Richland & Kabini rumours

[mrmt]

I hope you can read since it doesn't say ~10% , it says more than 10% (or >10%). More than means it can be 11% but it can be 30% too.

I think we have someone here who never prepared a ppt to a customer or to senior management. If you have 30% increase on your product performance, why would you write >10%? That's right, you don't. It's outright stupid. If you have 30% on the sleeve what you do now is to generate a lot of hype to the product without disclosing your secret sauce. Promise and deliver 30% puts you in a very nice spot. In business language, and you can bet your paycheck that AMD senior management reviewed what the guy can or cannot talk, >10% means 10%, 11%, 12%, maybe 13%. If you can reach 14% you'll write ~15%, not >10%. It never means 30%. Never.

Looking back at x86 history, I can't recall any other time beyond Netburst => Conroe when we had >30% increases on IPC, and even then clocks were down by 40%. IPC increases of the same arch usually sits between 5% and 10%, so if they get 10% with only adding 10% on xtor budget, a near perfect scaling, that would be quite a feat already.

As for the clocks, going from 2 to 4 cores puts a great burden on the thermal budget, so doubling core counting and still being able to rise frequencies isn't something trivial. The first slide stating 10% doesn't seems too out of touch with reality. The assumption here is that a 28nm bobcat core with double the number of cores would have roughly the same frequencies.

That's what you do to obscure the information about the product that's not yet launched(and to protect it due to competitive reasons).

I surely wouldn't want everyone knowing totally the merits and flaws of my products, as it would give my competitors an unneeded head start to counter my product, but a rough but accurate performance estimate without disclosing what I did to get there generates the kind of good hype for the product that brings OEMs in, that makes key people and companies interested in knowing what I have, and even better, some very good news to announce to investors at the EC. It is the "how" that cannot be disclosed.

 

[Abwx]

That slide tells us nothing about 28nm Bobcat . What they say is basically >10% gain. Of course it's more than 10% gain, but 15% is >10% and 35% is also more than 10% . The new information is that >10% refers to 28nm Bobcat ,not the one we have today(and the lead of Jaguar project states so in the video,in no uncertain terms).

This slide say that Kabini frequency will be increased by more than 10% ,
wich put it at about 2Ghz max , while IPC will be increased by more
than 15% , the combined numbers point to more than 26% better
perfs than a 1.7Ghz Bobcat at the said 2Ghz frequency , surely
for the 25W part , though.

Other meaningfull enhancements are obviously FP , as you mentionned it ,
as well as an impressive instructions set extension from wich only FMA
is missing compared to a Piledriver.

Cuting the frequency by 50% while reducing drasticaly the Vcore ,
getting rid of two cores and of half the GPU while working on single
channel memory would put it in tablet TDP territory.

Edit : The perfs of an hypothetical 28nm Bobcat would be cetainly also improved by a 10% freq. upgrade , as for IPC ,
we can just take the AMD rep. words at face value...

 

[mrmt]

JFAMD was a Director of Product Marketing, Server Business Development for AMD, he was his job to sell you CPUs. Jeff Rupley presentation was about Information on the Jaguar Architecture, and had nothing to do with PR or sales.

Yeah, you either can't understand the meaning of "sell", the meaning of a technical guy in an open event explaining the architecture all those people or you don't want to, either way, I'm out.

It is obvious that you havent understood what he(Rupley) said. Let me explain, he said 10% more over an imaginable Bobcat at 28nm. That is, IF we had a 28nm BobCat, Jaguar would have a +10% more.

If you have a look at the AMD slide abwx posted before, an official AMD slide for that matter, it states design goals for Kabini. It is safe to say that a 28nm bobcat would have performance targets sitting between the 40nm bobcat mentioned on the slides and the 28nm Kabini otherwise Kabini wouldn't make any sense at all. So while we don't have a 28nm bobcat to benchmark, we can have a reasonable performance estimate from that point.

