Kabini Rumors

[Exophase]

The scores are incredibly broken even neglecting the uselessness of the benchmark.. how does a 2.4GHz Core 2 Duo score twice higher than a 2GHz Core 2 Duo?

 

[Centauri]

Pro tip: this is an English-language forum.

...warum kannst du nicht Deutsch lesen?

 

[SiliconWars]

The scores are incredibly broken even neglecting the uselessness of the benchmark.. how does a 2.4GHz Core 2 Duo score twice higher than a 2GHz Core 2 Duo?

You need to relax your accuracy detection a little bit and just use some common sense. Look at the numbers and discard the ones that don't make sense instead of making them proof of the whole benchmark being broken.

With that done you can work out where Kabini stands. It's not at Core level yet despite what the AMD fans want to see. 4 Jaguar cores will beat 2 Core cores at the same clock speed, and that is where AMD is marketing it - vs pentium and i3.

 

[Idontcare]

You need to relax your accuracy detection a little bit and just use some common sense. Look at the numbers and discard the ones that don't make sense instead of making them proof of the whole benchmark being broken.

Yeah, cuz that approach doesn't reek of confirmation bias, not at all...

People display this bias when they gather or remember information selectively, or when they interpret it in a biased way. The effect is stronger for emotionally charged issues and for deeply entrenched beliefs.

 

[monstercameron]

Are the variances due to different kernels?

 

[SiliconWars]

Yeah, cuz that approach doesn't reek of confirmation bias, not at all...

You can just look at the actual numbers and see a clear pattern where faster cpu's beat slower ones.

When I wrote about discarding the results I meant the kind of results that exophase mentioned - not discarding the results that made one side look better. If a 2.4 GHz cpu beats a 2 GHz cpu one of the results is probably wrong yeah? - so you need to look at the results around it to see what is more likely to be wrong. Did i say common sense? Maybe it's not as common as I wish it was.

 

[Idontcare]

You can just look at the actual numbers and see a clear pattern where faster cpu's beat slower ones.

When I wrote about discarding the results I meant the kind of results that exophase mentioned - not discarding the results that made one side look better. If a 2.4 GHz cpu beats a 2 GHz cpu one of the results is probably wrong yeah? - so you need to look at the results around it to see what is more likely to be wrong. Did i say common sense? Maybe it's not as common as I wish it was.

If some of the data are suspect then the method employed in generating the questionable data is likewise suspect, and that is relevant because it brings into question the validity of all the data, including the data you wish to be able to assume are correct.

If I tell you Pi() = 2.756, you would tell me I was wrong. If I then told you Pi() = 5.342, you would tell me I was still wrong.

You wouldn't tell me "well, one of those two numbers is probably right, so let's go with the smaller one, common sense tells me 2.756 is probably the right answer"...would you?

In science we have a saying - fit the theory to the facts, not the facts to the theory.

What you are doing, your special version of "common sense", is you are cherry picking and filtering the data such that what remains fits and conforms to your theory of what you expect the results to be.

Further, the process by which you select your data (while discarding the inconvenient data) is going to be tainted by confirmation bias.

 

[SiliconWars]

What I do is look at the scores, compare them to what I know about the CPU's and using that I can tell if it's reasonably accurate.

If you are using whetstone and dhrystone only then you could end up with really bad results. Look at the results, compare them to benchmarks we know are legit. You can see clear patterns where chips like the 2600K easily beats any AMD cpu, or an E6600, or an i5 750 etc. Is that any different from what we see in other benchmarks?

If you just look at it overall instead of trying to get uber-precise numbers then it's really not that bad. Yes some of the results don't make sense, but just throw them out instead of throwing out the rest based on a few being out of whack with the rest. If you are going to discard these results because of 1 or 2 strange scores then you'll miss a lot of good potential information.

 

[wlee15]

It should be noted that the T7300 benchmark was capped at 1200mhz.

 

[Abwx]

According to the the link above at P3D Kabini will also be included
in the server segment in a couple of months.

http://www.planet3dnow.de/cgi-bin/newspub/viewnews.cgi?id=1366198521

 

[SocketF]

If I tell you Pi() = 2.756, you would tell me I was wrong. If I then told you Pi() = 5.342, you would tell me I was still wrong.

