AMD Piledriver rumours ... and expert conjecture

[Reynod]

We have had several requests for a sticky on AMD's yet to be released Piledriver architecture ... so here it is.

I want to make a few things clear though.

Post a question relevant to the topic, or information about the topic, or it will be deleted.

Post any negative personal comments about another user ... and they will be deleted.

Post flame baiting comments about the blue, red and green team and they will be deleted.

Enjoy ...

 

[gamerk316]

And the significance of L3 is...?

Slightly faster then main memory. One could argue if its a waste of die space or not...

 

[noob2222]

No, in order to put AMD's CMT and Intel's SMT on an equal basis, I looked at how they scale up, ignoring 'real', 'almost real' and 'fake' 😛 core differences by comparing a single thread bench to an 8-thread bench for each CPU. There is no Cinebench review that I'm aware of comparing 4 BD cores to 8, which is why I did it that way - apples to apples.

still doesn't add up to what you claim

Intel single core cinebench 10 = 5991, *4= 23964. Actual = 22875. thats below 100%

As far as intel saying 30%, that means you need to do the same as you did with amd to get the same results. 130% for core+smt= 520% for a 4 core cpu. If intel isn't even scaling to 100%, thats less than 400% for core+smt for the cpu.

Now if you want to go and say "but for these 2 Intel's HT performs higher than the advertised max while AMD's CMT lower than the advertised max." you might want to get the analogy correct. if you consider intel's k cpu as 8 threads, then your figures are right, but wrong in the conclusion.

Intel scales 48% to 56% per thread as AMD scales 67% to 73%. (apple to apples)

Thats not Intel performing higher and as I stated, nowhere near +30% boost intel claims, as an "improvement" means that much higher. Intel isn't saying that a core+smt will perform like 30% of a dual core like amd's statement

 

[dogman_1234]

Slightly faster then main memory. One could argue if its a waste of die space or not...

Any L4 R&D? Might consider it for marketing?

 

[lilcinw]

The last thing AMD needs is more marketing bullet points (MOAR COARS!!1!).

They need solid results not talking points *political rant redacted*.

If BD has taught us anything it is that more is not always better. I would advocate less/rebalanced, more responsive cache, not another layer.

 

[ctbaars]

Why is the locked "Bulldozer reviews and opinions, old bulldozer rumors" still Sticky? Lets have positive, forward thinking.

 

[jimmysmitty]

still doesn't add up to what you claim

Intel single core cinebench 10 = 5991, *4= 23964. Actual = 22875. thats below 100%

As far as intel saying 30%, that means you need to do the same as you did with amd to get the same results. 130% for core+smt= 520% for a 4 core cpu. If intel isn't even scaling to 100%, thats less than 400% for core+smt for the cpu.

Now if you want to go and say "but for these 2 Intel's HT performs higher than the advertised max while AMD's CMT lower than the advertised max." you might want to get the analogy correct. if you consider intel's k cpu as 8 threads, then your figures are right, but wrong in the conclusion.

Intel scales 48% to 56% per thread as AMD scales 67% to 73%. (apple to apples)

Thats not Intel performing higher and as I stated, nowhere near +30% boost intel claims, as an "improvement" means that much higher. Intel isn't saying that a core+smt will perform like 30% of a dual core like amd's statement

You still assume perfect scaling, which there never is.

Any L4 R&D? Might consider it for marketing?

I would imagine Intel could call stacker DRAM that if they wanted to.

 

[Tijok]

Hey All,

New to the thread, but I thought I might throw in a bit of info I stumbled across that seems very pertinent to the future of PD, especially in it's Trinity incarnation.

http://www.hardcoreware.net/amd-vs-intel-bulldozer-vs-sandy-bridge/4/

The tl;dr is that BD works better that SB in OpenCL, but more so than that, works better in tandem with an AMD graphics card than even on it's own (when used for OpenCL processing).

So it seems that AMD's strategy in the OpenCL area is already paying off (for OpenCL code) if configured correctly, and it would follow that it will do even better in PD, as the GPU can have even tighter and more coherent integration.

Other note: It seems both PII and SB had a similar (negative) scaling when working in tandem with a GPU, while BD somehow enhanced it's own processing ability in that case. Not sure what black magic that is, but it certainly plays perfectly into a Fusion future!

 

[amdfangirl]

It will definitely be an interesting comparison, but the future PD for desktops should be a better performer than the PD derived core in Trinity, simply because Trinity will lack the L3 cache that desktop PD will have.

