Richland & Kabini rumours

[pablo87]

The 2 year old Brazos design sold something like 5MM units last quarter. Seems like it has established its own niche and competes with itself for the most part. If we're to believe the leaks, its replacement Kabini is better in every way.

GPU - Closest thing we have to benchmark Kabini's GPU is 8790 which is >> 5650 which in turn was >> 5450 (the architecture in Brazos). And more bandwidth. Potentially twice as fast.

FCH on-die integration is another plus. Its not only that AMD gets to save the cost of the FCH die and packaging, but OEMs benefit further on PCB real estate, jelly beans, machine time, etc...Win Win.

CPU - twice as many cores, faster FP, improved IPC and higher TDP range. Major improvement over Brazos across the board.

 

[NTMBK]

Which atom is that the 22nm process or the 14nm process. Your fond of saying 2014 / Put up or shout up make a wager . I will bet that 22nm atom is in phones and on the market in 2013. You name the stakes

22nm- and I need no reward other than more proof that you're totally wrong. How's that earth destroying rogue planet coming along, by the way?

 

[Nemesis 1]

Everything I've seen so far seems to indicate the Avoton (sp?) will be the only 22nm Atom released in 2013. Kind of a disappointment, but most OEMs haven't even managed to get their Clover Trail-based tablets out the door yet...

Ya . The Cedar trail chips having problems getting out the door . Exactly. How many tablet wins did intel have in 2011. How many in 2012 . So now that the plarforms are built changing chips is all that remains . Same as phones Motorola was way latter than ornge at bringing atom phones to market. Because Ornge used Intel referance design . Motorola will be ready for the new atom as soon almost as intel starts shipping which is around this time . You like will not see an announcement of 22nm processors coming in the phones other than what we learn next week . Than it will be silent again.

 

[SiliconWars]

Today Brazos isn't really Atom competitor only. Brazos also competes with Celeron/Pentium in the bottom market in notebooks, and more recently, desktops (it sells 10x more than 8C Bulldozer). So Kabini as Brazos successor will compete against Core and Atom.

Yes and can I remind you that Brazos was supposed to be only a competitor to Atom initially.

A 25W Kabini will have 3x-4x the graphics performance of 18W Brazos with over double the multithreaded performance. Even 14nm Celerons will be nowhere near that kind of performance increase over the current SB Celerons. And we still don't even have 22nm Celerons so really you're talking 2 years before these 14nm Celerons are actually going to be competing with Kabini, by which stage it has probably been shrunk or we'll be near the successor.

 

[SiliconWars]

You are probably correct

If it matches up with 128SP discrete part next year then it will probably be close to this spec GPU-wise. Can we now say ~40-50% faster than Brazos? It sure looks like that

128 SP is the only number that makes sense. 128 GCN on 28nm vs 80 VLIW5 40nm must be closer to 3x the increase in performance I'd reckon.

 

[inf64]

I doubt 3x is remotely possible. GCN is about 5% more efficient in games Vs VLIW4(see techpowerup) and GPU clocks will probably be similar or at best some 10% higher on Kabini. Still , if it ends up being close to 50% faster at same or better power draw it will be a massive step ahead and a tough rival to beat,especially in perf./watt department.

 

[SiliconWars]

Brazos is VLIW5 though. 64 VLIW4 should be about equal to 80 VLIW5, so just doubling the SP's should get us near double performance already.

Then we have the real problem before, which is that Brazos' igp was horribly cpu limited. The extra two cores are bound to alleviate most of that problem. I will be very suprised if it isn't at least twice as fast graphically as Brazos at the same TDP.

 

[inf64]

Yeah I forgot Brazos was VLIW5. The difference between GCN gen1 and VLIW5 is around 22-24% according to techpowerup. In OpenCL I expect it to be even higher. As for CPU bottleneck maybe the better performing cores will contribute somewhat to better utilization of the SPs but I doubt 2 more x86 cores will be adding much in terms of games. There are some games that can benefit from 2+ threads though so I'm not saying it won't help.

 

[NTMBK]

A 25W Kabini will have 3x-4x the graphics performance of 18W Brazos with over double the multithreaded performance.

Have we got confirmed numbers on the Kabini graphics parts? I haven't seen anything yet.

 

[SiliconWars]

Yeah I forgot Brazos was VLIW5. The difference between GCN gen1 and VLIW5 is around 22-24% according to techpowerup. In OpenCL I expect it to be even higher. As for CPU bottleneck maybe the better performing cores will contribute somewhat to better utilization of the SPs but I doubt 2 more x86 cores will be adding much in terms of games. There are some games that can benefit from 2+ threads though so I'm not saying it won't help.

