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

[Manicdan]

for a laptop i think the most important parts for the cpu is idle, and low thread count perf/watt
very few people load up 4 cores with a laptop, some games might, but most still just use a few threads, and if they use more than 2 the extra are like half used is most cases

im just going to wait for the real reviews to pop up and see how well they improved the total 35w package.

[xdan]

So same process
same TDP

resulted in:

better cpu performance
better idle power consumption
better gpu performance

clearly a fail...

Trinity is a major step forward compared to llano.

Yes 12% with about 30% higher core frecuency. And if it ends like BD which meant major power consumption and TDP increase at a slight overclock, what then.
A bad IPC is a bad IPC no mather you put it.
Let say oc on Trinity from 3.8ghz to 4.5ghz, that's about 20% but on Llano performance it's like raising with 7% the frecuency.
Let's compare again so A8 3850 at 3250Mhz(+12%) should equal Trinity at 3.8ghz(don't even count TURBO).
Furthermore A8 3850 at 3500Mhz( 3250Mhz + about 7%) should be equal with Trinity Cpu at 4.5ghz .
The IGP part it's irrelevant, it's clear a good path as in desktop Gpu's.

So clear from cpu side a FAILDOZER.
AMD in 9 months was not capable to make some tweaks to Llano sillicon to make it do 3.2-3.6ghz(like Phenom II) with overclock to 4-4.2ghz. It should have been a much better choice.
Llano oc >= Trinity oc.

3DMark 06 CPU Score:
A10-5800K: 4304 Marks
A8-3850: 3814 Marks

FAIL !!!

http://hwbot.org/submission/2193039_...50d_8027_marks
A8-3850 @3.321Mhz - 4577.

[flyck]

Yes 12% with about 30% higher core frecuency. And if it ends like BD which meant major power consumption and TDP increase at a slight overclock, what then.
A bad IPC is a bad IPC no mather you put it.

IPC is irrelevant without clock and clock is irrelevant without ipc.
llano is hardly reaching 3GHz from a product point of view. 3.5GHz is waay out of its reach... That can only be obtained by a huge amount of voltage at this point. (lThey can't even manufacter the 3GHz part in decent quantities with such a high voltage).

Let say oc on Trinity from 3.8ghz to 4.5ghz, that's about 20% but on Llano performance it's like raising with 7% the frecuency.
Let's compare again so A8 3850 at 3250Mhz(+12%) should equal Trinity at 3.8ghz(don't even count TURBO).
Furthermore A8 3850 at 3500Mhz( 3250Mhz + about 7%) should be equal with Trinity Cpu at 4.5ghz .
The IGP part it's irrelevant, it's clear a good path as in desktop Gpu's.

So wait. increasing one cpu by 20% results in the same improvement as increasing another cpu by 7%? You do know this is wrong from every metric possible? (ipc is already included in the frequency and thus the frequency increase...)
llano needs just as many % frequency increase as bulldozer to keep the same level of performancedifference.

so when using valid mathematics:
2.9*1.11*1.2 = >3.8GHz. Thats using your metric for trinity... llano would need to run at 3.8GHz or higher... a clockspeed which is an exception to be reached in overclocking...
(also note: your first calculations for trinity are also wrong since you wanted to discard turbo.)

So clear from cpu side a FAILDOZER.
AMD in 9 months was not capable to make some tweaks to Llano sillicon to make it do 3.2-3.6ghz(like Phenom II) with overclock to 4-4.2ghz. It should have been a much better choice.
Llano oc >= Trinity oc

FAIL !!!

Clearly the current 32nm gate first approach doesn't allow stars to reach the clockspeeds it would have reached on 45nm. They would need a ~20% clock improvement while lowering the voltages by just as much to compete with trinity. So clearly they have no choice but to go with the bulldozer design, because it is clearly better. It is able to outperform the previous generation on 45nm, where the previous generation on 32nm doesn't stand a chance.

Does that make llano/bulldozer/trinity great cpu's? No they lack due to huge constrainst by the process (not related to yields). But trinity and even BD fare alot better than the ph2 derivatives on the same process node.

[BeepBeep2]

Llano is not the same "STARS" as Phenom II.
Piledriver isn't Bulldozer either. They are similar, yes but are not the same.

Lets face it, if you put Llano + L3 in AM3+ it probably would have been better received than Bulldozer.

Outside of this, we aren't really comparing apples to apples, more like apples to pears.

