Richland & Kabini rumours

[VirtualLarry]

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.

100W Jaguar? What would that entail? If a quad-core Jaguar is 25W, would that mean a 12/16-core Jaguar? That would certainly be interesting, if only from a marketing POV.

"16 CORE JAGUAR CPU!"

 

[Centauri]

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.

It makes sense. AMD needs to focus on a select number of designs, and Jaguar has a better chance of cheaply becoming a TDP-competitive high end offering than Bulldozer does of becoming TDP competitive in any segment.

 

[inf64]

100W Jaguar based server CPU?Why not. But a desktop one? Wishful thinking.

 

[nehalem256]

100W Jaguar? What would that entail? If a quad-core Jaguar is 25W, would that mean a 12/16-core Jaguar? That would certainly be interesting, if only from a marketing POV.

"16 CORE JAGUAR CPU!"

It depends on how much of 25W is the cpu cores. If it takes say 7W to go from 2 to 4 cores it might only take say 75W for a 16-core Jaguar. Although you would have to also probably go to a dual-channel memory controller to keep the kitty fed.

 

[AtenRa]

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

There will be quad core Kabini at 18W.
Having a Full Node shrink, gives them the ability to ~double the transistor count and at the same time keep the same power consumption. That means we could have double the Brazos(quad core, 160 Radeon cores) at the same 18W. Add that the Jaguar cores are more efficient and with an even more efficient iGPU than VLIW5 that Brazos has and you can have a very powerful SoC at 18W.

 

[Ventanni]

100W Jaguar? What would that entail? If a quad-core Jaguar is 25W, would that mean a 12/16-core Jaguar? That would certainly be interesting, if only from a marketing POV.

"16 CORE JAGUAR CPU!"

Just hope it doesn't break down after pulling it out of the "driveway", err case, lol.

But anyways, Brazos was a stripped down K10 core if I remember right. Jaguar is an improvement on Brazos, so really this is just an evolution of where the original K10 architecture could go. It's not fair to directly compare to what Intel did, but the Bulldozer story is awfully similar to the Netburst one. I think architectures like these (Bulldozer/Netburst) are, or could be, a theoretically superior product on paper, but they both run into the same problem; in order to be great, they run into a thermal wall, or at least they break the physical limits of how fast a silicon transistor can realistically switch on a mass produce-able, reliable scale.

Also, if you have a 100w server product, you have a 100w desktop product too. Both Intel and AMD already do this.

Back to where I'm going with this though, it just seems like Brazos/Jaguar is a reimagining of K10. They've stripped it down, improved on some aspects, integrated it with a GPU, and unlike Bulldozer, they've actually made money off of it. And, they've also integrated it into an SoC design (hello tablet world!). It just seems natural that Jaguar-Core-Next (lawlz I made that up) will be the next primary architecture for AMD after Steamroller/Excavator.

But that's just me. I could be a bit too excited about it, so I'll admit that.

 

[inf64]

Bobcat is a new design and not a K10 derivative.

 

[NTMBK]

If im not mistaken Kabini will have integrated FCH.

You're mistaken, I'm afraid.

The 2nd Gen Low Power APU is Kabini, 1st Gen SoC is Temash.

 

[pablo87]

@ pablo

AMD cannot design and produce "15" designs. What you suggest takes engineering resources and time(and money). They CAN make GPU-less Trinity.

From various reviews we see that at most Vishera can be 15% faster than L3-less Piledriver bu this is corner case. In reality the difference is closer to ~4-5% on average.

GPU-less 2M Pilderiver @32nm would take about half of Trinity's die area(so ~120mm^2). 4M Piledriver with no L3 would in turn take ~200mm^2(count in only 2 additional modules and a bit more of "dead area"). 200mm^2 on 32nm sounds a lot better than 320mm^2 ,and both are with automated tools involved.

Imagine if AMD had access to intel's money and fabs: PD or even better SR on 22nm intel-level fab tech with hand optimized layouts and 8 different dies(2/4M with and without L3 for server/desktop; 1/2M with GPU and without L3). Now since AMD has no money and cannot make such variations in their product mix in timely manner,they are doomed to die salvage what they can. Unfortunately.

Inf64,

I was suggesting that were it not for the WSA, Bobcat becomes the architecture for APUs (AMD's market is different than Intel - their lack of corporate business means their APUs are more or less relegated to Consumer/SMB and in that space, good enough is good enough) and that all of it is fabbed at TSMC.

Do you think I'm wrong to think this would save engineering resources (porting GPU cores to GF among other things) that could be applied to permutations of Brazos (higher performing GPU, with perhaps external GDDR)?

Also, my suggestion was not to have Trinity at all, just have Piledriver on AM2. At which point, it begs the question, is AM3 needed? So you end up cutting 2 efforts - Trinity and Vishera/AM3, so would there not be more engineering resources saved by AMD?