Depends on the accurancy of your whole calculation. My physics teacher in school was very keen on the the significant figures (http://en.wikipedia.org/wiki/Significant_figures ) in the calculations.

You wouldn't tell me "well, one of those two numbers is probably right, so let's go with the smaller one, common sense tells me 2.756 is probably the right answer"...would you?

If you have bad/imprecise data, then I would say that "Pi equals 3" would be a very good approximation and that it will still be useful to calculate with it.

In science we have a saying - fit the theory to the facts, not the facts to the theory.

Yes, but a "guesstimation" is not a science ;-)

What you are doing, your special version of "common sense", is you are cherry picking and filtering the data such that what remains fits and conforms to your theory of what you expect the results to be.

Further, the process by which you select your data (while discarding the inconvenient data) is going to be tainted by confirmation bias.

I tend to use worst cases in such situations, i.e. use the very best numbers of the competition and assume that the result in question is also a very good one.
In the case of that specific benchmark, I would assume that the Kabini results are not that good, bc the i7 950 machine, which also runs the 3.8 linux kernel has rather bad Whetstone results.

Try to proof me wrong, however, you cant, because the whole guessing-thing is totally unscientific. ;-)

It is more of a geeky bet than a precise calculation, simply because there is no precise data base

 

[Abwx]

What matters for the near future :

“We have started volume shipments of Kabini in the first quarter and have a strong portfolio of high volume entry-level design wins based on its substantial performance and battery life improvements,” said Rory Read, chief executive officer of AMD

Technical numbers :

“This year we will have the broadest range of mobile processors in our history spanning from sub-4W to 35W.

Temash looks good...

http://www.xbitlabs.com/news/cpu/di...o_Ship_Low_Power_Kabini_APUs_for_Revenue.html

 

[Idontcare]

What matters for the near future :

Technical numbers :

Temash looks good...

http://www.xbitlabs.com/news/cpu/di...o_Ship_Low_Power_Kabini_APUs_for_Revenue.html

That is good. Means 28nm is getting off the ground.

Now if only they'd get piledriver ported to 28nm, be it a dumb shrunk type affair or a reworked steamroller core. They need their fat cores to get put on a die-size and power-consumption diet. Piledriver on 28nm process would be nice. Kabini's power numbers show us the 28nm process is good for that.

 

[frozentundra123456]

What matters for the near future :

Technical numbers :

Temash looks good...

http://www.xbitlabs.com/news/cpu/di...o_Ship_Low_Power_Kabini_APUs_for_Revenue.html

He is the CEO of amd. What is he supposed to say "our cpus are crap and no OEM wants to: use them"?

Kabini may in fact be a fine product. Point is, the CEO of and is hardly an objective observer.

 

[Abwx]

That is good. Means 28nm is getting off the ground.

Now if only they'd get piledriver ported to 28nm, be it a dumb shrunk type affair or a reworked steamroller core. They need their fat cores to get put on a die-size and power-consumption diet. Piledriver on 28nm process would be nice. Kabini's power numbers show us the 28nm process is good for that.

On a perf/mm2 perspective Piledriver is not competitive
with Jaguar even at the same node , given that perf/watt
matters more than anything else in mobile it is better to focus
on Jaguar and apparently Mr Read target the mainstream/entry
level markets of whatever use a CPU or APU.

 

[Abwx]

He is the CEO of amd. What is he supposed to say "our cpus are crap and no OEM wants to: use them"?

Kabini may in fact be a fine product. Point is, the CEO of and is hardly an objective observer.

AMD is not in good shape but still , he has a very agressive
strategy given the meand at his disposal and GF poor execution
that did delay all current and coming products by six months
if not one year in the case of Bobcat porting to 28nm.

 

[tincart]

That is good. Means 28nm is getting off the ground.

Now if only they'd get piledriver ported to 28nm, be it a dumb shrunk type affair or a reworked steamroller core. They need their fat cores to get put on a die-size and power-consumption diet. Piledriver on 28nm process would be nice. Kabini's power numbers show us the 28nm process is good for that.

Not with GloFo's arm halfway up AMD's money-hole.

 

[inf64]

Ok a reality check.
Brazos is not "crap".
Trinity is not "crap".
FX (especially 63xx/83xx) is not crap, quite the opposite.
Kabini/Temash is polar opposite of crap, to say the least.
PD based Opterons, according to spec (industry standard) are definitely not crap.

