Kabini Rumors

[SiliconWars]

Tigerdirect has some, it seems to be VisionTek that makes most of it from what my search is showing.

http://www.tigerdirect.com/applicati...s=1.65v+sodimm

You're right that there isn't a lot of it around though, guess I was wrong on that.

 

[Exophase]

what about workloads that fit in 512k vs 2mb when you have 4 of them concurrently.

It's the same thing. The extra cache helps, but not having the extra cache doesn't automatically mean you'd instead be helped by more main memory bandwidth.

 

[itsmydamnation]

It's the same thing. The extra cache helps, but not having the extra cache doesn't automatically mean you'd instead be helped by more main memory bandwidth.

I bet you an internet beer because of the L2 prefetcher more main memory bandwidth increases CPU performance in more then just corner cases.

 

[Exophase]

I bet you an internet beer because of the L2 prefetcher more main memory bandwidth increases CPU performance in more then just corner cases.

Meaning you expect to see an improvement in what, most benchmarks tested? And you're saying, substantially greater scaling than what you see with scaling on high-end IB? Just for heavily multithreaded or for everything?

IB has prefetching too, probably better than what Jaguar so that doesn't make it better at being able to utilize available bandwidth.

With 64-bit DDR3 1066 overall bandwidth is 8.5GB/s. For a single 1.4GHz processor to saturate that it has to use 6 bytes per cycle. That's a large percentage of cycles that aren't logic, branching, memory operations that reuse in the cache, or wasted cycles/stalls due to branch mispredicts and main RAM latency that couldn't be hidden. Some parts of some programs will occasionally look for this but not for very long. It's such that statistically speaking, these parts don't tend to happen much on multiple cores simultaneously. You'll still see some of these little bursts get faster but for most programs it'll just result in a relatively small boost when averaged over the whole runtime of the program. Some entire tasks are highly stream oriented but they're not very standard so don't expect to see much of it.

 

[itsmydamnation]

Meaning you expect to see an improvement in what, most benchmarks tested? And you're saying, substantially greater scaling than what you see with scaling on high-end IB? Just for heavily multithreaded or for everything?

IB has prefetching too, probably better than what Jaguar so that doesn't make it better at being able to utilize available bandwidth.

No not at all, im not talking about IVB. I'm talking about jaguar. im saying that im willing to bet one internet beer that Jaguar CPU performance in more then just corner cases will see improved scaling with memory frequency, 1600 vs 1866 vs 2133.

BD/PD/SB/IVB dont seem to care about memory bandwidth once around 1333/1600. In the few benchmarks we have seen around jaguar is beating trinity per clock(FP/SIMD, int would be interesting), now given that we do see performance differences between 1066 vs 1333 vs 1600 on the CPU side with BD/PD/SB/IVB I think i have an even shot of not owing you an internet beer .

 

[Exophase]

I was just saying I don't expect the scaling to be much higher (if at all) than it is for IB, and I don't think that the difference in cache will be made up for with more main memory bandwidth very often. Stuff that scales well going from 1066 to 1600 on IB may scale well on Jaguar too.

This is completely ignoring anything that uses the GPU.

 

[itsmydamnation]

The other thing to consider is latency, jaguar doesn't have the same amount of OOOE resources as a big core and is clocked low. So assuming same timings 1066 vs 2133 1/2 the memory access time on a cache miss could start to add up.

 

[Exophase]

The other thing to consider is latency, jaguar doesn't have the same amount of OOOE resources as a big core and is clocked low. So assuming same timings 1066 vs 2133 1/2 the memory access time on a cache miss could start to add up.

Yes, and having more bandwidth alone does nothing to hide latency. When do you see DDR3 1066 and 2133 set with the same timings (in clock cycles)?

 

[itsmydamnation]

Yes, and having more bandwidth alone does nothing to hide latency. When do you see DDR3 1066 and 2133 set with the same timings (in clock cycles)?

