Kabini Rumors

[sm625]

I remember Bobcat was boasting nearly 2x better IPC than Atom in pre-release BOINC numbers (also Dhrystone), in the real world you never saw anything close to that.

Having witnessed the performance of both platforms I can safely say that there are plenty of real-world scenarios where one does feel twice as fast as the other. Javascript on brazos, for example, runs twice as fast, clock for clock, vs the atom.

 

[parvadomus]

I want a 12 core Kabini for desktops, AM3+ at 3Ghz at least.

 

[Arkadrel]

I want a 12 core Kabini for desktops, AM3+ at 3Ghz at least.

I rather AMD not go full retard with MOAR COARS

I feel like 8 cores is plenty for now.
Id love to see the performance level of the GPUs inside the APUs raise though.

PC version of the PS4 chip would be nice for mid-low level pcs (rumor'ed to come true).
GDDR5 system memory means no AM3+ motherboard compatibility, but for the extra memory bandwidth for the APU its totally worth it, id think.

 

[Idontcare]

I rather AMD not go full retard with MOAR COARS

In a lot of ways AMD has sort of modeled itself after the old SUN business model. More cores, more threads, but weaker IPC and questionable power-consumption. In the meantime they make up for falling sales by increasing the rate of blogging.

I personally had assumed that SUN and AMD would merge at some point, one buying the other, but that was before Oracle entered the picture and scooped up SUN.

 

[SocketF]

In a lot of ways AMD has sort of modeled itself after the old SUN business model. More cores, more threads, but weaker IPC

Hasn't Jaguar a higher IPC than Piledriver? I think it should be at least similar

@parvadomus:
12 is an odd number, lets say 16 ;-)

 

[(sic)Klown12]

Hasn't Jaguar a higher IPC than Piledriver? I think it should be at least similar

Jaguar looks like it could be a good starting point for AMD to try and scale up. Kind of like Intel using Yonah as a starting point to replace Netburst.

 

[Idontcare]

Jaguar looks like it could be a good starting point for AMD to try and scale up. Kind of like Intel using Yonah as a starting point to replace Netburst.

I've been saying that since Bobcat. Only JFAMD poo-poo'ed that idea here in these forums when I mentioned it.

Mind you, just because I think it is a good idea doesn't mean I have any reason to think such.

My reasoning is/was entirely superficial as I am not a microarchitecture guy. I just make associations and grand generalizations and liked the parallels between Intel's Merom/Yonah and AMD's situation with Bobcat (then) and Bulldozer (then).

I don't know if there is anything intrinsically scalable about Jaguar's microarchitecture such that it could potentially rival the performance of say Piledriver or Steamroller but at a far lower power-consumption. We saw Intel do that, doesn't mean AMD can though.

Very different R&D levels and priorities between those two companies. Can't assume they will arrive at the same results.

 

[inf64]

Hasn't Jaguar a higher IPC than Piledriver? I think it should be at least similar

@parvadomus:
12 is an odd number, lets say 16 ;-)

Similar IPC but 2x less max. clock headroom with somewhat better power spec at the low volt/clock range and same thread count.
Remember that Jaguar should be compared to SR running at lower clock as it's more fair comparison. Both at 28nm, both are a real next gen from their predecessor cores (BD/PD->SR and Bobcat->Jaguar).

 

[insertcarehere]

I rather AMD not go full retard with MOAR COARS

I feel like 8 cores is plenty for now.
Id love to see the performance level of the GPUs inside the APUs raise though.

PC version of the PS4 chip would be nice for mid-low level pcs (rumor'ed to come true).
GDDR5 system memory means no AM3+ motherboard compatibility, but for the extra memory bandwidth for the APU its totally worth it, id think.

I have doubts that any GDDR5 system memory will be used effectively by pc programmers, as it has never been seen before. I don't think an 8 core (or less) jaguar @1.6ghz will hold up in pc gaming as well as a ps4 anyway. Still, will be curious to see anything like this to come out of AMD.

 

[Ventanni]

Will Jaguar IPC be that good? I was able to dig up some benchmarks from the link below. They're not perfect, but they show an interesting comparison between K10, K8, and Bobcat. It looks like Bobcat prioritized integer performance over FPU performance, which should largely be addressed by the substantially improved FPU unit in Jaguar.

http://www.xtremesystems.org/forums...Core-performance-analysis-Bobcat-vs-K10-vs-K8

Looks promising if Jaguar can be scaled to the 3-4ghz range.

 

[(sic)Klown12]

Looks promising if Jaguar can be scaled to the 3-4ghz range.

A potential problem I can see is that Jaguar's design might be so tied to being a low speed that to increase it would need a much longer pipeline that would negate a lot of the increase the additional speed would add.

 

[inf64]

Will Jaguar IPC be that good? I was able to dig up some benchmarks from the link below. They're not perfect, but they show an interesting comparison between K10, K8, and Bobcat. It looks like Bobcat prioritized integer performance over FPU performance, which should largely be addressed by the substantially improved FPU unit in Jaguar.

http://www.xtremesystems.org/forums...Core-performance-analysis-Bobcat-vs-K10-vs-K8

Looks promising if Jaguar can be scaled to the 3-4ghz range.

