Kabini Rumors

[AtenRa]

On AMDs own site. I wrote a lengthy post about it here:

http://forums.anandtech.com/showthread.php?p=34720015&highlight=#post34720015

I was totally shocked about it in a worst possible way. But given AMDs tendency of adhering bad news, that's what I expect from Temash until proven otherwise.

This is the Tablet power consumption(Including the screen etc). For the 28nm Temash to have higher consumption in HD Video Playback than 40nm Hondo it means that either the Screen is bigger in size or the measurement was made at a higher Brightness settings, or both.

 

[Olikan]

Dresdenboy found some benchmarks....=)

http://citavia.blog.de/2013/04/17/2-ghz-amd-jaguar-benchmarks-15761535/

direct link...for the comparison
https://www.osadl.org/CPU-benchmarks.qa-farm-cpu-benchmarks.0.html

jaguar is in the botton.... the "r9s1" on the list....first table is single core, the other is multi

at lest on those benches....jaguar have ipc similar to nehalem O.O

 

[Gideon]

This is the Tablet power consumption(Including the screen etc). For the 28nm Temash to have higher consumption in HD Video Playback than 40nm Hondo it means that either the Screen is bigger in size or the measurement was made at a higher Brightness settings, or both.

AMDs materials contain more than System Only power consuption. There is also the APU power consumption and in fact it's almost 1 on 1 the same as Z-60 + Northbridge.

Let me reiterate it, this time looking at APU consumption alone. This is Brazos 2.0 based Z-60:

And from AMDs own materials for A4-1200 this time APU only power consumption :

http://www.amd.com/us/products/notebook/tablets/Pages/tablets.aspx#4

The AMD “Larne” reference platform is projected to measure APU power at 1.2 W at idle, 1.40 W during web browsing, 2.35 W during video playback and .02 W during a system S3 “sleep” state. Total system power for the reference platform is projected at 2.8 W at idle, 3.7 W during web browsing, 5.3 W during video playback and .07 W during a system S3 “sleep” state

So let's compare the two:

Z-60 (Brazos 2.0) 2 cores @1 Ghz:
idle: 0.75 + 0.55 = 1.3W
web: 1.12 + 0.6 = 1.72W
video: 1,57 + 0.68 = 2.25W

A4-1200 (Temash) 2 cores @1 Ghz
idle: 1.2 W
web: 1.40 W
video: 2.35 W

So yeah, it seems I was indeed too pessimistic, claiming that there is NO change as web-browsing power draw seems to decrease 19%. While HD-playback gets worse by 5% i can imagine it can actually deliver smooth 1080p framerates (unlike Bobcat in some cases).
However the change is still very very small (nowhere near the 50% improvement from the TSMC slides you posted for instance). Idle power draw is particularly disappointing IMO.

 

[Gideon]

FYI: they talk about system power consumption.

In all those graphs the APU has the smallest impact on overal powerconsumption. So you are running against diminishing returns. even a 50% APU power reducation will not lead to significant different charts.

And by all this you are also comparing exactly the same workloads under exactly the same system running at the same temperature in the same environment with the same OS.

The footnotes also include APU only powerdraw.

Bear in mind that all this data is from internal AMD Performance Labs and it is for reference systems. I would really hope they use as similar settings as possible for these, otherwise what's the point ?

I'm also rooting for Temash to become an excellent Tablet CPU smack in the middle of Core and Atom lineup with better prices. However AMDs information makes me very careful on the power-consumption side for now. I'd really like for them to release more numbers to clear any doubt. Hell i'd really like to see some real-life measurements on Z-60 as well while that's the case.

 

[Arkadrel]

Idle power draw is particularly disappointing IMO.

Yeah its kinda sad they couldnt get that lower.

However the change is still very very small (nowhere near the 50% improvement from the TSMC slides you posted for instance).

IPC arnt the same for both though (CPU vs CPU).
So even if both are 1ghz, the larne platform will give better performance.
The GPU on the jaguar one is much more powerfull too, which I figour also again means smoother web browseing.

So even if web browseing went from 1.72w -> 1.40w (~23% reduction),
the performance at 1.40w is going to be higher.



Still:

http://www.anandtech.com/show/6529/busting-the-x86-power-myth-indepth-clover-trail-power-analysis/2

Total platform power consumption (start screen - long idle):

Microsoft Surface RT (tegra 3) (10.6" screen ?) = ~3.2watts

Total system power for the reference platform is projected at 2.8 W at idle, 3.7 W during web browsing, 5.3 W during video playback and .07 W during a system S3 “sleep” state.

