Trinity vs Ivy Bridge

Anand said no turbo modes are used for A4-5000.

 

Oh, right, silly me. Case closed, then.

 

Intel's TDP is more of an average, isn't it ?

Anyway, The Core i3 has 50% higher performance than Kabini in F1, so the 70% higher power consumption isn't that unexpected, IMHO.

Still, Kabini looks fairly good in its niche.

Cheers

 

What if you have a 1.6GHz CPU with multiple cores, and in a situation one core is running at 1600MHz while the other ones are clocked down to 800MHz. This feels like turbo to me. But all four cores can run at 1600MHz, here's a difference. But what if power/temperature are monitored and you find yourself, exceptionnally or because of GPU use throttling down frequency to respect the TDP.. this feels like turbo again.

"Turbo" is a pretty arbitrary notion to me, though I wouldn't complain (at all ) if I have a processor with it.

 

Uhh, that's the full system power they are measuring right? i.e. not the chip power (17W).

 

Intel TDP is not TDP.
AMD TDP is.

If my memory holds true, Intel uses since years a fake measurement (the typical values under typical load) whereas AMD tells the value for worst load case.

So, ~double or such (ok, maybe a bit less...) Intel TDP all times...

 

Uh, yes, Intel's TDP is TDP. Otherwise Intel chips would burn.

 

Yes, without the laptop's own screen.

This is why we keep seeing weird results for the "17W" ULV i3's depending on system. Because the maximum turbo clock speed is based on temperature, better cooling will allow for higher clocks, but based on those results (and other's I've seen) there's no way that these HD4000 ULV's are only drawing 17W at maximum graphics turbo. It's more like 25W at least - there is simply no way the rest of the system without screen is drawing another 17W on top of the supposed maximum TDP of the chip.

Most power consumption tests do not test during gaming so Intel is getting away with advertising these as 17W chips, which is true for CPU loads but not for GPU loads unless the cooling is only adequate to dissipate 17W, at which point the turbo will be a lot lower and performance with it.

edit - that's my theory, might be bullshit.

 

It's only allowed to exceed the chip TDP for very short periods of time - it can't sustain an average TDP higher than the design point.

There's no need for you to speculate on this... just download Intel GPA and look at the power numbers for the GPU/CPU/Socket yourself in the system analyzer.

It's not correct. Like I said, download some monitoring tools and look for yourself. Not sure why you're on a crusade here in any case... process tech shines more the more TDP limited stuff is, and quite frankly, HD4000 is simply not designed to be wide enough to scale up to higher TDPs. Haswell GT3 will increase that of course, so stay tuned.

 

How short are benchmarks though?

 

By short I'm talking like ms level; at most seconds, but I don't recall the details. Definitely not long enough to skew benchmark runs. It also will typically happen - if at all - in the "warm up" phase of benchmarks where the results get thrown out anyways, as that's what heats up the chip quickly. Once it hits a steady state it obviously can't do that.

 

I'm sure I read in some Intel white paper that it was capable of turbo'ing past the TDP limit for up to a minute. You'll know better than me I guess but i'll try to find it again.

Like I said though I've seen this kind of thing before with older 35W chips, where the intel systems are performing much worse in terms of graphics efficiency (the other benchmarks like web browsing etc are generally tied as you would expect from 35W parts).

http://techreport.com/review/21099/amd-a8-3500m-fusion-apu/7



There's no doubt Intel has caught right up in performance since those days but I've always been suspicious of how exactly they got there while lowering TDP so much as well. In the end there is a lack of data on intel's power efficiency while gaming.

Anand says turbo can exceed the TDP limit "until the silicon gets up to temperature" - http://www.anandtech.com/show/6695/microsoft-surface-pro-review/7

A CPU turbo slide mentioning 1 minute extra -

 

Well, mostly by starting to actually care beyond just being "good enough"

Agreed, but same for everyone. Power efficiency for a given fixed task is an area that is really poorly handled by reviewers at the moment. Ideally you want some sort of "joules per time-demo" style measurement between systems, normalizing other factors, but that is not always easy.

