[mczak]

That's the thing I'm not sure there are really "performance bugs". Sure there are performance issues but it's unclear if these are really the result of bugs as such and not just unbalanced architecture. (I would have considered the low li1 associativity a likely candidate for causing performance issues but apparently amd didn't think so...)
Maybe some changes like being able to execute some more instructions in the AGUs go in the right direction, but those are not in Piledriver.

[3dilettante]

I think someone said the leaked die shot showed extra SRAM in the area of the branch predictor. I'm not entirely sure.

The bulk of the improvement seems to be around reducing the power draw of the cores and better clocks. Llano had some notable weaknesses here that Piledriver's more advanced turbo and circuit implementations should improve upon.

Llano's inflexible turbo option doesn't set the bar that high.

[fellix]

Originally Posted by 3dilettante

I think someone said the leaked die shot showed extra SRAM in the area of the branch predictor. I'm not entirely sure.

Yes, that area is where the branch prediction logic resides. Probably the BTB size has been increased, along with other still unknown tweaks.

[eastmen]

I'm going to assume a more mature 28nm process will make up the performance diffrence.

[3dilettante]

I thought Trinity was 32nm.

[mczak]

Originally Posted by fellix

Yes, that area is where the branch prediction logic resides. Probably the BTB size has been increased, along with other still unknown tweaks.

According to the SOG all family 15h processors have the same number of BTB entries.
That's why I said there doesn't really seem to be many architectural changes, the SOG really has quite some details. But yes, better clocks / turbo could help quite a bit.

[Lightman]

Apparently 5%-8% better IPC for new module. Rest of performance jump is made by better power management / Turbo / Clocks.

[almighty]

Originally Posted by Lightman

Apparently 5%-8% better IPC for new module. Rest of performance jump is made by better power management / Turbo / Clocks.

So it'll be on par with Phenom 2 and around 3 generations behind Intel latest and Greatest.

At this pace Intel will be a good 60%+ faster per clock.

[eastmen]

Originally Posted by almighty

So it'll be on par with Phenom 2 and around 3 generations behind Intel latest and Greatest.

At this pace Intel will be a good 60%+ faster per clock.

Well they are saying a 17watt Trinity will be as fast as a 30w Llano. That should put it pretty close to Sandy bridge IMO.

The jury is still out on what increases if any we will see with Ivery bridge

[almighty]

Originally Posted by eastmen

Well they are saying a 17watt Trinity will be as fast as a 30w Llano. That should put it pretty close to Sandy bridge IMO.

The jury is still out on what increases if any we will see with Ivery bridge

Leaked trinity over at overclock.net show that its no faster then Llano

[DavidC]

Originally Posted by almighty

Leaked trinity over at overclock.net show that its no faster then Llano

What, the OBR result? He updated later to say it was a Chinese fake. I've read from somewhere else it was a Bulldozer chip, not Piledriver.

[denev2004]

Originally Posted by mczak

I wonder how AMD expects to get 10% higher performance from Piledriver. According to that Software Optimization Guide, there's really not that many performance enhancing changes. Ok 10% more entries in load/store queue (44 instead of 40) is nice, as is the doubled l1 dtlb size (from 32 to 64). There's also a couple more supported instructions but otherwise it seems virtually unchanged - even keeping the very lame l1i associativity of Bulldozer for both Trinity and Vishera.
Of course you can't directly estimate performance by this guide but since everything seems to be so extremely similar I really don't expect much (except maybe higher clocks).

Not really. It could happens on Latency as well as CR
I hope the 32nm of Piledriver will be a later version. Remark that the first version of 45nm of GF is not that good, while the latest of which is really great.

[AlexV]

Originally Posted by eastmen

Well they are saying a 17watt Trinity will be as fast as a 30w Llano. That should put it pretty close to Sandy bridge IMO.

The jury is still out on what increases if any we will see with Ivery bridge

No, what they're cleverly saying is that perf/watt should be sortof equal...in a pretty specific setting. And the jury is far less out on what Ivy Bridge (please make an effort and at least use proper codenames and proper form in written communication, there's honest mistakes and just sloppiness) than it is on what Piledriver brings to BD. Let's not build up unrealistic expectations again only to have them go kaboom.