The main thing is that the OS won't be using so much of the cpu performance available to it. Brazos' two cores are very easy to choke with even light loads so even games that are only using two cores are easily bottlenecked by it. I would expect the 3rd and 4th cores alone to give Kabini a 30%+ uplift in the performance of many games simply due to it having two free cores to use for gaming.

http://www.tomshardware.com/reviews/multi-core-cpu,2280-11.html

That's quite an old article but it shows how the increase from 2 to 3 cores affects gaming performance (15%). Clearly nobody is actually writing games for 3 cores so the uplift is solely down to having two free cores to utilise. That was back in 2009, it should be more profound these days.

Have we got confirmed numbers on the Kabini graphics parts? I haven't seen anything yet.

Nothing confirmed but if Kabini is "10% bigger than Brazos at 28nm", then it basically has to mean it has 128 SP's. If it was only 64 SP's then it would be a lot smaller. I expect Temash will have 64 SP's though, ~50mm2 total die size.

 

[Nemesis 1]

http://www.youtube.com/watch?v=VeCJd3CfeNo

 

[AtenRa]

I don't think Kabini will have such an easy time as Brazos had.

Yes, with half the power consumption, 25%+ higher performance for the CPU cores and at least 50% higher performance for the iGPU over Brazos, Intel doesnt have a chance.

First and foremost Kabini and Temash doesn't seem to have a TDP low enough to compete with Clover Trail or ARM A15, which means that AMD chips will be on tablets, but most likely niche tablets.

Well, Temash Win8 tablets will be the only product in the market that you would be able to play current DX-11 games, use your PC applications and more. Temash tablets will literally put PCs in your fingertips

Kabini won't have an easy time on laptops too, as it will clash with IVB Celeron and Pentiums. The fact that AMD is redirecting part of their Brazos sales to desktops are an indicative of how tough are current conditions for AMD low-end chips, and while Brazos might improve the situation a bit, competing against Core and Trinity won't be easy.

Kabini will have two more cores and much better iGPU than Ivy Celeron/Pentiums have. Kabini will also be much cheaper than those Intel CPUs. It seams that 2013 will be very difficult for Intel in the low end.

But more important, AMD will be stuck on 28nm until at least 2015, when they will move to a half-baked 20nm process. This at the time when Intel will be on the second generation finfet at 14nm, and Atom will be on this process already.

Atom buy that time could be irrelevant for smart phones and for x86 tablets/mobile/desktop. 28nm/20nm TSMC and 14nm-XM GloFo ARM SoCs in 2014 could be the end of ATOM in the phone market.

We already saw how poor Trinity fared against Core, I don't see any reason to think that Kabini/Temash will fare any better when subjected to the same handicaps (Intel chips will have smaller die size, lower TDP and better CPU performance) regardless of how good AMD GPU is.

We have already seen how great Brazos fared against ATOM and low budget Celerons/Pentiums, it is only logical to expect that Kabini will do only better with half the TDP, 25% higher CPU performance and 50%+ iGPU performance over Brazos. And all that, almost in the same die size of Brazos.

 

[inf64]

Nothing confirmed but if Kabini is "10% bigger than Brazos at 28nm", then it basically has to mean it has 128 SP's. If it was only 64 SP's then it would be a lot smaller. I expect Temash will have 64 SP's though, ~50mm2 total die size.

10% bigger x86 cores ,not the die containing iGPU.

 

[SiliconWars]

10% bigger x86 cores ,not the die containing iGPU.

Ah right. Well either way, it's either a ~50mm2 chip with 64 SP's or a ~80mm2 chip with 128 SP's. I'll be quite surprised if they've only gone with 64 SP's for this, especially with the 25W TDP.

 

[Nemesis 1]

AMD is over with . The only thing left are diehard fans ,. We can compete with arm and intel . Sorry Amd has not proved it can compete with either. Its like the days befor 2006 the AMD guys just don't get it . Haswell is a better big chip and Silvermont will be a better little chip . I know this will be as I say . You Amd guys are on your knees praying to a god you don't believe in to 1 up intel . It won't happen . I seen AMDs quarterly reports . Their is alot of fantasy in this thread.

 

[zlatan]

Again you don't know what your talking about . The Atom phone chip will use Imagination gpu. The tablet chip will use Intels HD4000 graphics and were are talking below 8 watts chips . So you best take the cheer leading skirt off and try doing some real research.

I don't deal with phone chips, so I don't know what iGPU will they use. But Bay Trail is a tablet, mobile, and desktop Atom. There will be also an industrial version for embedded systems.
The chip will use 4 execution units, so it will not really an HD Graphics 4000 iGPU (with 16 EUs). The architecture is the same.

 

[inf64]

Nemesis sorry for asking this obvious (but kind of redundant) question but why are you here posting in this topic? I really can't see the reason. You don't bring anything valuable to the discussion and you only either flame others (directly) or post offtopic stuff in Kabini thread(atom this,atom that etc.).