Seems close to the ~2.5watts or so, the Atom z2760 has (Acer w510 has a 10.1" screen).
The A4-1200 (Temash) inside the AMD "Larne" platform, was supposedly with a 13.3" screen.

Its hard to do a apples to apples comparison when the platforms vary so much.

 

[Gideon]

Wow, nice find. While I'm aware, it's apples to oranges then looking at that Anandtech test I'm a bit more hopeful

It would be really helpful, if some independent review site could benchmark Z-60 power/performance. I'm curious however if Vizio and Toshiba plan to release their Z-60 tablets at all, considering that Temash could arrive in Q2 already. After all the first Temash prototypes were shown in January and a really nice OEM one in February.

 

[Exophase]

You have just said that going to 28nm Dual Core BobCat @ 1GHz will get you half the power, meaning adding two(2) more cores will get you the same power as a Dual Core 1GHz @ 40nm. And we haven’t count the higher efficiency of the Jaguar cores.

No, 40-50% lower means value / 1.4 to 1.5, not value / 2. TSMC 40nm to 28nm does not double power efficiency.

The design is more efficient, yes. It also has higher perf/MHz. Those two things tend to work against each other when estimating power consumption at some clock speed. All the other TDP numbers for Kabini also speak against this idea of something like 2.3x power efficiency.

 

[AtenRa]

AMDs materials contain more than System Only power consuption. There is also the APU power consumption and in fact it's almost 1 on 1 the same as Z-60 + Northbridge.

Let me reiterate it, this time looking at APU consumption alone. This is Brazos 2.0 based Z-60:

And from AMDs own materials for A4-1200 this time APU only power consumption :

http://www.amd.com/us/products/notebook/tablets/Pages/tablets.aspx#4



So let's compare the two:

Z-60 (Brazos 2.0) 2 cores @1 Ghz:
idle: 0.75 + 0.55 = 1.3W
web: 1.12 + 0.6 = 1.72W
video: 1,57 + 0.68 = 2.25W

A4-1200 (Temash) 2 cores @1 Ghz
idle: 1.2 W
web: 1.40 W
video: 2.35 W

So yeah, it seems I was indeed too pessimistic, claiming that there is NO change as web-browsing power draw seems to decrease 19%. While HD-playback gets worse by 5% i can imagine it can actually deliver smooth 1080p framerates (unlike Bobcat in some cases).
However the change is still very very small (nowhere near the 50% improvement from the TSMC slides you posted for instance). Idle power draw is particularly disappointing IMO.

HD Video Playback on the Hondo test was made at 720P with 60nits of Brightness when Temash was tested at 1080p with 100nits of Brightnes.

http://www.amd.com/us/products/notebook/tablets/Pages/tablets.aspx#4

Power projections based on calculations carried out by AMD Performance Labs measuring total system and individual component power at Windows Idle and under various system loads while web browsing and/or viewing a 9:57 minute online video in h.264 format, viewed at 1080P setting at 100 nits.

 

[sm625]

Dresdenboy found some benchmarks....=)

http://citavia.blog.de/2013/04/17/2-ghz-amd-jaguar-benchmarks-15761535/

direct link...for the comparison
https://www.osadl.org/CPU-benchmarks.qa-farm-cpu-benchmarks.0.html

jaguar is in the botton.... the "r9s1" on the list....first table is single core, the other is multi

at lest on those benches....jaguar have ipc similar to nehalem O.O

It scores the same dry/whet as a Xeon E3-1220L, a 2 core 4 thread sandy that is close to an i5-2450M and faster than an i5-3437U. I find this extremely difficult to believe. The multi-threaded Excel test result is higher than the i7-920. Sure lol.

 

[sontin]

No, 40-50% lower means value / 1.4 to 1.5, not value / 2. TSMC 40nm to 28nm does not double power efficiency.

Sure it does: For example Kepler and Tegra 4.

 

[Exophase]

Sure it does: For example Kepler and Tegra 4.

Okay, let me try to make this clearer. What I'm saying is that the same design (a direct shrink, same clocks) won't result in 2x the power efficiency going from TSMC 40nm to 28nm. If you also change the design to use substantially more transistors as is afforded by the shrink you can gain higher power efficiency. But that requires a problem which has a lot of untapped parallelization that can be exploited. That's much more true for GPUs than it is for CPUs, and outside of doubling the SIMD width Jaguar isn't going substantially wider. Nor have they increase their transistor budget by a huge amount.