Additionally, Trinity's clockspeed compared to future PD desktop CPU's will almost certainly be limited due to Trinity's TDP having to share a GPU, unless AMD decides that all CPU's from here on will have a GPU integrated.

I would argue better GPU and battery life are more important than CPU performance in a consumer laptop.

Wow... this thread has to be setting new records for mod derailment 😛

You should see the Linux section. That amdfangirl is having an ego trip.

Why is the locked "Bulldozer reviews and opinions, old bulldozer rumors" still Sticky? Lets have positive, forward thinking.

Because it turned into a massive flamewar. Hardly civil one.

 

[Chad Boga]

I would argue better GPU and battery life are more important than CPU performance in a consumer laptop.

Okay, but I'm not sure how that is relevant to what you have quoted from me, as that was about trying to gauge how Piledriver might shape up on the desktop.

 

[amdfangirl]

I just stating the perf. of Trinity CPU side doesn't really matter. It is a laptop processor first and formost.

I get to not make sense at this crazy time of day.

 

[gamerk316]

Any L4 R&D? Might consider it for marketing?

To be fair, most of my old books still talk about the L2 as Main Memory...

An L3 kinda makes sense if its a global cache and the L1/L2 are per core, but a private L3 or a public L2/L3 doesn't make much sense to me.

 

[gamerk316]

Hey All,

New to the thread, but I thought I might throw in a bit of info I stumbled across that seems very pertinent to the future of PD, especially in it's Trinity incarnation.

http://www.hardcoreware.net/amd-vs-intel-bulldozer-vs-sandy-bridge/4/

The tl;dr is that BD works better that SB in OpenCL, but more so than that, works better in tandem with an AMD graphics card than even on it's own (when used for OpenCL processing).

So it seems that AMD's strategy in the OpenCL area is already paying off (for OpenCL code) if configured correctly, and it would follow that it will do even better in PD, as the GPU can have even tighter and more coherent integration.

In english: AMD does better in OpenCL because it has the stronger built-in GPU.

 

[-Fran-]

I just stating the perf. of Trinity CPU side doesn't really matter. It is a laptop processor first and formost.

I get to not make sense at this crazy time of day.

Not quite agree there... You *could* argue that notebooks won't ever compare to desktop in a 1 on 1, but saying that the CPU side doesn't matter I really don't agree there.

Notebooks nowadays are more than just for reading the email and doing word processing. They have been broadening a lot more in what you use them for. I won't trade my CPU (in a notebook) for more video card if I want to make use of CPU intensive tasks, right? Llano has a really sweet spot for a "jack of all trades" APU, but if PD doesn't deliver the same performance, even giving more GPU muscle, there's no way it's going to replace Llano.

I'm kind of biased there, though, cause I love to program "on the go" in my lappy, but I'm sure you get my point.

So... Don't say CPU power doesn't matter in a notebook, please >_<

Gaming wise... The power envelope is so low that I'm not sure that PD will deliver more perf than Llano to be honest. The CPU penalty will be a lot IMO and might not be able to deliver the same through output as a whole. BF3 at least, I'm sure will perform quite good though.

Cheers!

EDIT: Missing word

 

[amdfangirl]

My reasoning is that if you wanted CPU-heavy, GPU light you would have bought Intel 😛.

 

[BaronMatrix]

In english: AMD does better in OpenCL because it has the stronger built-in GPU.

He said BD\PD not Trinity. I could have and probably did tell everyone that AMD optimized for new SW and ISAs like OpenCL. Put AVX\XOP under FMAC and you'll see BD.

 

[-Fran-]

My reasoning is that if you wanted CPU-heavy, GPU light you would have bought Intel 😛.

Got an i3 (2.5Ghz) actually, lol. Slapped 8GB in there and it's just like a desktop. It matches (except threaded heavy) to the C2Q at the office at least, which is a lot to say. I can even game on the thing, damn good.

A friend got an A6 + discrete, which I find a lil' redundant, but oh well. It's, like I said, a nice jack of all trades and "cheap" as well (more than my i3, but less than the i5 that comes after my i3). Plays everything and doesn't disappoint CPU wise. It's quite close to the i3 in that regard, so P/P wise it's a stalemate IMO; hard to say which is better at what with the discrete card in it.

My point was, just don't say that the CPU "can be" lesser than the previous gen in terms of through output (I have a hard time writing that... u_u), even more when Intel keeps pushing faster CPUs at low TDP (which is king in mobility). I really believe that if PD can't match Llano from a CPU stand point, it's going to fail. It won't matter the GPU in it, trust me. And I'm not comparing it to Intel's line up at all, it's Llano vs PD I'm talking about. Even more when Llano + discrete would be better than PD + discrete; I'm sure of it. Moreover, when they're built at the same 32nm node. There won't be a significant improvement (if there is) in TDP.

