[PCPER] beema and mullins

[inf64]

This smells Richland all over again, when we got promised everything but in reality got only small improvements.

Not really.

PCmark8- Mullins @ 4.5W scores 1809pts while Kabini A6-1450 @ 8W scores 1343pts. 35% higher score with almost 2x less specced TDP.

3dmark11- Mullins(4.5W) scores 22% higher vs Kabini A6-1450. GPU clock is likely again ~20% higher on Mullins Vs A6-1450, TDP spec of Mullins SoC is almost 2x lower.

Pcmar8- Beema (15W) scores 2312pts while Kabini A6-5200 (25W) scores 1861pts. Beema scores 24% higher while having 40% less TDP than A6-5200. This implies AMD is able to clock up Beema part to ~2.5Ghz vs 2Ghz on Kabini while consuming roughly 40% less power at that clock Vs Kabini @ 2Ghz. That's massive improvement if true (unless someone would like to suggest that tweaked Kabini, which is Beema has much higher IPC- highly unlikely scenario). Score of 2312pts(Vs 1861 of Kabini) is there, there are just two ways to get it: clock or IPC or combo of both (IPC being minor part in combo).

3dmark11- Beema(15W) scores 20% higher vs Kabini A6-5200 (25W). GPU clock is likely again ~20% higher while SoC's TDP spec is supposedly 40% lower than A6-5200.

 

[Gikaseixas]

Graphics seems to be the savior for these puppies. I believe the cpu side will continue to lag behing Intel offerings

 

[mrmt]

Not really.

PCmark8- Mullins @ 4.5W scores 1809pts while Kabini A6-1450 @ 8W scores 1343pts. 35% higher score with almost 2x less specced TDP.

If the article is to be believed Puma is an improved Jaguar: More or less the same units, same number of units some improvements on the uncore and that's it and they are built on the same process. How are they getting 2x the same performance per watt?

My guesses:

- Kabini/Temash power management was the crappiest thing on earth after the first P4.

- They are messing with the TDP specs of the new chips, and with turbo it is going *way* beyond the TDP for far longer than the reasonable time.

- They are going for some fancy process, like 28nm FDSOI and the thing is really worth it.

There is no way to get twice the performance per watt on the same process with just an improved design.

 

[inf64]

Note that they list the true TDP and not SDP in the image I attached. SDP for that Mullins part is actually 2W while "TDP" is 4.5W. These two are clearly not one and the same.

FD-SOI is not in line with AMD's previous statements about going 28nm bulk across all APU products in 2014.

The clock speeds of both Mullins and Beema are easy to extrapolate. They are almost exactly 20% higher on both CPU and GPU in the respective benchmarks (PCmark-> CPU is boosted to 2.5Ghz Vs 2Gz; 3dmark11-> GPU is boosted by 20%).

I found this nice table on SA forum, credit to user "van".

Mullins part (TDP 4.5W, SDP 2W) has 97% performance of Kabini A6-5200 which is 25W TDP part in Pcmark8 (stresses CPU) and 80% performance of A6-5200 in 3dmark11 (which is a GPU benchmark).

 

[mrmt]

The clock speeds of both Mullins and Beema are easy to extrapolate. They are almost exactly 20% higher on both CPU and GPU in the respective benchmarks (PCmark-> CPU is boosted to 2.5Ghz Vs 2Gz; 3dmark11-> GPU is boosted by 20%).

So they got 20% more clock with almost 50% less TDP, at the same process? And you find this credible?

 

[monstercameron]

So they got 20% more clock with almost 50% less TDP, at the same process? And you find this credible?

maybe it is conditional, maybe fully stressed we see kabini power levels but idle and moderate loads we see a drop in power from overall soc power...

 

[Vesku]

I agree with the idea it's in large part due to better, any at all in the case of A4-5000 and A6-5200, turbo modes and running benches that only stress the CPU or the GPU and not both.

 

[inf64]

So they got 20% more clock with almost 50% less TDP, at the same process? And you find this credible?

I'm just telling you what the footnote states. I have no way of knowing the details and the real power draw.

 

[Ancalagon44]

I for one am very interested in Beema. I hope some good 11.6 laptops using it come out next year.