There is no chance for Jaguar to even clock to high 2Ghz range, let alone 3+Ghz. It's just not designed with such clocks in mind. AMD's goal was to somewhat improve clocking headroom from Bobcat's limit, which was 1.75Ghz. They managed to hit 2-2.2Ghz range for QC offerings which is 25%, a big increase in clocking ability (pipeline depth adjusted) while increasing the IPC substantially at the same time.

 

[Arkadrel]

Looks promising if Jaguar can be scaled to the 3-4ghz range.

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):

 

[SocketF]

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

Yes, but these are no unlocked chips. Chipset-, PCIe-clock or some other clock probably limit the core clock, not the cpu itself.

The pipeline depth is by the way quite long, it shouldn't limit the clock in the 2-3 GHz range, Bulldozer's pipeline is not much longer.

 

[parvadomus]

How much die area 2 jaguar core uses vs a BD module? Or transistors? How many cores can we put in the same 315mm of zambezi

 

[Arkadrel]

If they really could get a jaguar running at 4ghz... that would be intresting.
Would very much like to see how it does at those speeds.

How much die area 2 jaguar core uses vs a BD module? Or transistors? How many cores can we put in the same 315mm of zambezi

3.1mm^2 for 1 core.

 

[parvadomus]

That would be like 50mm2 for a 16-core Jaguar based CPU, They could use like 150mm2 more for L2, L3, I/O etc.

 

[nehalem256]

I would think the real market for a 16-core Jaguar would be servers.

16-core Opterons clock lower than desktop chips making the likely clockspeed handicap less.

 

[Arkadrel]

That would be like 50mm2 for a 16-core Jaguar based CPU, They could use like 150mm2 more for L2, L3, I/O etc.

Not really sure you can compair like that.

Hans de Vries blew up a picture, and estimates the Kabini to be around ~110mm^2.

Thats only for a 4core APU.
Not sure how much of that is GPU, L2/L3, I/O FCH ect.

 

[mrmt]

I personally had assumed that SUN and AMD would merge at some point, one buying the other, but that was before Oracle entered the picture and scooped up SUN.

Two rocks do not float

 

[SocketF]

Not really sure you can compair like that.

Hans de Vries blew up a picture, and estimates the Kabini to be around ~110mm^2.

Thats only for a 4core APU.
Not sure how much of that is GPU, L2/L3, I/O FCH ect.

You can easily compare the die sizes by just using the module sizes.

A Quadcore-Jaguar module with 2 MB L2 has 26.2 mm² in 28nm
A Bulldozer-module with only 2 cores has 30.9 mm² in 32nm

Including the scaling effect from 32 -> 28nm, you can more or less say, that 4 Jaguar cores equal 1 BD-module with 2 cores.

So AMD could build a combined 16core-Jaguar-Vishera below 300 mm² by using 28nm. Power consumption and clock are then the big questions. Probably the same around ~125W but only at ~2GHz. Most important question for enthusiast would be the max clock for the turbo-mode of 1 Jaguar-module / 4 cores, but that also depends on the manufacturing process.

 

[Dresdenboy]

I have doubts that any GDDR5 system memory will be used effectively by pc programmers, as it has never been seen before. I don't think an 8 core (or less) jaguar @1.6ghz will hold up in pc gaming as well as a ps4 anyway. Still, will be curious to see anything like this to come out of AMD.

Not as long, as there are just a few threads, leaving most of the cores idle.

Any high bandwidth GDDR5 system memory shouldn't be a problem as it's working transparently, just a bit faster.

 

[Olikan]

Not as long, as there are just a few threads, leaving most of the cores idle.

Any high bandwidth GDDR5 system memory shouldn't be a problem as it's working transparently, just a bit faster.

one thing that i noted in the benchmarks... the shared l2 cache is really slowing the MT scaling

 

[Dresdenboy]

one thing that i noted in the benchmarks... the shared l2 cache is really slowing the MT scaling

A single DDR3 channel might have a bigger impact.

Could also be the inapplicability of a turbo mode with all cores loaded.

 

[Exophase]

You can easily compare the die sizes by just using the module sizes.

A Quadcore-Jaguar module with 2 MB L2 has 26.2 mm² in 28nm
A Bulldozer-module with only 2 cores has 30.9 mm² in 32nm

Including the scaling effect from 32 -> 28nm, you can more or less say, that 4 Jaguar cores equal 1 BD-module with 2 cores.

So AMD could build a combined 16core-Jaguar-Vishera below 300 mm² by using 28nm. Power consumption and clock are then the big questions. Probably the same around ~125W but only at ~2GHz. Most important question for enthusiast would be the max clock for the turbo-mode of 1 Jaguar-module / 4 cores, but that also depends on the manufacturing process.

I wonder how much of this is due to density differences on the separate processes. ie, if Jaguar is made on GF 28nm how large would it be, and how large would a BD or PD module be on TSMC 28nm..

I don't think a 16-core 2GHz Jaguar would use anywhere close to 125W - no reason it shouldn't scale like 4 quad-core 2GHz Kabinis would, and individually those should use under 20W for the CPU part alone. So I'd expect under 80W to be doable, but this is assuming that the uarch can handle this kind of scaling in the first place. I don't know what the limitations of the coherency interconnect capabilities are, or the compute units that can hang off the memory bus, etc.

 

[Pilum]

Erste Linux-Benchmarks von AMDs Kabini-APU aufgetaucht
source: www.planet3dnow.de

Pro tip: this is an English-language forum.

Also... Dhrystone! How exciting!! Finally some meaningful benchmark data!!!!