I'm not going to get into claims with Tegra 4 because I don't want to turn this into another nVidia argument.

 

[Arkadrel]

It scores the same dry/whet as a Xeon E3-1220L, a 2 core 4 thread sandy

Xeon E3-1220L @ 2.20 ghz (~20w TPD) = 50953 dry , 5222.5 whet
Kabini 4C 25W (9830) @2.00 ghz (~25w TPD) = 53,497 dry , 4901.9 whet

Why is that so hard to believe?

The Xeon is a 32nm intel cpu, the Kabini is a 28nm AMD cpu.

The Intel CPU is from 2011, the AMD Kabini isnt even launched yet.


it matches up with this:

Another Sandy bridge (2c/4t) vs Temash (4c/4t) both at 1.4ghz.
Where the Temash is faster.

Maybe HT just isnt as good as haveing a real core? and in highly multi-threaded workloads, AMD can beat Sandy bridge CPUs in perf/watts (if they have a 2c/4t vs 4c/4t senario).

 

[inf64]

Xeon E3-1220L @ 2.20 ghz (~20w TPD) = 50953 dry , 5222.5 whet
Kabini 4C 25W (9830) @2.00 ghz (~25w TPD) = 53,497 dry , 4901.9 whet

Why is that so hard to believe?

The Xeon is a 32nm intel cpu, the Kabini is a 28nm AMD cpu.

The Intel CPU is from 2011, the AMD Kabini isnt even launched yet.

Kabini is an APU . It shares the TDP budget with iGPU, while that Xeon has no iGPU and whole TDP budget is assigned to its x86 cores. Also Xeon(or better yet Core design that makes Xeon) was designed for server/high perf. desktop market first and then was pushed down into lower power segment while Kabini is designed not for performance but for perf./watt. The x86 dedicated die area for Kabini and that Xeon (even if it was on 28nm) would be very much different, Xeon taking up much bigger die area for x86 cores(cache not counted in).

 

[Arzachel]

@Gideon: Isn't a straight Hondo to Temash APU powerdraw comparison moot, because the later has the FCH on die?

 

[Arkadrel]

@Arzachel the figours hes useing include the FCH.

Also Xeon(or better yet Core design that makes Xeon) was designed for server/high perf. desktop market first and then was pushed down into lower power segment while Kabini is designed not for performance but for perf./watt.

^ yep. Thats why I think its possible for a low powered Temash to beat a sandy bridge at equal wattage (atleast in certain senarios).

Thats just the issue though.... Intel wont be useing 32nm Sandy bridges downclocked.
Intel has newer products too at low power that ll probably beat the temash.


That said.... The jaguar cores are nice little things. Im surprised they can scale it so well.

 

[mikk]

E3-1220L has two cores while Kabini four cores. This is a massive advantage in benchmarks with good core scaling.

 

[Arkadrel]

@Mikk

We know that, its just...
People have a image of AMD as being much much slower than Intel.
Also of AMD being power in-effecient.

And of sandy bridge as being something magical with IPC so high even twice the cores of a AMD cpu, wont beat it.

Which apparntly isnt true.

You see a 20w TPD Intel Xeon sandy, get beat by a 25w TPD Temash, in multi threaded Apps.
(the xeon is without a GPU? and no FCH, unlike the temash).

 

[happysmiles]

I think it's going to be a good year for AMD APU buyers (unlike the past 3 years.)

 

[Exophase]

Dhrystone is a worthless benchmark and using a bunch of separate Dhrystone instances to try to simulate a multithreaded workload is even worse.

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.