 

[mrmt]

I'm just telling you what the footnote states. I have no way of knowing the details and the real power draw.

And I'm asking your opinion. Do you think this is credible?

 

[inf64]

Well it's not fake if that is what you are asking. Or are you saying now AMD is straight out lying to everyone, investors included?
You have the scores, you have the specced TDPs. Draw your own conclusions. As for me, no I don't think AMD is lying, they really got those scores and the 2x perf./watt improvement they state in slides means they also really got those low power levels.

 

[JustMe21]

I think if the battery life is good and the heat generation is minimal, then AMD could make it in the tablet market. What I would like to see them do is encourage OEMs to make tablets with 4 GB RAM standard, offer 64/128 GB drives instead of 32/64 GB like Atom has, and install Windows 8.1 64-bit.

They would also need to make a deal with Microsoft to be able to offer Office 2013 for free like Atom tablets have, because one of the weakest point of Windows 8 is the lack of apps in the Store.

In the end, Mullins' performance and Bay Trail's performance aren't earth shattering, so we need battery life and features.

 

[AtenRa]

I agree with the idea it's in large part due to better, any at all in the case of A4-5000 and A6-5200, turbo modes and running benches that only stress the CPU or the GPU and not both.

Much like Intel Tablet BayTrail review

 

[Gideon]

And I'm asking your opinion. Do you think this is credible?

I doubt the actual power-draw is 2x less for the same performance, however TDP, why not ?

About A6-1450:

Cinebench R11.5 Single (~1.4 GHz CPU): 13.7 Watt
Cinebench R11.5 Multi (~1.1 GHz CPU): 14.9 Watt
Prime95 large FFTs (~1.0 GHz CPU): 16.9 Watt
FurMark (~1.2 GHz CPU): 21.2 Watt
Prime95 + FurMark (~0.85 GHz CPU): 21.9 Watt
3DMark06: 19.5 Watt

In Jaguar, the CPU power draw is nowhere near as high as GPU-s, when the other one is totally idle. With an aggressive turbo mode they could easily fix this.

And considering that jaguar perf/W probably isn't that great as some ARM chips done on the same process, it certainly SHOULD be some room for power/pref optimizations. (Yeah I know that comparison is very hard to make, as we don't know the ARM chip TDPs, nor can we compare X86 to ARM easily, but still)

After all this IS a new iteration of the chip, not just a respin like Brazos 2.0. There is no way the would have gotten a A5 core and connected standby otherwise. Some uplift can also be attributed to there

 

[monstercameron]

I doubt the actual power-draw is 2x less for the same performance, however TDP, why not ?

About A6-1450:


In Jaguar, the CPU power draw is nowhere near as high as GPU-s, when the other one is totally idle. With an aggressive turbo mode they could easily fix this.

And considering that jaguar perf/W probably isn't that great as some ARM chips done on the same process, it certainly SHOULD be some room for power/pref optimizations. (Yeah I know that comparison is very hard to make, as we don't know the ARM chip TDPs, nor can we compare X86 to ARM easily, but still)

After all this IS a new iteration of the chip, not just a respin like Brazos 2.0. There is no way the would have gotten a A5 core and connected standby otherwise. Some uplift can also be attributed to there

it seems they will have to fiddle with units to get power just right on a per workload basis...maybe this just wasn't ready with temash/kabini and maybe temash had a prototype version of the turbo mechanism....

 

[mrmt]

Well it's not fake if that is what you are asking. Or are you saying now AMD is straight out lying to everyone, investors included?
You have the scores, you have the specced TDPs. Draw your own conclusions. As for me, no I don't think AMD is lying, they really got those scores and the 2x perf./watt improvement they state in slides means they also really got those low power levels.

IMO they are being very flexible with their TDP levels, like they were with the FX 8350 *and* they are taking a very optimum case for their turbo implementation.

But the second hypothesis of AMD lying out to everyone, investors included, isn't too far fetched, is it? Remember Barcelona (Senior VP saying "40% over clovertown"), Bulldozer (CFO saying 35% over Thuban), AMD and OpenCL (one year before their consumer drivers support it), Kaveri (1050 GFLOPS, and that was the new management team), and others.

