AMD - Piledriver chitchat

[watercooled]

I've just been digging through some news articles and noticed, the 15 core IVB-EX die appears to have the same stats as the existing 12 core EP die - 541mm2 and 4.31B transistors. Although they have different feature sets, I wonder if they're just fusing them off where appropriate? The die shot we were given for the 10C EP (notice, we weren't shown a 12C die ) looks an awful lot like the 15C die minus the extra slice of 5 cores. That 12C 'die' did bug me considering they have the 10C one; would making a separate die for two extra cores be that sensible? At the time it would've made sense to assume the 10C was a fused-down 12C had it not been for the die shot and memory controller differences.

Releasing 12C first helps them to ramp yields I guess (oh I forgot, Intel 22nm yield is so perfect they don't need to bin do they? ). The EP and EX differ significantly in some areas, like extra QPI links, memory controller, etc, so that's a fair amount of unused die either way it's fused. However it may still be cheaper to do that than to have two slightly smaller dies, losing some potential for binning, etc. And those die size/transistor counts, unless incorrect, are identical as I said, which would be extremely unlikely even assuming rounding. Plus the lack of that 12C die shot...

[DanceswithUnix]

Interesting find. Intel were selling a 6 core CPU made from an 8 core die which was supposedly nothing to do with yield and all to do with thermal limits.

Perhaps the same was true here. They could hit decent speeds with 12 cores, but not cope with the full 15. A while down the road, process improvements have moved the goalposts.

There was a figure of about $1M to make up the masks for a new chip on a modern process. That is a lot to claw back, so selling the same chip re-branded over and over is to be expected.

[watercooled]

Yeah that makes sense. Do you mean the extra surface area of the larger die helps thermally i.e. a 12C CPU from a 15C die would be better in that respect than a native 12C? Or besides that, I guess using a 15C die would allow them to potentially pick the best-clocking cores while disabling the others; with a 12C die it would be all or nothing.

My comment in brackets was meant in jest obviously, just mirroring the sensationalism I see dotted around certain other forums. However, even with a very mature process, with a die this large I'd still expect a considerable number of partly defective dies which could be recovered by binning, but otherwise lost if they had no core-disabled parts on the market.

[DanceswithUnix]

Massive die yes, but lots of cache on there which may well be repairable with spare cache lines to keep the yields up.

The thermal bit is basically this: You can run all the cores at low speed, or some of the cores at high speed and generate the same TDP. Different people will have uses that suit one or the other, so one part can be sold as two SKUs.

Anandtech give a map of how the die is fused off for different configurations, and yes knocking out alternating cores does spread the heat out nicely.

As for sensationalism, my first thought on this was that if AMD released a 155W part people would be pointing and laughing.

[DanceswithUnix]

Scan and Ebuyer both seem to be out of stock on the A10-7850K. Hopefully this is an indicator of good sales, not supply issues.

[ik9000]

Massive die yes, but lots of cache on there which may well be repairable with spare cache lines to keep the yields up.

The thermal bit is basically this: You can run all the cores at low speed, or some of the cores at high speed and generate the same TDP. Different people will have uses that suit one or the other, so one part can be sold as two SKUs.

Anandtech give a map of how the die is fused off for different configurations, and yes knocking out alternating cores does spread the heat out nicely.

As for sensationalism, my first thought on this was that if AMD released a 155W part people would be pointing and laughing.

well they released a 125w part then pre over clocked it to 220w and reviews label it a thermo nuclear cooker... and no mobo officially supports it without limitations and caveats...

[HalloweenJack]

well they released a 125w part then pre over clocked it to 220w and reviews label it a thermo nuclear cooker... and no mobo officially supports it without limitations and caveats...

my 9590 on a £35 cooler seems to be working well on a sabertooth 990fx

[CAT-THE-FIFTH]

I wonder what the TDP and power consumption of overclocked FX8320 and FX8350 CPUs would be?? Nice to see the reviewing world not testing something else.

