondo was always an optimized Brazos. It will be two cores @ 1 GHz and 80 shaders @ 280 MHz with a TDP of 4.5W. The chipset will have the biggest change and be severaly castrated, but still more then enough for tablet use

[informal]

@Smartidiot89

Are you sure they explicitly said "IPC" performance when they mentioned 10-15% uplift? And not clock+IPC?

[Smartidiot89]

@Smartidiot89

Are you sure they explicitly said "IPC" performance when they mentioned 10-15% uplift? And not clock+IPC?

I am sure that is what I was told, but I see other media are saying clocks+IPC. But 10-15% increase in IPC might as well be in well optimized software. These are PR-people and after Bulldozer I definitely won't trust it for certain I think 10-15% yearly is a realistic goal, looking at how Bulldozer is today.

Regarding Llano -> Trinity by the way we are forgetting one thing which should also be held accountable for the 20% increase. Llano was supposed to have been a >3,0 GHz design and it launch at 2,9 GHz, again probably due to GlobalFoundries. So the jump from Husky -> Piledriver should be a mix of IPC and clocks.

EDIT: Just noticed your signature, Terrace was right all along

[undone]

I am sure that is what I was told, but I see other media are saying clocks+IPC. But 10-15% increase in IPC might as well be in well optimized software. These are PR-people and after Bulldozer I definitely won't trust it for certain I think 10-15% yearly is a realistic goal, looking at how Bulldozer is today.

BD is a buggy product, we have to regard it as a prototype rather than a chip. BD is another Willamette(Pentium4) IMO
If you take these problem fix into account then 10-15% is too less. But 20% stronger than Llano is somewhat surprise, cuz it put 8 core Piledriver to position that 20% stronger than 8 core Llano(or K10), oh wait should it was BD's goal?

edit: IPC or frequency is not important now, cuz we cannot figure out what BD SHOULD be, not to mention PD.

[Smartidiot89]

BD is a buggy product, we have to regard it as a prototype rather than a chip. BD is another Willamette(Pentium4) IMO
If you take these problem fix into account then 10-15% is too less. But 20% stronger than Llano is somewhat surprise, cuz it put 8 core Piledriver to position that 20% stronger than 8 core Llano(or K10), oh wait should it was BD's goal?

edit: IPC or frequency is not important now, cuz we cannot figure out what BD SHOULD be, not to mention PD.

It's hard to look at Bulldozer as a prototype chip, they already had this at 45nm which was an 8-core Bulldozer. They didn't launch it and instead focused on a 32nm part. AMD f**ked up, big time.

AMD aren't doing these projections based on "how it should've been", but for what it is today. So it is 10-15% based on the consumer products in the FX-series out today. So the increase is over todays Bulldozer. But with process maturity the overall increase in performance should be more. I'd say process maturity should account for the biggest increase in performance over FX-8150 and a coming FX-8170, since AMD have been open about that they aren't happy with where GlobalFoundries is today.

[undone]

AMD aren't doing these projections based on "how it should've been", but for what it is today. So it is 10-15% based on the consumer products in the FX-series out today. So the increase is over todays Bulldozer. But with process maturity the overall increase in performance should be more. I'd say process maturity should account for the biggest increase in performance over FX-8150 and a coming FX-8170, since AMD have been open about that they aren't happy with where GlobalFoundries is today.

It should have been somewhere that if GlobalFoundries do well in 32nm, but GF miss the target, I still believe BD architecture is good, and wait new revision(B3/C0) and Piledriver to be released.

[Manicdan]

im curious about the chip size and transistor count of trinity vs llano. zembezi is quite large for not being much faster than thuban, would be good if they can pack in those compute units into a tighter space and save some wafer costs.

[mrcmtl]

What will also be interesting is to see how PD actually performs without the massive L3. The 30% increase in GPU performance could be due to the increase in SP by switching to VLIW4.

[FlanK3r]

If clock will be simillary, 10% is "good", But Im a bit pesimistic...maybe max 5% and other will be higher clocks for Piledriver.

[Mechanical Man]

If its faster than quadcore Llano clock for clock with two modules, im surprised. With change to vliw4 arch it will get more efficiency from same area but memory bw increase is where most of gpu potential is. 30% for GPU should be easy.

[informal]

If its faster than quadcore Llano clock for clock with two modules, im surprised. With change to vliw4 arch it will get more efficiency from same area but memory bw increase is where most of gpu potential is. 30% for GPU should be easy.

I don't think this will happen. It may be faster in few specific tasks,but overall I think it will end up being slower than Llano. This is not such a terrible thing since PD will auto-clock like crazy and this will make it a much better APU then Llano is. Add in new ISA support,covering all available instruction sets,new power management and 30% better GPU performance and you get a real IB challenger. GPU will be a clean victory for Trinity,IB will be no mach ,not by a long shot. But CPU side will be noticeably slower at similar clock. It depends how high PD (and IB) can clock so in the end it may come close I guess.

[Mechanical Man]

I don't think this will happen. It may be faster in few specific tasks,but overall I think it will end up being slower than Llano. This is not such a terrible thing since PD will auto-clock like crazy and this will make it a much better APU then Llano is. Add in new ISA support,covering all available instruction sets,new power management and 30% better GPU performance and you get a real IB challenger. GPU will be a clean victory for Trinity,IB will be no mach ,not by a long shot. But CPU side will be noticeably slower at similar clock. It depends how high PD (and IB) can clock so in the end it may come close I guess.

Yes there was that magical "im surprised if"

IB is hard one, if of 77W TDP for high end are true, low end will have awesome TDP too.. With hybrid gpu technologies (was it optimus for nVidia and something other for AMD) its hard to beat. Ofc low price will be where AMD has stand, but also low margins...

[TESKATLIPOKA]

Mechanical Man

IB is hard one, if of 77W TDP for high end are true, low end will have awesome TDP too..

Certainly, IB will be a TDP beast so we can expect higher clocks and not to forget a minor IPC increase, but its not so bad for Trinity, the only thing they need is high clocking ability and some IPC increase. The problem with mobile Llano is the lack of frequency but if Trinity can operate on a considerably higher frequency lets say 2.4Ghz for a 2module in the same TDP as Llano then at least in multi it should be quite close to Intel because this will go mostly against 2core+HT.

With hybrid gpu technologies (was it optimus for nVidia and something other for AMD) its hard to beat.Ofc low price will be where AMD has stand, but also low margins...

You would need a pretty strong Nvidia GPU to beat this one and hardly every notebook has a discrete gpu much less one performing better than this one, not to mention Nvidia won't sell them for free either, its like comparing Atom+ION2 against Brazos and I think its more or less pointless. The cost for the cpu and mobile GPU will be certainly higher than for Trinity also you need just one chip instead of two and this will be really tempting for OEMs, even if they have lower margins for one sold APU they will sell more. BTW I think enabling Eyefinity is a tempting feature to boost the sales.

[PatRaceTin]

what next chip set series?

900FX to 1000FX ?

[FlanK3r]

for Piledriver, think, still 990FX, but I hope for new mobos

[Manicdan]

L3 has no purpose for client uses?
then why is Phenom stronger than Athlon with the 45nm chips?

[Smartidiot89]

L3 has no purpose for client uses?
then why is Phenom stronger than Athlon with the 45nm chips?

That was true back then, but this is a whole new architecture. We shall see, but derivates will most likely have different dies for consumers and servers where things such as cache (and unnessary amounts) will be adjusted depending on market.