 

[inf64]

Well we will have to wait and see about this, right?
A question. Would you like this to be true or not?

 

[mrmt]

Well we will have to wait and see about this, right?
A question. Would you like this to be true or not?

It doesn't matter, does it? Me wanting something does not make AMD results better or worse.

 

[inf64]

I'm asking you that because the way you post (to me) it seems as if you just wish this to be untrue. Just my observation.

 

[mrmt]

I'm asking you that because the way you post (to me) it seems as if you just wish this to be untrue. Just my observation.

Nope, I just smelled something fishy on the slides and AMD historical data doesn't help. But if Intel comes tomorrow saying that Skylake will have double the performance per watt of Broadwell you'll find me having a similar reaction, same with Qualcomm, Samsung, etc.

 

[Arkadrel]

I doubt the actual power-draw is 2x less for the same performance, however TDP, why not ?

About A6-1450:

Cinebench R11.5 Single (~1.4 GHz CPU): 13.7 Watt
Cinebench R11.5 Multi (~1.1 GHz CPU): 14.9 Watt
Prime95 large FFTs (~1.0 GHz CPU): 16.9 Watt
FurMark (~1.2 GHz CPU): 21.2 Watt
Prime95 + FurMark (~0.85 GHz CPU): 21.9 Watt
3DMark06: 19.5 Watt
In Jaguar, the CPU power draw is nowhere near as high as GPU-s, when the other one is totally idle. With an aggressive turbo mode they could easily fix this.

And considering that jaguar perf/W probably isn't that great as some ARM chips done on the same process, it certainly SHOULD be some room for power/pref optimizations. (Yeah I know that comparison is very hard to make, as we don't know the ARM chip TDPs, nor can we compare X86 to ARM easily, but still)

After all this IS a new iteration of the chip, not just a respin like Brazos 2.0. There is no way the would have gotten a A5 core and connected standby otherwise. Some uplift can also be attributed to there

^ you realise that those are the power loads of the intire laptop right?

maximum performance profile, maximum brightness, Wi-Fi on: 10.3 Watt

^ idle laptop uses 10.3 watts @ maximum perf. profile.
(wi-fi on, max brightness display)

6watts or so is alone from the display + wifi.
Then im sure the ram&hdd use some power too, same with motherboard and chipsets.

And considering that jaguar perf/W probably isn't that great

Intire laptop uses 21.2 watts of power, when running Furmark.

10-13 of those watts are probably "not" the APU, but the rest of the system (motherboard, ram, hdd, chipsets, wifi, display ect).

A ARM system would have the same amount of watts used on those parts as well.

 

[Abwx]

I'm asking you that because the way you post (to me) it seems as if you just wish this to be untrue. Just my observation.

Short sellers and other HFs often bet on a company s bankruptcy,
you wont even have to buy back the short sold stocks , 100% ROI.

 

[Arkadrel]

Mullins part (TDP 4.5W, SDP 2W) has 97% performance of Kabini A6-5200
(which is 25W TDP part) in Pcmark8 (stresses CPU) and 80% performance of A6-5200 in 3dmark11 (which is a GPU benchmark).

Haveing their next 4.5watt APU almost give same performance (80-97%) as the old 25watt TDP one, is really impressive.

So much so I can understand why some would be skeptical.

 

[monstercameron]

Haveing their next 4.5watt APU almost give same performance (80-97%) as the old 25watt TDP one, is really impressive.

So much so I can understand why some would be skeptical.

with aggressive turbo it is surely possible,

if an a6-1450 could turbo a single core upto 2ghz and all 4 cores upto 1.5ghz with the gpu section downclocked, they could stay within tdp and still perform like the faster 25W apu in certain cpu-bound scenarios...

the same could be said for the gpu, the a6-5200 is clocked at 600MHz, if the a6-1450 could boost upto 600MHz while possible keeping a single core at 1.5GHz or all cores at 800MHz, in certain gpu bound scenarios they could be close...

 

[krumme]

What about TSMC HPM process in play here?

Look at the QQ s600 to s800 transition. Same core, but the speed gains were incredible. Powerusage was way down also.