[watercooled]

They're essentially the same CPU, and power draw tends to go up drastically with increasing voltage, so I'd expect it to be in the same ballpark.

http://www.tomshardware.com/reviews/...cy,2850-9.html
Although the writers can't seem to figure out why idle power is similar. I don't know why a reviewer wouldn't know about power gating...

http://www.bit-tech.net/hardware/cpu...idge-review/11

Overclocking really doesn't go hand-in-hand with efficiency...

[CAT-THE-FIFTH]

Going back to the above picture again. It seems those power measurements are at the PCI-E connectors.

It seems to be boosting beyond its TDP spec.

I remember talking just before the GK104 launch the non-deterministic Turbo would cause problems,and I was right to some degree.

[DanceswithUnix]

It seems to be boosting beyond its TDP spec.

As long as it doesn't *average* over TDP spec, then why is that not a good thing? Or more to the point, given that lots of these cards have non standard cooling then they should cope fine as long as they aren't ignoring the power regulation requirements and chip junction temperatures?

Edit to add: I am too lazy to go check THG to see just what sort of card they are testing. Doesn't seem to have thermally throttled during this though, which was one of your concerns.

[watercooled]

I'd like to see more of an investigation into it though, how it compares to other cards, how much of it is caused by boosting, etc.

(Yes I've just edited a big bit out while I check something)

[CAT-THE-FIFTH]

As long as it doesn't *average* over TDP spec, then why is that not a good thing? Or more to the point, given that lots of these cards have non standard cooling then they should cope fine as long as they aren't ignoring the power regulation requirements and chip junction temperatures?

Edit to add: I am too lazy to go check THG to see just what sort of card they are testing. Doesn't seem to have thermally throttled during this though, which was one of your concerns.

Most of the reviews are done on open air test benches or if you are lucky high end ATX cases with loads of cooling fans.

What about smaller cases with less cooling?? Many people are lucky to have more than one fan in their system.

This card is being pitched for use in smaller cases,which are thermally more constrained.

People are more interested in the cards with the smallest coolers,not the larger ones,so they can plonk it in a HTPC or SFF rig. This is what I gather from people's comments on why they want the card on a few forums.

So what happens when a single slot or low profile version is released with a weedier cooler?? If it needs to boost that much past spec,its worrying.

Plus a tiny cooler,in a smaller PC with lots of dust buildup,means cooling performance will be reduced.

I do mostly SFF builds,so this is why I am worried the testing is not good enough.

Also,the tests are lucky to exceed 3 minutes and usually are at most 30 to 60 seconds.

The throttling problems were seen over longer time periods. What about 15 to 30 minutes in a small case??

If it is boosting so highly,then it means that it will need better cooling than a card like an HD7750. They do not boost and consume less power.

The reference GTX760 throttled by 140MHZ in 15 minutes in a normal case IIRC. Geforce Titan with the best reference cooler ever made,throttled over longer periods.

There were no reference GTX770 cards to test,so it means the two Nvidia Boost 2.0 cards tested with worse cooling had issues,due to its aggressiveness. Even though AMD Boost is not the same,we saw what happened to the R9 290 under suboptimal cooling conditions with the reference cooler. The previous version hardly boosted,so really was of no use for scores.

The boost is not small with newer cards - it adds a decent chunk to performance.

[CAT-THE-FIFTH]

Another thing:

http://en.wikipedia.org/wiki/GeForce_600_Series
http://en.wikipedia.org/wiki/GeForce_700_Series

Look at the lists carefully.

The mobile versions of the Nvidia GPUs have no boost.

That includes the GK106 too.

That indicates that reasonable cooling is required for boosting and Nvidia knows this.

[HalloweenJack]

if its boosting over TDP spec then how much power over the slot limit is it going? danger to vrm`s?

[DanceswithUnix]

The throttling problems were seen over longer time periods. What about 15 to 30 minutes in a small case??

Perhaps they should do some hot box testing like you get with decent PSU reviews. That would get you ideal results from open frame tests (I want those, in my personal main PC I have tons of airflow) but also give you an idea of tolerance to use in a harsher environment.

The mobile part thing is interesting. Laptops have OK cooling but almost no thermal inertia as people get stressy about having their lap/naughty bits cooked. I imagine without a decent lump of heatsink to dump energy into boosting isn't really going to work.