I hope some moderator reads your posts since it's really becoming tiresome and all can agree on this,I'm sure.

 

[AtenRa]

Nothing confirmed but if Kabini is "10% bigger than Brazos at 28nm", then it basically has to mean it has 128 SP's. If it was only 64 SP's then it would be a lot smaller. I expect Temash will have 64 SP's though, ~50mm2 total die size.

Jaguar core is 10% larger than BobCat core at 28nm, not the Kabini SoC over Brazos.
But having at least 128 SPs is 99% possible, either VLIW4 or GCN over 80 VLIW5 in Brazos.

With 128SPs (VLIW4 or GCN) at the same frequency as 80SPs (VLIW5), we could have close to 50% more performance.

 

[inf64]

@ AtenRa

Fudo reports it can crossfire with 128SP next gen 8700M discrete part,for what it's worth. Whether this means it has to have 128 GCN SPs or not is unknown.

 

[AtenRa]

Ah right. Well either way, it's either a ~50mm2 chip with 64 SP's or a ~80mm2 chip with 128 SP's. I'll be quite surprised if they've only gone with 64 SP's for this, especially with the 25W TDP.

Well im expecting a Quad core Kabini with 128SPs for the iGPU to be larger than Brazos(75mm2), close to 80-90mm2 because of the larger Jaguar cores and larger Fron-End of the iGPU over the VLIW5.

Also, we still dont know if it will have dual memory channels or single like Brazos.

 

[Ventanni]

With 128SPs (VLIW4 or GCN) at the same frequency as 80SPs (VLIW5), we could have close to 50% more performance.

That, and some strong OpenCL performance too. Although, how much OpenCL is being used on heavy tablets?

 

[SiliconWars]

Well the 6950 is basically equal to or slightly faster than the 5870, that's 1408 VLIW4 SP's @ 800 MHz vs 1600 VLIW5 SP's @ 850 MHz. 64 GCN SP's should be slightly more powerful or at least as powerful as 80 VLIW5 at the same clocks, so I think that only 50% uplift for 128 SP's on 28nm is being a tad conservative.

If you add it all up - the extra shaders, the extra x86 cores, higher clocks, slightly higher bandwidth etc then 2x graphics performance seems the likely minimum entry point to me at the same TDP.

 

[NTMBK]

Yes, with half the power consumption, 25%+ higher performance for the CPU cores and at least 50% higher performance for the iGPU over Brazos, Intel doesnt have a chance.

Wait and see, wait and see! We don't know if these things will pan out- I can see a combination of 2 out of the 3 being true, but I'm not convinced they'll pull off all three. Let's hope they can (and that Intel can catch up too).

Temash tablets will literally put PCs in your fingertips

That's a messed up mental image D:

 

[AtenRa]

@ AtenRa

Fudo reports it can crossfire with 128SP next gen 8700M discrete part,for what it's worth. Whether this means it has to have 128 GCN SPs or not is unknown.

Yes i have read it thx

Well, either way(VLIW4 or GCN) it will destroy everything Intel have(iGPU) or will have in the next 2 or more years in the 9W and bellow.

Also, just imagine Tablets playing DX-11 games(1024x768) like DIRT, Civ 5 and more. With USB3/HDMI and the ability to use any Windows device like mouse, keyboards, Printers etc, and use your Windows applications.

You could have your PC with you 24/7 at the size of a 10-11" tablet

 

[Nemesis 1]

Nemesis sorry for asking this obvious (but kind of redundant) question but why are you here posting in this topic? I really can't see the reason. You don't bring anything valuable to the discussion and you only either flame others (directly) or post offtopic stuff in Kabini thread(atom this,atom that etc.).

I hope some moderator reads your posts since it's really becoming tiresome and all can agree on this,I'm sure.

I wouldn't be in here except all the comparing to intel chips . As I said you guys want to compare chips keep it to AMD products. I hate when You compate future AMD to present Intel . Compare future to future not future to past .

According to our estimates Airmont will offer 10 times more computing power and graphics performance compared to current chips and a 20 times more performance when compared to the Atom N270 . Atom N270 released in 2008 delivers a yield of 3856 MIPS (million instructions per second) at 1.6GHz. 14nm Atom Airmont at same time yields 79,000 MIPS performance which is same performance as that of AMD Phenom X6 1100T (yes, it sounds exaggerated, but that is the projection). Another good thing is Airmont will not consume power more than 15W.


14nm Airmont, 32nm Saltwell, 22nm Silvermont Release Date Info
However, before the arrival of Airmont in 2015, Intel has rescheduled their roadmap and would release Saltwell (32nm) in September in order to replace the current 45nm chip Bonnell, and 32nm chip Medfield. Also 22nm chip Silvermont will be released in 2013 which will be the first to use Tri-Gate Transistor technology like Intel Ivy Bridge Processor. These chips will also have significant improvements in Power consumption.