The Jaguar Core Architecture is on the designing board for more than a year now, the decision to make it have been made even farther in the past and recent economic conditions played no part in that product because it is already finished(the design). The product will be released within the next 6 months.

Oh dear, and here I thought that the cuts Rory Read did on his first month at the helm were due to the bad economic situation AMD was in. Now, thanks to you, I know that AMD situation in 2011 was fine. But now I can't really understand the reason for the job cuts. Maybe it was just the new CEO doing his rite of passage, you know, firing 10% of the company and collecting the bonus without feeling bad at night.

 

[Abwx]

If you have a look at the AMD slide abwx posted before, an official AMD slide for that matter, it states design goals for Kabini. It is safe to say that a 28nm bobcat would have performance targets sitting between the 40nm bobcat mentioned on the slides and the 28nm Kabini otherwise Kabini wouldn't make any sense at all. So while we don't have a 28nm bobcat to benchmark, we can have a reasonable performance estimate from that point.

Whatever the numbers of the definitly unreleased 28nm Bobcat
the real question is rather if the promised iteration will have a
significant market share.

I dont know the die area of a 4C Kabini but if they manage to keep
the same area or so as a 2C Bobcat then it s likely that this chip
will find its way on mass produced low cost notebooks and tablets
for 2C versions.

 

[Hans de Vries]

Found a Video from HC24 on Jaguar presentation . Interesting Q&A (I embedded the time,just click). Head of the Jaguar project basically states they achieved >10% mircoarchitectural frequency gain versus Bobcat if it was done on 28nm . That was very surprising to me since what we heard before from sites covering HC was just : >15% singlethread IPC gain(~2x fp throughput), 2x more cores, >10% clock gain (assumed versus 40nm Bobcat). Now this new info from the guy in the video is revealing in a way that we have no clue at what clock would a 28nm Bobcat run. It could be the same 1.7Ghz as top product was running at or it could have been significantly more. Also in the very presentation the fellow mentioned that by doing various optimizations to the added die area(more xtors) they've increased the die size of Jaguar by about 10% versus a Bobcat if it was done on 28nm (also a very interesting information).

So new information (to me at least): core is 10% bigger than what 28nm Bobcat would have been if it was on 28nm ; clock gain by adding stage here and there(decoder and fp scheduler portions of pipe) cost them a bit IPC(didn't say how much,probably a few %s) while providing a >10% clock gain to the design. So a question is : how high does it clock in the end and how much power did they save(for Temash). Bobcat at 28nm would certainly clock much higher than 40nm Bobcat and/or have higher Turbo boost state at least. This new Jaguar core may very well significantly outclock 40nm Bobcat(20+%?) while performing much better at the same clock(SSE stuff would see major uplift,basically K10 level if not better).

In the end we may have pretty potent chip in Jaguar for even a desktop user,with decent GPU and decent CPU. Since it was supposed to be launching in the same year alongside SR core performance delta would have been in the favor of later(but by how much). As it is now, IF Richland is 28nm it could save it in direct comparison with this little fella,but even if Richland is a Piledriver+,from IPC standpoint(by as much as 5%), Jaguar would probably still own it in IPC department . Fun times ahead!

Nice find, thanks!

 

[cbn]

Welcome to the forum!!

 

[IntelUser2000]

Its true what mrmt is saying. It's unlikely that they'll be going far higher than 10% clock increase in total(that means 28nm Jaguar vs 40nm Bobcat).

You gotta also account for power use. You get way more % decreasing power than increasing clocks. While it may be possible to clock Jaguar at same process 10% higher than Bobcat, things like improved Uncore, Graphics, and even CPU core might eat that power budget alone. Leaving the real improvement up to process.

It's rare that what is claimed before the product turns out to be better in average in real life. Often, it doesn't even meet targets.

 

[AtenRa]

If you have 30% increase on your product performance, why would you write >10%?

