AMD - Piledriver chitchat

[DanceswithUnix]

IMO this sort of testing makes more sense in a scenario where you expect most candidates to pass (high yielding) - if you end up struggling to meet demand because of low yield, it may make more sense to run everything through this test in order to capture as many high-performing chips as possible. Then again, I suppose there could be some level of initial candidate selection/filtering based on performance of previous tests i.e. predicting which are more likely to pass based on, for example, voltages needed to reach the 'standard' clocks.

Yes, it does imply to me that they have plenty of 6800K chips to be getting on with. The testing does sound very targeted too, I think the way the statistics goes is that if the first few on a wafer are a bit slow then they probably all are on that wafer as if an alignment is out then a stepper will reproduce that all the way across.

The one thing I would love to know, is whether the 100W parts are intentionally made as higher leakage fast parts or if they just turn out that way.

[kalniel]

There's usually a table that you get which shows voltages required for each speed bin for a particular chip. Each chip will require a slightly different voltage setting, so you do get chip to chip variation and the voltage for the rated speed is encoded in the chip for BIOSes to read I think.

If you get a chip that requires less voltage then you either bin it as a low wattage part, or can increase the clockspeed further. As yields improve then you can more easily supply these binned parts, but might still release these 'good' chips under normal frequencies for market reasons.

[watercooled]

I don't think there's really a 35W difference between the 65W and 100W parts in reality; AFAIK everything from top-bottom, whether desktop or mobile, are made from the same die and chosen by binning. I'm not certain of course. However, the 100W parts make up a fairly part of their line-up; would volumes be high enough to justify a separate process?

[watercooled]

Sorry for completely changing the subject, but I've just spotted something interesting about the desktop dual-core Haswell (i.e. i3) CPUs - they appear to be using the 177mm2 GT2 quad-core die. This is different from SNB/IVB where they released a dual-core die for i3, Pentium, Celeron. Well, Haswell does still have a dual-core die at 181mm2 but it includes the larger GT3 GPU, but it doesn't seem to be used in any socketed desktop parts.

Anandtech posted the die variants used here, on Haswell's release: http://www.anandtech.com/show/7003/t...4560k-tested/5

I wonder if this has anything to do with Intel supposedly using the new Atom cores for Celeron and Pentium, and seemingly ditching the Atom brand in this area - that would likely drastically reduce the volume of the small GPU, dual core die, so now Haswell has done away with that die and is down to 3 variants - GT3 4C, GT2 4C, and GT3 2C, with what would've been the GT2 2C segment being filled by GT2 4C on desktop, and it seems the GT3 2C part in some of the lower TDP parts. (See Wikipedia).

Edit: Without meaning to read too much into this, as my above theory could well cover the decision, but I wonder if the following could be another reason? The previous strategy would mean any dies with only partially flawed quad CPU or high-end GPU would've been scrapped since the lower tier SKUs already had their dedicated dies - now they can 'die harvest' some of those for the 'virtual GT2 2C' parts.

I'm not sure how the costs would work out, but if yields were sufficiently high, a load of fully-working, larger (therefore more expensive to produce) dies would have to be sold as cheaper SKUs. I guess there would be a point, in terms of yield and volume, where it would just be cheaper to do as Intel have done previously and have a smaller, cheaper die for these SKUs? But then that leads back into my previous hypothesis - maybe the lower volume without Pentium/Celeron crosses that cost line?

[CAT-THE-FIFTH]

Sorry for completely changing the subject, but I've just spotted something interesting about the desktop dual-core Haswell (i.e. i3) CPUs - they appear to be using the 177mm2 GT2 quad-core die. This is different from SNB/IVB where they released a dual-core die for i3, Pentium, Celeron. Well, Haswell does still have a dual-core die at 181mm2 but it includes the larger GT3 GPU, but it doesn't seem to be used in any socketed desktop parts.

Anandtech posted the die variants used here, on Haswell's release: http://www.anandtech.com/show/7003/t...4560k-tested/5

I wonder if this has anything to do with Intel supposedly using the new Atom cores for Celeron and Pentium, and seemingly ditching the Atom brand in this area - that would likely drastically reduce the volume of the small GPU, dual core die, so now Haswell has done away with that die and is down to 3 variants - GT3 4C, GT2 4C, and GT3 2C, with what would've been the GT2 2C segment being filled by GT2 4C on desktop, and it seems the GT3 2C part in some of the lower TDP parts. (See Wikipedia).