Moreover, is it possible that there is an opportunity cost to developing PD Trinity and supporting 2 sockets, that could have been applied to core development - maybe L3 less SR on AM2 comes out faster?

A little off topic - holistic approach always seems to be.

 

[pablo87]

Ehm, they are doing that already...

They are called Athlon II X4 730/740/750K. Yes the 750K has an unlocked multiplier...

And so no one can claim they are not available:

http://www.proshop.dk/CPU/AMD-Black-Edition-Athlon-II-X4-Processor-2382433.html

In stock and for sale... :whiste:

Yeah but no stock in Aarhus, maybe we can get Lego to make us some?

My target price would be $60, lets see that would be 3x as profitable as OEM Trinity.

 

[ShintaiDK]

Yeah but no stock in Aarhus, maybe we can get Lego to make us some?

My target price would be $60, lets see that would be 3x as profitable as OEM Trinity.

You know what "fjernlager" means?

And Lego is in Billund.

 

[IntelUser2000]

There will be quad core Kabini at 18W.
Having a Full Node shrink, gives them the ability to ~double the transistor count and at the same time keep the same power consumption. That means we could have double the Brazos(quad core, 160 Radeon cores) at the same 18W. Add that the Jaguar cores are more efficient and with an even more efficient iGPU than VLIW5 that Brazos has and you can have a very powerful SoC at 18W.

Sure, there will be a 18W quad core, but the frequency will be significantly lower. The 7W increase is there to get similar frequency as the dual core Brazos at 18W. I wouldn't be surprised we'll see a dual core 25W part that clocks 10-15% higher than the quad too.

The days where you go 50% reduction in power every process is gone. It's now ~30%. Also you have the memory controller and I/O that scales even worse and still takes up power.

You can multiply amount of transistors many times as you want, but it'll be limited by power use how high it'll perform.

 

[IntelUser2000]

......

It looks like a good chip, but there's nothing that makes it stand out. It'll have to play between rock(Clover Trail) and a hard place(Core).

There's a technical reason for that. Care to guess why?

 

[pablo87]

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

After all the promises and failure to meet expectations, it would be about time for AMD to promise something that is acceptable to buyers, and then follow thru by exceeding those promises.

Its called managing expectations and its common sense/mgmt 101.

Though common sense is something I've come not to expect from AMD.

 

[pablo87]

You know what "fjernlager" means?

And Lego is in Billund.

No, I'm afraid not.

Yes I know.

 

[krumme]

As Pablo says the WSA clouds all rational decisions.

But if we look at the difference to bobcat what strikes is the new powergating, on top of the lower leaking 28nm process. When a core is like 4mm2 ?, its basicly free to add more cores concerning production cost and also in regards to battery life. That spells low end laptop and also desktop market, as addition to the tablet market.

We know that the Hyderabad team was about 80 persons putting all the IP together. So its a arch that is very cheap to scale, and technically very obvious to do it.


I dont think AMD today have the organization and cash to maintain a more complex arch. There is not point in beeing number 2 in this market. Kabini obviously have a huge target market where low power - and expensive - Atom, nor expensive Haswell will be able to compete.

 

[AtenRa]

Sure, there will be a 18W quad core, but the frequency will be significantly lower. The 7W increase is there to get similar frequency as the dual core Brazos at 18W. I wouldn't be surprised we'll see a dual core 25W part that clocks 10-15% higher than the quad too.

The days where you go 50% reduction in power every process is gone. It's now ~30%. Also you have the memory controller and I/O that scales even worse and still takes up power.

You can multiply amount of transistors many times as you want, but it'll be limited by power use how high it'll perform.

According to TSMC, 28nm HP has 45% more speed, ~2x transistor density at the same leakage as 40nm G process that Brazos is using.

Take as an example the Radeon HD7970 produced at 28nm HP at TSMC against the HD6970 at 40nm G process.

HD7970 has ~63% more transistors, higher frequency with smaller die and close the same power consumption as HD6970 at 40nm G.

That means that Brazos at 28nm could have almost double the CPU cores, double the Radeon cores, almost same frequency 1.6-1.7GHz and same power consumption (18W) as Brazos at 40nm.
Now, add a 15% higher IPC (or higher for single core) and an improved more efficient with more Radeon Cores iGPU and you have a winner at 18W.

Of-course a dual core could be clocked higher(maybe up to 40-50%) and keep the same 18W TDP.
I will say that a ~2.5GHz Dual core Jaguar would be faster than a 2GHz Dual core Single Module Trinity at 18W (A6-
4455M 2.1GHz base with 256 Radeon Cores at 17W) at half the die size .

 

[Haserath]

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

They could try the Bulldozer strategy.

Promise the moon then explode on the launch pad.