He is talking about Frequency not Performance. He is clearly saying MORE (>) than 10% than a 28nm BobCat. Play the video again

 

[AtenRa]

Its true what mrmt is saying. It's unlikely that they'll be going far higher than 10% clock increase in total(that means 28nm Jaguar vs 40nm Bobcat).

You gotta also account for power use. You get way more % decreasing power than increasing clocks. While it may be possible to clock Jaguar at same process 10% higher than Bobcat, things like improved Uncore, Graphics, and even CPU core might eat that power budget alone. Leaving the real improvement up to process.

It's rare that what is claimed before the product turns out to be better in average in real life. Often, it doesn't even meet targets.

Bobcat(Brazos CPUs) was up to 18W, Kabini SoCs (Jaguar) will be up to 25W.

Not only that but, you have a full node shrink and a better architecture (more efficient). That two allows you to have higher frequency and higher IPC. Playing with voltage let you have more frequency at the same Wattage as 40nm BobCat or using lower Voltage will give you same performance as BobCat but with lower power consumption.
Clearly you can have more than 10% higher frequency and certenly more than 10% overall performance from BobCat to Jaguar.

ps: Ohhhh, and we havent touched the iGPU yet. It will be more power efficient than that on BobCat.

 

[AtenRa]

Oh dear, and here I thought that the cuts Rory Read did on his first month at the helm were due to the bad economic situation AMD was in. Now, thanks to you, I know that AMD situation in 2011 was fine. But now I can't really understand the reason for the job cuts. Maybe it was just the new CEO doing his rite of passage, you know, firing 10% of the company and collecting the bonus without feeling bad at night.

Jaguar has started before Rory became CEO at AMD. And since even Rory made clear that they are after the smaller Core dies, the design and the Team behind Jaguar remain unaffected buy the first firing.

 

[mrmt]

He is talking about Frequency not Performance. He is clearly saying MORE (>) than 10% than a 28nm BobCat. Play the video again

So what? >10%, frequency, something, die reduction, anything, if you have 15%, 20%, 50%, you don't put >10% on a presentation.

 

[inf64]

Ok mrmt,you are correct.It's 10%.

 

[AtenRa]

Welcome to the forum Hans, your expertise will be really appreciated.

 

[mrmt]

Jaguar has started before Rory became CEO at AMD. And since even Rory made clear that they are after the smaller Core dies, the design and the Team behind Jaguar remain unaffected buy the first firing.

So what? Look at what I said in the previous post:

As for the 28nm bobcat, the results shouldn't be very good. If 28nm bobcat were that good a cash strapped company like AMD would have gone that route without thinking too much. I.stead they went for the hard route, improving the uarch.

To which you answered:

The Jaguar Core Architecture is on the designing board for more than a year now, the decision to make it have been made even farther in the past and recent economic conditions played no part in that product because it is already finished(the design). The product will be released within the next 6 months.

I said that a cash strapped company like AMD wouldn't go for a new uarch if a Brazos die shrink would yield results too close to Kabini, so the performance of a hypothetical 28nm Brazos would sit between 40nm Brazos and 28nm Kabini, with a nice distance from the latter. Given that AMD official slides already compared Brazos and Kabini, we can have a nice idea of where a 28nm Brazos would sit.

I never mentioned recent economic conditions, but I said that a cash strapped company like AMD would be very cash conscious when deciding where to go on this fork. And did AMD become cash strapped today? No. AMD has been bleeding cash since 2006, with only a small hiatus in 2011. So recent economic conditions just made the situation a lot worse, but AMD didn't ever had cash in abundance since acquiring ATI.

So my point about being cash strapped when giving go-no go for Kabini still stand unless you go for years before 2006, where AMD had a reasonable amount of cash at their disposal.

 

[AtenRa]

By your logic AMD would have never designed Bulldozer. Jaguar being a smaller core is more easier, it needs less capital and less time to change/design it.

Clearly AMD have decided they needed to change something, I believe they have overhauled Jaguar in order to take the 25W place with a smaller CPU die design than Trinity and Kaveri.