Sometimes it's just not only the products that one needs, it's industry connections and underhand deals that play major role in acceptance of its products .

 

[Gideon]

What matters for the near future :

Technical numbers :

Temash looks good...

http://www.xbitlabs.com/news/cpu/di...o_Ship_Low_Power_Kabini_APUs_for_Revenue.html

Excellent news

I only hope they manage to get actual products out sooner than with Richland, which started shipping for renevue already in January and we still have no products =/

I would really hope for June, but Back-to-school timeframe is probalby more realistic.

 

[Exophase]

That is good. Means 28nm is getting off the ground.

Now if only they'd get piledriver ported to 28nm, be it a dumb shrunk type affair or a reworked steamroller core. They need their fat cores to get put on a die-size and power-consumption diet. Piledriver on 28nm process would be nice. Kabini's power numbers show us the 28nm process is good for that.

I'm confused, are you referring to GF's 28nm bulk process? Because Kabini is definitely on TSMC 28nm (http://www.techpowerup.com/img/13-02-19/95c.jpg)

On a perf/mm2 perspective Piledriver is not competitive
with Jaguar even at the same node , given that perf/watt
matters more than anything else in mobile it is better to focus
on Jaguar and apparently Mr Read target the mainstream/entry
level markets of whatever use a CPU or APU.

Well we don't really know how big Piledriver would be on TSMC 28nm. I've said it before, but I'm pretty convinced that the potential for density can be pretty different across different nodes even if they're using the same name..

For example you can find Cortex-A9s that take up less area on TSMC's 40nm (or even 45nm, I've heard) than on Samsung's 32nm. This could be down to physical implementation but I doubt that alone is so hugely different. You can compare L2 caches and the like in this regard too. I expect GF's process to be more similar to Samsung's than TSMC's. Not going to say anything definitively one way or the other, I just think it can't be assumed that they'd have similar density.

 

[CTho9305]

Bobcat stopped around 1,75ghz right? and people that overclocked them, usually didnt get them higher than 2,4ghz.

I dont think Jaguar's will go much higher, even at 28nm vs 40nm of bobcat.
Highest Id think these will come out at is probably there abouts 2.5ghz or something.


here is a bobcat (E-350 from 1.6ghz-> @2.36ghz) (stock aircooling motherboard came with):

1.648V on TSMC40G .

 

[Abwx]

Well we don't really know how big Piledriver would be on TSMC 28nm. I've said it before, but I'm pretty convinced that the potential for density can be pretty different across different nodes even if they're using the same name..

For example you can find Cortex-A9s that take up less area on TSMC's 40nm (or even 45nm, I've heard) than on Samsung's 32nm. This could be down to physical implementation but I doubt that alone is so hugely different. You can compare L2 caches and the like in this regard too. I expect GF's process to be more similar to Samsung's than TSMC's. Not going to say anything definitively one way or the other, I just think it can't be assumed that they'd have similar density.

TSMC 28nm density seems to largely match Intel s "22nm" ,
it is indeed superior to GF 32nm SOI but with the drawback
of a lower frequency so no wonder that AMD will use GF
28nm FDSOI for their high frequency CPUs even if it has not
as good density as TSMCs bulk.

 

[Exophase]

Do you have any official source confirming that AMD is actually using GF's 28nm FDSOI? I've only really heard this speculated.

 

[Abwx]

Seems that GF licesnsed STM 28/20nm FDSOI.

• The 28nm FD-SOI generation, currently in the industrialization phase, is scheduled to be available for prototyping by July 2012.
• The next node, the 20nm FD-SOI generation, is currently under development and is scheduled to be ready for prototyping by Q3 2013.

http://www.advancedsubstratenews.co...-st-st-technology-open-to-other-gf-customers/

28nm FD-SOI Process Design Kit (PDK) is available now, targeting risk production by mid-2012. Evaluation SPICE models are now available for the 20nm node, and full PDK is scheduled by end of 2012, with risk production for 13Q3

If spice models are available then dynamic simulations are possible.

 

[Exophase]

I know that GF has the capability of doing FDSOI at 28nm. I'm asking you if you have an official source confirming that AMD is using it.