[I.S.T.]

The only way a Trinity at 17W could compare to a 30W Llano anyway would be solely in the GPU. The CPU will be slower.

[itsmydamnation]

Originally Posted by I.S.T.

The only way a Trinity at 17W could compare to a 30W Llano anyway would be solely in the GPU. The CPU will be slower.

you know this based on what? bulldozer does very well at low power. quadcore llano at 35watt TDP is only 1.4 ghz. so a little bit of extra IPC over bulldozer and if they can hit 100-200mhz higher base clock and a better turbo then llano i can see it matching.

not saying that it will..... but you speak like its fact when reality is you have NFI.


the 35watt 8 core does have a base clock of 1.6 and a turbo of 2.8 so it could get close.

[Zaphod]

The analyst day slides are a bit inconsistent:

The score for the 2012 AMD A4-4355M (ULV-17w) on the "Pumori" reference design for PC Mark Vantage Overall benchmark is projected to score 3525
The score for the 2012 AMD A6-4455M (ULV-17w) on the "Pumori" reference design for PC Mark Vantage Overall benchmark is projected to score 4200

The former is certainly much less impressive compared to the A6-3400M (35w) on the "Torpedo" reference design scoring 4545 and the E2-1800 (another slight boost to Zacate?) on the "Torpedo" reference design scoring 2757.

[AnarchX]

Testing performed by AMD Performance Labs. The score for the 2012 AMD A10-4600M on the “Pumori” reference design for PC Mark Vantage Productivity benchmark shows an increase of up to 29% over the 2011 AMD A8-3500M on the “Torpedo” reference design. The AMD A10-4600M APU has a score of 6125 and the 2011 AMD A8-3500M APU scored 4764.

http://blogs.amd.com/fusion/2012/03/...9D-generation/

FX-4100 (3,6GHz+, 8MiB L3): 6113

[Raqia]

Originally Posted by AnarchX

http://blogs.amd.com/fusion/2012/03/...9D-generation/

FX-4100 (3,6GHz+, 8MiB L3): 6113

The lack of L3 seems to hurt for some tests.

[denev2004]

Anger Fox Test results for AMD Bulldozer processor

BTW the new version of The microarchitecture of Intel, AMD and VIA CPUs: An optimization guide for assembly programmers and compiler makers has been released

[fellix]

AMD Trinity (Piledriver) Outperforms Bulldozer in Integer and Floating-Point Benchmarks

[fehu]

two a8-4500m with different performance?

[Gubbi]

And a 4.2GHz multi core K8 derivative ?

Cheers

[hoho]

Originally Posted by fehu

two a8-4500m with different performance?

Different work units having different calculation requirements is my guess. Though without having actual links to the work units used to make that table I can't be sure.

[denev2004]

Originally Posted by Gubbi

And a 4.2GHz multi core K8 derivative ?

Cheers

It seems the clock here means the highest TB rate.

[Lightman]

Original source explains why:

Since we don't know the exact clock frequencies of the benchmark runs, it is difficult to find the correct value for calculating per GHz results. I estimated those based on turbo clocks, which might lead to skewed results. At least in the case of comparing Trinity with its Piledriver cores to the FX models, I hope that rather similar turbo mode behaviour should reduce the error margin.
OK, here comes the table comparing several values I filtered out of my collected BOINC results to have OS and client version the same. As you can see, Piledriver w/o L3 cache seems to perform a bit better than BDver1 based FX models:

Note: I used "Trinity vs. Bulldozer" to denote the difference between a L3-less Piledriver core and a Bulldozer core, which always had L3 available.
Another note (as of 04/10): In the Piledriver vs. Bulldozer columns I divided the Trinity value by the maximum of all FX values. Further the FP benchmark likely run at base clock frequency. I'll add more on that in a follow up article.

http://citavia.blog.de/2012/04/08/tr...ance-13460109/