 

[NTMBK]

Bobcat(Brazos CPUs) was up to 18W, Kabini SoCs (Jaguar) will be up to 25W.

Quick note - Kabini is not an SoC. It is an APU. Only Temash is an SoC (with integrated FCH).

 

[AtenRa]

If im not mistaken Kabini will have integrated FCH.

 

[mrmt]

By your logic AMD would have never designed Bulldozer. Jaguar being a smaller core is more easier, it needs less capital and less time to change/design it.

Yes, in my logic AMD would have never designed a Bulldozer, but regardless of my opinion, it is clear by now that Bulldozer was a *huge* strategic mistake for AMD, and the company is paying dearly for it. Rory's team seems to be committed to Jaguar and ARM too.

The current management doesn't seem fond of the thing too, they seem to be prioritizing Kabini and ARM over their big-core line.

Clearly AMD have decided they needed to change something, I believe they have overhauled Jaguar in order to take the 25W place with a smaller CPU die design than Trinity and Kaveri.

That's my understanding too. Problem is that, the way the company is structured now, there will be product overlapping, sales overlapping, and marketing overlapping, this in a moment that the market is shrinking and AMD sales are plunging. AMD is too small to have two product lines serving the same market.

Also AMD has nothing close of the clout Intel has for OEMs, and Intel know the dangers of mixing two lines. Regarding Core and Atom, limiting OEM specs the most they could, they still tight fenced the two products lines under different marketing umbrellas. Different names, different logos, different everything. When I see two 14inch notebooks with AMD logo and Radeon graphics logo, one with Brazos and another with Trinity, how do you explain the price and performance differences for Mr and Ms. Doe?

To be fair with AMD this overlapping might not be their fault entirely. One plausible reason for them to do that is the need to keep churning out Trinity chips from GLF in order to not pay another take or pay charge. Selling the thing at loss might be better than paying the charge.

 

[inf64]

I think it's best for you to try and apply for a job in AMD. You *obviously* know all about their mistakes and problems,just present your case and see what happens.

 

[AtenRa]

That's my understanding too. Problem is that, the way the company is structured now, there will be product overlapping, sales overlapping, and marketing overlapping, this in a moment that the market is shrinking and AMD sales are plunging. AMD is too small to have two product lines serving the same market.

You havent understand what i have said, maybe there will only be a single product under the 25W segment with a Jaguar core architecture. So, no overlaping.

When I see two 14inch notebooks with AMD logo and Radeon graphics logo, one with Brazos and another with Trinity, how do you explain the price and performance differences for Mr and Ms. Doe?

The same way you going to answer if one was an Intel Pentium and the other was an Intel Core i3.

The more expensive is the fastest :biggrin:

No need to explain the technical stuff to people that doesnt know about hardware, it will only confuse them more.

 

[mrmt]

You havent understand what i have said, maybe there will only be a single product under the 25W segment with a Jaguar core architecture. So, no overlaping.

They will overlap on 17w and 25w.

 

[AtenRa]

So you know for certain they(AMD) will release a 17W and 25W Kaveri ??

What if they will only release Jaguar products all the way up to 25W, after that they can have 35W up to 100W with Kaveri.

 

[IntelUser2000]

Bobcat(Brazos CPUs) was up to 18W, Kabini SoCs (Jaguar) will be up to 25W.

Yea, but because they have a quad core version. Probably the 25W will be 4 core while 18W would remain dual core.

 

[mrmt]

So you know for certain they(AMD) will release a 17W and 25W Kaveri ??

What if they will only release Jaguar products all the way up to 25W, after that they can have 35W up to 100W with Kaveri.

Because kabini will start with Richland, not with Kaveri. And unless you are assuming tgat AMD will phase out17w and 25w skus with Richland, we will see those chips overlapping with kabini.

 

[Ventanni]

I will admit my own over-excitement for Jaguar, but I think AMD will scale its Jaguar line up to 100w and replace the Bulldozer architecture altogether.