 

[VirtualLarry]

I will say that a ~2.5GHz Dual core Jaguar would be faster than a 2GHz Dual core Single Module Trinity at 18W (A6-
4455M 2.1GHz base with 256 Radeon Cores at 17W) at half the die size .

That would indeed be pretty impressive if they could pull that off.

 

[NTMBK]

According to TSMC, 28nm HP has 45% more speed, ~2x transistor density at the same leakage as 40nm G process that Brazos is using.

Take as an example the Radeon HD7970 produced at 28nm HP at TSMC against the HD6970 at 40nm G process.

HD7970 has ~63% more transistors, higher frequency with smaller die and close the same power consumption as HD6970 at 40nm G.

That means that Brazos at 28nm could have almost double the CPU cores, double the Radeon cores, almost same frequency 1.6-1.7GHz and same power consumption (18W) as Brazos at 40nm.
Now, add a 15% higher IPC (or higher for single core) and an improved more efficient with more Radeon Cores iGPU and you have a winner at 18W.

Of-course a dual core could be clocked higher(maybe up to 40-50%) and keep the same 18W TDP.
I will say that a ~2.5GHz Dual core Jaguar would be faster than a 2GHz Dual core Single Module Trinity at 18W (A6-
4455M 2.1GHz base with 256 Radeon Cores at 17W) at half the die size .

AtenRa, don't get too carried away. Don't BullDozer this. It looks like it has potential, but if you hype it as if it is going to be the next Conroe Moment then it's probably going to crash and burn, and people are going to lash out at it.

Wait until we have hard numbers, and until reviewers have silicon in hand and benchmarks to show us. Then if it turns out as brilliant as you had hoped, shout it from the rooftops! Don't get this worked up over a handful of slides and some very rough estimates from TSMC. There's plenty more involved in what clock speeds can be attained than just the process involved, you know that.

Chill out. Be upbeat about it, sure. But if you build it up this much, then people are just going to tear it down if it performs even slightly below the extremely high expectations you're setting for it.

 

[AtenRa]

NTMBK, im not hyping anything, what i said above is doable with a Full Node shrink and better more optimized microarchitecture design.

Dont forget that Trinity Module has a 10-20% hit in performance from the CMT when Jaguar does not. Add that Trinity have less IPC than Phenom II and we can speculate that Jaguar at the same frequency is very close. Put all together and you end up having the same or more performance at reduced die size.
It is from all of those that i have said that perhaps AMD will replace the low end 25W and bellow Trinity with Kabini.

The only thing we dont know yet is the Radeon Cores, im expecting between 128 and 256.

 

[Abwx]

It looks like a good chip, but there's nothing that makes it stand out. It'll have to play between rock(Clover Trail) and a hard place(Core).

There's a technical reason for that. Care to guess why?

Clover Trail is not in the same league , Kabini has four hard cores
with each being more efficient than a CT core.

Core offering is undoubtly of higher perfs but it is also more expensive
to manufacture.

What will be instrumental to its eventual success is its ability
to fill the needs of the value market wich the previous offering
was not capable given its low IPC , frequency and core count.

As a two cores CPU it would be completely wiped out by either
i3s or Trinity but the quad core sibling has exactly what is needed
to compete with the lower parts of the mentionned CPUs , wich
can be a problem if it is priced too low as it is obvious that
a 4C part will be more attractive than a 2C Trinity for any
average consumer.

 

[Insert_Nickname]

You know what "fjernlager" means?

It just means the Danish distributor (dunno who proshop uses though) has it in stock... but you're right it could be very far away...

I almost hate to write it, but with that pricing its actually looking semi-competitive with a G5xx/G6xx/G8xx system. That's if you're using a discrete GFX card. A55 based boards cost about the same as H61's do. It's even unlocked, so it could possibly do 4.2ish...

Only problem is that another 200,- DKK will get you a 3220...

 

[inf64]

It just means the Danish distributor (dunno who proshop uses though) has it in stock... but you're right it could be very far away...

I almost hate to write it, but with that pricing its actually looking semi-competitive with a G5xx/G6xx/G8xx system. That's if you're using a discrete GFX card. A55 based boards cost about the same as H61's do. It's even unlocked, so it could possibly do 4.2ish...

Only problem is that another 200,- DKK will get you a 3220...

It does ~4.5Ghz on air cooling with vcore bump . And when OCed it's more than competitive .

 

[beginner99]

It does ~4.5Ghz on air cooling with vcore bump . And when OCed it's more than competitive .

Same as always. If you OC you will need to spend another $40+ for better cooling and hence you could get a better, locked Intel CPU that performs just as well.

Besides that I don't see the point of a quad-core in low power devices. You don't want to trans code movies or anything with this anyway so a dual-core with higher IPC would be much better IMHO. Even in a desktop if you don't play games or do transcoding or similar stuff a dual-core is perfectly fine.