That said, as far as users are concerned, ultimately it comes down the performance level provided in a given form factor and for how long (battery). Another way to measure might be clamping to a fixed frame-rate that everyone can hit (vsync or half vsync) and seeing how long systems with similarly sized batteries run. The TechReport measurement is almost there, but failed to normalize the workload (i.e. time demo or frame cap). Their newer articles do a better job of this, but I haven't seen them do it for a gaming workload yet.

Like I've mentioned though, Intel already exposes the power measurements for the various parts of the chip in GPA and other tools via MSRs, so take a look if you're curious.

 

That is a fair point with game benchmarks, the overall time spent under load will be the same for each chip (unless it's a set number of frames they render, instead), so the ones that perform better cannot finish the task sooner. This is however, a real world scenario - unless, you cap frames to an obtainable frame rate.

Regardless, I think we are seeing the benefits of Kabini actually being a full SoC here, coupled with the lack of turbo which would sometimes leave the load somewhat unbalanced depending on whether the task is CPU intensive or GPU intensive. For timed tasks, this would simply mean it is slower but spends longer - and more power - completing the task. A theoretical Kabini with turbo should use the same amount of power by completing the task sooner. With most game benchmarks, this is not the case.

 

Supposedly the A8 Llano mobiles would throttle themselves if they ran on battery:

http://www.legitreviews.com/article/1636/5/

It's not so much as Intel running past their TDP as possibly AMD preventing their CPUs from reaching their TDP limits on battery.

 

That doesn't seem to be the case here. We have reviews comparing power use to bobcat and the improvements are more than the delta between their respective TDP's:

http://www.legitreviews.com/article/2197/4/

Some of this is due to the lack of a chipset, but we can already see that it performs more than 20% (rough IPC increase) better than bobcat in these same benchmarks, yet uses even less power than the TDP would suggest.

Whether Kaveri shares these same traits is unknown, but I think it's safe to say that Llano's throttling and poor turbo issues will not be coming back to haunt future AMD products, Trinity is already a fair testament to that, Richland is looking to be an even better one.

 

Anand released some gaming power numbers for the 2020M - http://www.anandtech.com/show/6981/...ality-of-mainstream-pcs-with-its-latest-apu/3

That's near HD 2500 level performance I believe, and I would suggest very lightly threaded games as well judging by how close the E-350 is to both. Power draw is double the A4-5000's.

 

It's a 35w chip, what else should we expect? The fact that we're talking about power use for the entire system makes it a little more impressive, but results for the chips alone should show Kabini using less than half the power of a Pentium 2020M. This shows generally just over half. The results just prove the stated TDP's are quite accurate.

 

The stated TDP's appear to be accurate in this case yes.

If we accept these results, how much of a difference would there have to be between this Pentium silicon and the i7 ULV silicon in order for the ULV to be 17W and beating the A4-5000 by as much as it did in the tomshardware review?





Is it likely that the ULV i7 is 50% faster than the A4-5000 while having the same power draw during gaming? How can HD4000 with even higher clocks than this 2020M's HD graphics - and two extra CPU threads - manage that while staying under 17W? Half the power for 50% more performance and a 4-thread cpu on top?

For me that's impossible, and the power numbers on Toms are more realistic. Tom's power delta for the A4-5000 vs this Pentium are similar to Anand's.

 

Look at the difference in performance on the techreport's review -

http://techreport.com/review/24856/amd-a4-5000-kabini-apu-reviewed/7





The Zenbook is nearly 2.5x faster and people think it's because of the single channel RAM on the X202E. While that is undoubtedly a factor I just can't believe it's having such an effect at low settings 720p. What's more the Sandra bandwidth test still shows it as being faster than the A4-5000 which is up to twice as fast in these benchmarks.



Then at the end Scott does a full set of battery tests sans gaming...