Well you don't of course, but absolute access time it can still be around 30% from for pretty "standard" 1066 (7-7-7) to 2133(11-11-11). If you want to pay for it you can get 9-x-x 2133 which would be 50-60% faster as well. Cant really buy high end 1066 anymore.

my 1333 3x2gb sticks i bought a very long time ago for an insane amount of money only to have voltage issues on my X58 MB are running 1080 8-8-8. If i took that same amount of money a bought ram now I could by 48gb of 2400 @ 10-12-12 ..........:ninja:

 

[Arkadrel]

http://ultrabooknews.com
Did testing of that Acer Aspire V5 122P, with a Temash@1.0ghz.

He got that turbo working eventually, and @1.4ghz in cinebench 11.5 it scored 1.23.
(that is with Lame running in background useing ~6-7% cpu load).

I think amd claimed it could reach around 1.39 points in cinebench (@1.4ghz),
when your not doing other stuff in background ect.


Cinebench 11.5:
Intel Celeron G530T (sandybridge) (2.0ghz) (35w) = 1.54
AMD A6-1450 (Temash) (@1.4ghz) = 1.39
Intel Atom D2700 (2.13ghz) = 0.70

Its a decent little netop cpu.

 

[Khato]

Cinebench 11.5:
Intel Celeron G530T (sandybridge) (2.0ghz) (35w) = 1.54
AMD A6-1450 (Temash) (@1.4ghz) = 1.39
Intel Atom D2700 (2.13ghz) = 0.70

Just for a bit of fun let's get some numbers from the mobile world for the Intel side as well? After all, there you can find a SNB Celeron and Pentium that run at the same 1.4 GHz.

Intel Celeron 877 (SNB, 1.4 GHz, generic 17w TDP) - 1.07
Intel Atom Z2760 (1.8GHz max, sub 3w TDP) - 0.55

Of course better power figures for the A6-1450 are necessary in order to get a good comparison on efficiency, but I'd tend to say it's quite safe to say that Temash is a fair deal more efficient than SNB based products. And from those numbers SNB is roughly 1.5x faster in single threaded performance, again, more or less as expected.

I included the Z2760 number in there mostly just for the fun of it. There's no question that the current Atom architecture is too anemic to really put up much competition, even though it's likely comparable in efficiency. But unfortunately I can't find any Cinebench 11.5 results for the Z2580 which was used as the reference point in Intel's recent Silvermont performance estimate. Since the Z2580 is essentially a 2GHz version of the Z2760 and it's all pure conjecture anyway I figure there's no reason not to use the Z2760 score and Intel's claimed 2.5x 4T performance increase to guess at around 1.375 for Baytrail at the same sub 3W power usage.

 

[Arkadrel]

I figure there's no reason not to use the Z2760 score and Intel's claimed 2.5x 4T performance increase to guess at around 1.375 for Baytrail at the same sub 3W power usage.

^ impressive if that does happend.

I think the guy at Ultrabooknews.com, said the Temash @1ghz uses about 3watts (more than idle) while running 100% cpu load doing Cinebench 11.5.

It only scored about ~0.85 points there.

If baytrail really can do about 1.3 points at 3watts,
then its like ~50% more cpu perf at the same power as a temash apu is.

Intels Fabs are the real winners though, 22nm & FinFETs.

Too bad we ll have to wait like 6months or more to see, how Intels Baytrail performs.
Would love to see a side by side compairison via benchmarks, of Baytrail vs temash.

 

[Vesku]

Jaguar IPC for Cinebench 11.5 is just a bit behind Core2/Phenom II. The high side of Intel's estimates would put Silvermont at Nehalem IPC. Will be interesting to see what actual hardware does, but for the next 6 months or more AMD is alone in this segment. Current Atom has been completely lapped.

 

[inf64]

I think there is zero chance that 2way Silvermont core will come anywhere near Nehalem core's IPC. If it comes close to 65nm Core2 it would be quite a feat.

 

[Vesku]

I think there is zero chance that 2way Silvermont core will come anywhere near Nehalem core's IPC. If it comes close to 65nm Core2 it would be quite a feat.