And in that same note, why the hell Notebook vendors haven't been selling more A8s without discrete cards at all? CPU + GPU (discrete) will be Intel's win; at least so far. I find it hard to believe that the A6 + 6670 is not as good as the i3 + GT425M at games when the i3 was/is cheaper! It kinda makes me angry to be honest >_>'

Cheers!

 

[noob2222]

You still assume perfect scaling, which there never is.

Intel is supposed to scale 65% per thread (100% + 30% improvement for smt, wich is 130%/2 since its 2 threads), and scales 48-56% where AMD is supposed to be 80% (module is 80% of a dual core) per thread and is 67-73%

Where on earth to you think I assume perfect scaling considering the numbers are lower than perfect?

Those are REAL WORLD SCALING, and neither are what they claim, but AMD is much closer to the target, not intel. Ultimately thats their goal is to hit what they claim thier cpu in theory (not perfect scaling) can do.


Now think about this. If AMD's core count scaling is much higher than Intel's smt, that means IPC improvements will have that much more impact on BD architecture.

 

[gamerk316]

Now think about this. If AMD's core count scaling is much higher than Intel's smt, that means IPC improvements will have that much more impact on BD architecture.

Correct. Because AMD has more "cores" [again, I REALLY dislike that term when discussing BD], even small IPC improvements can have significant impact, as the effect is magified by the number of cores used.

That being said, you are still at the mercy of SW scaling, and due to the way BD is designed, I don't see significant long-term prospects for IPC.

 

[fazers_on_stun]

still doesn't add up to what you claim

Intel single core cinebench 10 = 5991, *4= 23964. Actual = 22875. thats below 100%

You're assuming 100% scaling to the first physical 4 cores, which is almost assuredly not the case. Compare the 2500K - no HT - with its scores bumped up by 34/33 to get equal clocks. That yields an R10 score for 4 cores of 20,999. So in that case HT yields 22875/20999 = a whopping 9% positive gain 😛. But at least it's a positive gain..

Intel's HT works best with 'light' threads where there's plenty of idle clock cycles that would otherwise go to waste. It has been known since the early Nehalem days that turning off HT for heavily-threaded apps is generally better.

However the whole point of my original post was to dispel the argument that "The die size increases show CMT to be a worthy approach to get this extra perf with minimal die space and power usage.". I think the jury is still out to lunch on that one..

 

[fazers_on_stun]

I just stating the perf. of Trinity CPU side doesn't really matter. It is a laptop processor first and formost.

I haven't looked at any Trinity roadmaps in a long time - are you saying there's no 6 or 8-core versions on it?

Anyway, S/A has a rumor about tomorrow's "new directions for AMD" announcement by CEO Read. Actually it's pretty much the same story as the WIN article on THG's front page. However S/A thinks AMD is making a 'monumental mistake' by going with an ARM license and entering the cellphone/tablet market, due to even more competition and incredibly small margins. But that seems to be what the BoD wants, according to the rumors floating around as to why they fired Dirk Meyer back in January.

One item in the S/A article that caught my attention, was the mention of AMD laying off "all the fusion related engineering not in Austin". I dunno how many engineers that comes to, but it ain't good news for Trinity. So if this turns out to be true, then my guess that Trinity would be delayed will probably turn out to be accurate, despite all the flaming to the contrary from certain individuals 😛.. And, it also may mean that AMD is giving up on at least desktop and possibly server (except for those using hundreds of ARM-type CPUs).

 

[dogman_1234]

And, it also may mean that AMD is giving up on at least desktop and possibly server (except for those using hundreds of ARM-type CPUs).

I am not paying 1000 USD for a stupid Intel processor that is really 200 USD.

 

[fazers_on_stun]

I am not paying 1000 USD for a stupid Intel processor that is really 200 USD.

Exactly. Which is why Intel wouldn't jack up the price like that, despite the Internet's 'carved in stone' premise that Intel would do precisely that if not for AMD's competition.

Intel didn't get to where they are by being stupid. Well, not 100% of the time stupid 😀.

 

[dogman_1234]

Yet...it does not stop them...

 

[AbdullahG]

http://www.guru3d.com/news/amd-1090fx-and-1070-chipset-info-leaks/

 

[JAYDEEJOHN]

Thats old news actually
Dont mean to rain on your parade heheh
I guess theres alot out there thats not in here, maybe someone can change that?