AMD - Piledriver chitchat

[watercooled]

Interesting. I guess the big die helps with heat dissipation if the GPU part is disabled, but I imaging 4GHz would be some way off the end of the efficiency curve.

[watercooled]

Just finishing watching the Skylake TR podcast. At around the 1h21m mark they're talking about graphics architectures, some interesting points in there. Later on, about 1h24m, he mentions that some of the power efficiency of Maxwell may be down to stripping out a lot of the scheduling logic from Kepler, again fitting in with what a lot of people have said about it being more of a purely graphics-focussed architecture at the expense of compute performance. But of course compute has a fairly important role in some aspects of DX12...

[CAT-THE-FIFTH]

Take with a bucket of salt but supposedly early production Zen samples are around Haswell performance:

http://seekingalpha.com/user/19326301/comments

[watercooled]

Rumour about AMD using TSMC for Zen: http://wccftech.com/amd-contracts-ts...obalfoundries/

Personally I'd need a lot more info before considering it viable though, a big reason being the WSA unless they plan to fill that with GPUs or something, which seems even less likely. Still, stranger things have happened.

[DanceswithUnix]

Carrizo products? http://www.anandtech.com/show/9674/h...o-aseries-apus

[DanceswithUnix]

So you can get Carizzo with ECC memory with optional DDR4 support, but only for digital signage use.

Ever get the feeling AMD is trying to beat customers away with a pooh covered stick?

http://www.anandtech.com/show/9726/a...h-ddr4-support

[scaryjim]

... Ever get the feeling AMD is trying to beat customers away with a pooh covered stick? ...

No, but I do get the feeling that they're seeing themselves more and more as a server/industrial/custom chip manufacturer before a consumer one. I guess at the moment they probably think that they don't have enough CPU performance to persuade the enthusiasts away from Intel and DDR4 isn't yet cheap enough for a mainstream consumer platform. iirc Kaveri didn't bring consistent IPC improvements over Richland but was task and instruction-set dependent, and Richland wasn't much of an improvement over Trinity, which is the same Piledriver core that's already in use in the FX -300 series chips?

So even if they took a deep breath and decided to risk going to an entirely new platform across the board, unifying APU and CPU lines, pushing DDR4 for the consumer space; they'd be doing it with an FX CPU series that wasn't significantly faster than their incumbent parts. The only place they could realistically make gains would be through slightly faster memory and a newer version of PCIe, and people would have to invest in an entirely new platform to do it.

Given typical industrial product lifecycles that's not such a big issue; you're probably replacing kit that's already 5 - 10 years old so you're going to be changing platform anyway, and you're going to be willing to invest in technology that will give you the maximum replaceable lifespan; if your signage application's RAM develops a fault in five years will it be easier to source DDR3 or DDR4 at that point?

In the consumer space it's a much harder sell, because people like cheap upgrades, carrying over existing components, and familiar technology. AMD went to both DDR2 and DDR3 later than Intel; with DDR2 they said the delay was down to DDR2 not offering any benefit over DDR. I imagine the issue now is the reverse: moving to DDR4 won't squeeze any more performance from piledriver, or even excavator, cores.

So, AMD's taking the new tech to the people who they think are most likely to want it - and pay for it - first. There's a risk to the strategy, but if they get it right it could put them in a better financial position to do a proper release through consumer products. Only time will tell if that's the case, though.

[DanceswithUnix]

TBH DDR4 support is the least bit of interest to me, though if it let them release AM4 motherboards before Zen comes out it would be nice.

These could make a lovely home server with the ECC support.

I expect they would make a really good Steam box as well.

The E350 came out on small motherboards, I don't even really require a socket. But as the saying goes, the best way to lose a fight is to not turn up.


Edit to add: Sapphire are doing a couple of ITX boards for this, hope they turn up in retail. 35W and HDMI 2.0 could make these of interest to the HTPC crowd.

http://www.sapphiretech.com/page.asp?IDno=5997&lang=eng

[scaryjim]

... if it let them release AM4 motherboards before Zen comes out it would be nice. ...

I think how well Zen tapes out may end up being the deciding factor on that; a good Zen ramp may negate the need for an interim solution, but this does at least demonstrate that they're production-ready with DDR4.

If Zen looks like being delayed further I really hope AMD suck it up and push the new platform out with excavator-based CPUs/APUs. Wanting to grow the semi-custom and embedded side of the business is all well and good, but if they do that at the expense of the consumer side they may find themselves doing a Via....

EDIT: re the sapphire boards: pictures and spec here: http://www.sapphiretech.com/productd...9AEC3&lang=eng

[DanceswithUnix]

http://arstechnica.co.uk/tech-policy...er-core-count/

Is it possible to have class action against all people so monumentally stupid that they cause normal people stress?

I had no respect for people who class action sued Intel for buying the lemon that was P4 without bothering to read the reviews first. But this is just tripe. The chip was faster than Phenom II for the same sort of money (unlike P4 which was slower than P3 for more money) so I can't see how harm was done.

[watercooled]

Yeah I think monumentally stupid about sums that up. I'm not often surprised by the idiocy of certain legal systems any more but really I can't see how this forms much of a case. Bulldozer does have eight integer cores and eight independent integer pipelines. Since there is, to my knowledge, no definition of what a 'core' is, how does this make any sense at all?

What about GPUs claiming to have 'thousands' of cores? In one sense it's not incorrect, but they're not what you'd call cores in CPU terminology.

In other news, provided this isn't marketing fluff it sounds like good news for GloFo and AMD: http://hexus.net/tech/news/industry/...infet-success/

It's not too surprising to hear about LPP production as it fits in with Samsung's timeline but this is GloFo after all. Provided the claims can be taken at face value, even H1 2016 doesn't seem unlikely as a release date TBH.

[watercooled]

Another one of my die 'analyses', I spotted this slide: http://i.web4.hu/apix_collect/1509/m...3_original.png

On the right is what looks like the quad core Skylake die we're familiar with, but to the left is a smaller dual core die.

So with some pixel counting, I've again estimated the die size using the quad core one as a reference. They look to be identical in width (71px) but differ in height (105 vs 80px). So we're looking at 7455 vs 5680 pixels area. The area of the quad core is given as 122.4mm2 by anandtech so that makes the dual core pictured roughly 93.3mm2!

I wonder if this die is used in things like the desktop i3/Pentium/Celeron?

Edit: Anandtech already had the die size listed in their Skylake launch article: http://www.anandtech.com/show/9582/i...cture-analysis

And on the slide titled 'most scalable ever' Intel have listed the dies used in each product, so in answer to my own question it looks like the 2+2 die is used on desktop.

[watercooled]

Something I spotted linked on Twitter: http://reviews.llvm.org/D13979

An Intel product dropping x87? Whatever will we do without SuperPi?

[shaithis]

Luckily it's only a quark SoC...

[CAT-THE-FIFTH]

It looks the Core i3 6100 IGP is still much slower than the IGP in the A10 7870K:

http://www.clubedohardware.com.br/ar...i3-6100/3192/8

[CAT-THE-FIFTH]

I didn't see this one coming:

http://www.anandtech.com/show/9832/s...he-tech-report

Interesting person for AMD to hire!!