[xdan]

IPC is irrelevant without clock and clock is irrelevant without ipc.
lano is hardly reaching 3GHz from a product point of view. 3.5GHz is waay out of its reach... That can only be obtained by a huge amount of voltage at this point. (lThey can't even manufacter the 3GHz part in decent quantities with such a high voltage).

Yes and BD completes all these points.
I mean it cannot reach SB frecuencys on air 24/7 4.8ghz-5ghz(some good SB) and it's IPC it's much worse than Phenom II.
In fact for 24/7 BD no much better than Phenom II. So what you can do with FX 8150 4.6 24/7 without raising to heaven power consumption and latest C3 revision X4 965/970 did 4.3-4.4ghz 24/7. So 200-300Mhz for nothing.
Remember Phenom C2 X4 920 2.8ghz, and C3 revision X4 965 3.4ghz?
So what on 32nm it's not possible a revision to improve clocks and voltages?

So wait. increasing one cpu by 20% results in the same improvement as increasing another cpu by 7%? You do know this is wrong from every metric possible? (ipc is already included in the frequency and thus the frequency increase...)
llano needs just as many % frequency increase as bulldozer to keep the same level of performancedifference.

Yes that's kind of difference in IPC for 12% Llano performance increase you need 30% Trinity increased clocks - 900Mhz(2900-3800) it's 30% increase above 2900Mhz Llano clock) . It's not 7% it's my mistake it's 8%.
12/30% stands with 8/20%.

[BeepBeep2]

Good BD can reach 4.8-5 GHz. The problem is, there are hardly any good BD.
I think by the end of 32nm we may see PD doing 5.2-5.4 GHz stable on air/water. We've already got BD doing 4.6-4.8.

First Phenom II's were lucky to do 3.8 stable, we quickly made that 4 GHz and then 4.2, then really good chips near the end doing 4.4.

IMO if PD Vishera can get 15% more performance in same TDP as BD plus some top end speed I'll be happy. (Apparently resonant clock mesh can lower power consumption up to 10% on the whole)

[xdan]

Continuation....

so when using valid mathematics:
2.9*1.11*1.2 = >3.8GHz. Thats using your metric for trinity... llano would need to run at 3.8GHz or higher... a clockspeed which is an exception to be reached in overclocking..

I'm sure my math it's good.
It's all about parity in IPC Llano(Athlon) / Trinity AT 4.4-4.5ghz (BD, Pilledriver whatsoever).
http://hwbot.org/submission/2264266_...d2_13015_marks
3dmark 2006 Cpu score
A8 3870K @ 3.776Mhz - 5261 marks

A10 5800K 3.8GHZ - 4304 marks
FX 4100 @ 4.5ghz - 4728 marks

With Pilledriver core but without L3 could be 5000-5100marks at 4.5ghz which is about 18-20% above 4300 score as is frecuency.

Anyway you put it Llano at 3.6-3.7ghz should be equal to Trinity. Llano 3.5ghz 24/7 it's possible in general.
And you don't need to change motheboard to new socket FM2 and so on.
9 months from Llano were enough to make a new revision with improved clocks and better IGP.
And it isn't any prove that Pilledriver with IGP will do better clocks. And because of IGP overclocking could be affected.
I mean 4-4.2ghz could be but 4.5ghz > not that sure.

[informal]

Llano is not going to hit anywhere near 3.5Ghz with onboard GPU. The design has no room to grow,let it die.
Trinity with no L3 cache already shows pretty solid improvements(6-10%) over BD parts with L3. When Vishera comes,it will have no onboard GPU so there is room for 2 more modules and L3. Around ~10% over BD parts with Vishera and around 10-15% higher clocks (Vishera vs FX8150) could put AMD in pretty neat spot and that is 30-35% faster than 1100T overall. There is no doubt that it should have been BDver1 to have these performance characteristics but better later than never. If we get another 10-15% in IPC with Steamroller and if 28nm gets AMD near 4.5Ghz mark in clock speeds, they will be matching SB-E's single thread performance(stock vs stock) without any problems.

[Smartidiot89]

that's funny trinity was support to have 25% high cpu vs bulldozer. but bulldozer was 15% slow then phenom II and abuot 35-45% slower then sandybridge.

Improvements will come from both IPC improvements and higher clock frequencies. The difference will be about ~10% for desktops from Llano, while for notebooks it will be about ~25% if i remember correctly.

[Kristoferr]

But Vishera from Bulldozer?