 

[Olikan]

Also Xeon(or better yet Core design that makes Xeon) was designed for server/high perf. desktop market first and then was pushed down into lower power segment while Kabini is designed not for performance but for perf./watt. The x86 dedicated die area for Kabini and that Xeon (even if it was on 28nm) would be very much different, Xeon taking up much bigger die area for x86 cores(cache not counted in).

yes, and?
it's xeon that is pretending to be a low power here...there is ALOT of transistors used just to it reach >3Ghz

it just means that amd made a good low-power cpu, and high power ones can't directly compete... like every damm low-power cpu should be in the first place

 

[Khato]

Dresdenboy found some benchmarks....=)

http://citavia.blog.de/2013/04/17/2-ghz-amd-jaguar-benchmarks-15761535/

direct link...for the comparison
https://www.osadl.org/CPU-benchmarks.qa-farm-cpu-benchmarks.0.html

jaguar is in the botton.... the "r9s1" on the list....first table is single core, the other is multi

at lest on those benches....jaguar have ipc similar to nehalem O.O

Let's expand that comparison to see if we can't draw some actual conclusions. r2s0 is the 2.2GHz 2c/4t SNB based Xeon, r3s8 is the 3.2GHz 6c Phenom II X6 1090T, r4s6 is the 3.4 GHz i7-2600k, and lastly r9s1 is reportedly Jaguar.

Box    kDhry    Whet      Execl     kCopy    kPipe    Index    Effective
r2s0   51290    5221.2    7489.2    202.69   1840.0   1246.9   17.04. 06:11
r3s8   69032    14252.7   16161.8   172.46   4928.5   2028.7   17.04. 06:12
r4s6   180738   31378.1   17927.1   284.95   4218.0   2950.4   17.04. 06:11
r9s1   53497    4901.9    6504.6    80.34    1378.2   825.8    17.04. 06:11

So first, what happens if we normalize frequency to 2GHz?

Box    kDhry    Whet      Execl     kCopy    kPipe    Index    Effective
r2s0   46627    4746.5    6808.4    184.26   1672.7   1133.5   17.04. 06:11
r3s8   43145    8907.9    10101.1   107.79   3080.3   1267.9   17.04. 06:12
r4s6   106316   18457.7   10545.4   167.62   2481.2   1735.5   17.04. 06:11
r9s1   53497    4901.9    6504.6    80.34    1378.2   825.8    17.04. 06:11

Now why don't we also go ahead and normalize to 4c or 2c/4t?

Box    kDhry    Whet      Execl     kCopy    kPipe    Index    Effective
r2s0   46627    4746.5    6808.4    184.26   1672.7   1133.5   17.04. 06:11
r3s8   28763    5938.6    6734.1    71.86    2053.5   845.3    17.04. 06:12
r4s6   53158    9228.9    5267.7    83.81    1240.6   867.8    17.04. 06:11
r9s1   53497    4901.9    6504.6    80.34    1378.2   825.8    17.04. 06:11

My conclusion? Yeah, you can't compare these scores to one another as they clearly aren't keeping all other parameters constant. In kDhry, it's reporting Jaguar at 1.86x the IPC of the Phenom II. Then the Whet score for the normalized i7-2600k is 1.95x that of the Xeon despite them both being SNB based... whereas the remaining four tests reverse that and have the Xeon anywhere from 1.29x to 2.2x faster than the i7 (not too surprising given that they're throughput tests and hence not likely influenced by core count as much/at all.) (Not included here, but another good laugh is comparing the r0s4 Atom N2800 results against the r0s5 Atom N2800 results )

Edit: Oh, and since I guess not everyone might know such... Dhrystone 2 is a benchmark from 1988 and the original Whetstone was written in 1972. To say that they're outdated benchmarks is an understatement.

 

[SocketF]

My conclusion? Yeah, you can't compare these scores to one another as they clearly aren't keeping all other parameters constant.

Try to limit the view on x86_64 platforms and 3.6 kernels or newer. Otherwise the numbers are really strange.

 

[Khato]

Try to limit the view on x86_64 platforms and 3.6 kernels or newer. Otherwise the numbers are really strange.

I only didn't limit it as such to make exactly that point. Not only are the benchmarks themselves of limited use, but they are not necessarily comparable to one another due to other variables not being held constant. That's not to say that they're necessarily bad, they're simply meant for a different audience.

 

[Olikan]

Try to limit the view on x86_64 platforms and 3.6 kernels or newer. Otherwise the numbers are really strange.

to be fair, the "Ipc" increase from bulldozer to piledriver, was very close to reality...http://citavia.blog.de/2012/04/08/trinity-piledriver-performance-13460109/

maybe this benchmark is kinda like superPI... it will show if something got better or worst, but it only works well using similar cores...and comparison using diferent arqs, are useless

 

[inf64]

it just means that amd made a good low-power cpu, and high power ones can't directly compete... like every damm low-power cpu should be in the first place

Oh I don't dispute that at all . I agree with it.