Definitely no point in waiting 6+ months for it. If it hovers around Core 2 IPC then it won't be much different from AMD Jaguar other than a high likelihood of being more efficient due to process advantage. If it's faster than that it will be interesting to see if it supplants Intel's ULV Celeron and Pentium lines.

 

[Arachnotronic]

Definitely no point in waiting 6+ months for it. If it hovers around Core 2 IPC then it won't be much different from AMD Jaguar other than a high likelihood of being more efficient due to process advantage.

It'll fit in a 7" tablet.

Kabini/Temash won't.

 

[itsmydamnation]

It'll fit in a 7" tablet.

Kabini/Temash won't.

complete bull, their is a 3.6 watt TDP dual core.

 

[Arachnotronic]

complete bull, their is a 3.6 watt TDP dual core.

And that'll hold up against a 4C Silvermont in the sub-3W power envelope, how well?

Also, no connected standby for Temash...

 

[Khato]

Intels Fabs are the real winners though, 22nm & FinFETs.

Too bad we ll have to wait like 6months or more to see, how Intels Baytrail performs.
Would love to see a side by side compairison via benchmarks, of Baytrail vs temash.

No question there - Intel could easily be getting half of its efficiency advantage off its process.

Too bad we have to wait so long still for Baytrail. But AMD is definitely going to have a good offering until that point in the low end small notebook and tablet space with Kabini/Temash. Sure once you up the power envelope it'll have to compete against Haswell, but there's a marked price difference there. And on the lower end it's going to simply dominate the current Atom core.

 

[Arkadrel]

Are 2ghz models as high as the jaguar's go? anyone know? How high do those turbo to?

 

[Vesku]

The embedded Jaguar press release doesn't list turbo. Highest SKU is Quad 2 GHZ 25W TDP which should be similar to the consumer 25W model. Still possible the consumer version will have some sort of turbo function.

http://www.amd.com/us/press-releases/Pages/amd-embedded-gseries-2013apr23.aspx

 

[itsmydamnation]

And that'll hold up against a 4C Silvermont in the sub-3W power envelope, how well?

Also, no connected standby for Temash...

in all likelihood for what you use a 7inch win 8 tablet for absolutely fine. You think a 4C Silvermount SOC is going to run @ ~1.2ghzunder 3W TDP? Connected standyby is a rather large "meh", its not likel im using my 7inch tablet as a phone. Jaguar has much improved power gating over bobcat. So long as its battery life is comparable to other SOC's in the field it will be fine.

But that's getting away from the initial point anyway.

 

[SiliconWars]

Apologies if already posted.

http://www.cpu-world.com/news_2013/2013051101_Future_Kabini_APUs_spotted_in_HP_255_laptop.html

 

[monstercameron]

here is an update

http://www.umpcportal.com/2013/05/amd-temash-a6-1450-performance-acer-aspire-v5/

 

[inf64]

here is an update

http://www.umpcportal.com/2013/05/amd-temash-a6-1450-performance-acer-aspire-v5/

Thanks for the link. Finally some hard data when it comes to power:

Test done on high-performance profile under battery power.
Test 2 – ALU
Standing usage immediately before test: 7.9W (7% utilization, ave 1Ghz clock in quiescent state)
Test max power drain (ALU Test): 12.3W (94% CPU load by Crystalmark Process)
Test power: 4.4W
Normalised: 790mw/Core/Ghz
Test 2 – FPU
Standing usage immediately before test: 8.1W (7% utilization, ave 1Ghz clock in quiescent state)
Test max power drain (FPU Test): 13.1W (94% CPU load by Crystalmark Process)
Test power: 5W
Normalised: 900mw/Core/Ghz

Estimated max power usage of 4 Temash CPU-cores at 1Ghz, 100%: 3.6W (+/- 10%)

Estimated max power usage of 4 Temash CPU-cores at 1.4Ghz, 100%: 5W (+/- 10%)