[AlleyViper]

OT: Funny, my 754 @2.6GHz from 04' does 989 CPU on 3DMark 06. That's about 1/4 stock 3.6GHz FX4100.

[tbone8ty]

as long as vishera can hit higher clock speeds (~5-5.2ghz on air ) at much lower voltage/power consumption ill be happy.

plus i think somewhere i read its a different version of piledriver in the desktop chips(vishera) than trinity

[FlanK3r]

FAIL !!!

http://hwbot.org/submission/2193039_...50d_8027_marks
A8-3850 @3.321Mhz - 4577.

u can not compare OC 3850 with stock (and we do not know RAM clocks at Trinity) Trinity...If you make OC at Llano, for best 3DMarks u need: Windows XP (XP score is higher than windows7 for 3D01-3d06), high clocks at memory, high clock integrated NB part. So 3.2 GHz Llano "stock" is not as 3.2 GHz Llano tweaked.

[Kristoferr]

Those differences are minor. We are not HWbot.org scoring for points here.

[xdan]

A8-3850 @3.321Mhz - 4577

Memory is at 1832 CL 9-11 so no big deal, 114 FSB, OS Windows 7 Ultimate so the score is pretty normal, no major tweaks.

[Jethro]

http://www.techpowerup.com/163887/Tr...for-Clock.html

[BeepBeep2]

OT: Funny, my 754 @2.6GHz from 04' does 989 CPU on 3DMark 06. That's about 1/4 stock 3.6GHz FX4100.

Stock 5600+ with DDR2-800 gets 2050 in 3DMark06.

[AlleyViper]

Well, that's expected for a 2.8GHz X2 A64 w/ dual channel DDR2, given my example was a single core 2.6GHz 1MB, single channel that scores about half. Btw, that 3D06 score was on W7 x64 with crappy timing 3x1GB UCCC at 2T; my older 2x512MB BH-5 and XP just aren't usable nowadays.
Still, it's awful to see better thread performance per clock on 9 year old cores under some circumstances.

[BeepBeep2]

I just think its funny how FX scores though...

Barely beating out A64 in some things

[informal]

Yeah but FX4xxx is just a half of bulldozer. It was designed as 8 threaded machine,cutting it in half is going against the initial perf. targets . Of course with 2 FPUs it(FX4xxx) can't score so high. Remember that it has 2x256bit FPUs that are shared between 4 integer cores. In case of X4 chips(llano/phenom) you have 4 cores each having 2x128bit pipelines on their disposal. Bulldozer was designed for integer processing and it's very obvious when you see its FPU. Let's hope SR core brings us at least some improvement in that regard.

[Darakian]

I have high hopes for piledriver. I've got a real use for 8 or more integer cores.

[informal]

Some interesting news about AMD's future core decisions from Vr-zone.

Talking about moving forward, the feel is that AMD is there to survive and thrive once again. Rather than repeating the bland statements of 'we don't want to compete in the high end, just on volume' which make no sense since a fabless company cannot compete on volumes against a competitor with seven large fabs at its disposal, this time there's clear indication that, after Piledriver, there will be substantial changes in both cores and system architecture from Steamroller onwards, that should help make AMD competitive closer to the top. I was told that delaying the socket migration beyond the AM3+, C32 and G34 to new socket is a good move,since AMD can design more aggressive, rather than stop gap, sockets for future platforms with better features like more memory and HyperTransport channels, as well as integrated PCIe v3, for greater future scalability. For the first time, some execs do acknowledge that Bulldozer approach may not have been the best one at the time, and things need to change. I was told that there is some good frequency scalability in the Piledriver core which should help gain some per-core performance ground.

So, in the near term, AMD will use APU to keep its presence in desktop and mobile market, and even low power 5 - 10 W part derivatives or Trinity may arrive for high-end HD++ tablets. The CPU core radical refresh is expected to complete within two years from now, along with brand new socket platforms, proving AMD a new base from which to attack the high end, again, just like in the good old Opteron/Althon 64 early times. The GPUs will continue to be the crown jewel of the company till then, though... expect new high end mobile HD7900 series later this month, and Sea Islands by yearend.

[Tomasis]

I long back to good old Opteron/Althon 64 early times

[haylui]

Trinity A10-4600M PassMark score?
http://www.chiphell.com/thread-446793-1-1.html

[Kristoferr]

Must be some bug in those tests. I do not belive that it can take on 2600K on any of CPU tests.