Edit: Without meaning to read too much into this, as my above theory could well cover the decision, but I wonder if the following could be another reason? The previous strategy would mean any dies with only partially flawed quad CPU or high-end GPU would've been scrapped since the lower tier SKUs already had their dedicated dies - now they can 'die harvest' some of those for the 'virtual GT2 2C' parts.

I'm not sure how the costs would work out, but if yields were sufficiently high, a load of fully-working, larger (therefore more expensive to produce) dies would have to be sold as cheaper SKUs. I guess there would be a point, in terms of yield and volume, where it would just be cheaper to do as Intel have done previously and have a smaller, cheaper die for these SKUs? But then that leads back into my previous hypothesis - maybe the lower volume without Pentium/Celeron crosses that cost line?

It could be down to yields - there was noise about Intel having some issues with 22NM. There are probably enough failed quad core parts,which makes them viable. OTH,it does mean the gap between AMD and Intel sub £100 mainstream CPUs has narrowed somewhat(177MM2 against around the 240MM2 mark).

Edit!!

CPU-Z pictures for Kaveri ES samples leaked:

https://scontent-b-dfw.xx.fbcdn.net/...89011012_n.jpg
http://cdn.overclock.net/9/99/9933b8...1656005_n.jpeg

[AETAaAS]

So, I think it is a down clocked 6800K with less testing. Feeling lucky?

Ooo... sounds so tempting, an extremely competent chip with a healthy discount... but but... I'm eagerly waiting for Kaveri, and if CAT's pics show CPUZ reading it right (I don't know how CPUZ draws data), AMD finally went down to 28nm. ARRRGGHHH. No. No. I promised myself I'd wait for an FM2+ mini ITX board first...

[DanceswithUnix]

No. No. I promised myself I'd wait for an FM2+ mini ITX board first...

You mean like this one? http://www.scan.co.uk/products/gigab...-dvi-hdmi-mitx

But yeah, this late in the game I would wait for Kaveri

[Terbinator]

So not quite the 'PS4 on a chip' Shame, solid start for a cheap Steam/MOBA box, though.

http://hothardware.com/News/AMD-Anno...-512Core-Part/

http://www.pcworld.com/article/20624...n-january.html

Edit: Could certainly get it closer to an Xbone though with 2133+MHz RAM.

[CAT-THE-FIFTH]

So not quite the 'PS4 on a chip' Shame, solid start for a cheap Steam/MOBA box, though.

http://hothardware.com/News/AMD-Anno...-512Core-Part/

Edit: Could certainly get it closer to an Xbone though with 2133+MHz RAM.

It seems to run BF4 much better!

[Terbinator]

That's true enough. Obviously the SR(!) cores are potent and not TDP bound like the Jag cores in the consoles, so i'm curious to how so much ground is made up with a decidedly 'weaker' GPU @ only 512 cores/800MHz.

[CAT-THE-FIFTH]

This is a review of the HD7750 GDDR3:

http://ht4u.net/reviews/2012/msi_rad...st/index36.php

The core and RAM runs at 800MHZ.

Trinity with 2133MHZ RAM is quite close to an HD6670 GDDR3 on average,and with 1600MHZ GDDR3 RAM is closer to an HD6570 GDDR3.

IIRC,the HD6670 and the HD6570 GDDR3 cards HT4U used,have slightly faster 900MHZ GDDR3 IIRC. Hence with 1600MHZ GDDR3 you are looking at least a 30% speedup over Trinity by my estimates,perhaps higher with faster RAM and in newer games. With the suspected improvements in the memory controller and memory handling,it probably would be more overall IMHO.

Edit!!

The part tested was the A10-7850K:

http://www.pcworld.com/article/20624...n-january.html
http://www.anandtech.com/show/7507/a...p-january-14th

It has a 3.7GHZ clockspeed and the GPU runs at 720MHZ.

Second Edit!!

They might be using a newer version of GCN which the HD7750 does not.

Third Edit!!

It seems AMD demoed the A10 7850K running the BF4 SP campaign at 1920X1080 on medium settings. It was running at 28FPS to 40FPS whereas the Core i7 4770K using an HD4600 IGP could only run the same sequence at 12FPS to 14FPS.

[CAT-THE-FIFTH]

New AMD servers from HP:

http://www.techweekeurope.co.uk/news...oonshot-131552

TressFX 2.0 detailed:

http://www.polygon.com/2013/11/11/50...-grass-and-fur

Grass and fur animations too!!

[CAT-THE-FIFTH]

Kaveri has a smaller die size than previous APUs:

http://img.new.livestream.com/events...46bed391ce.png



Kaveri against a Core i7 4770K with a GT630:

https://www.youtube.com/watch?v=HjAM2zYNqko

[scaryjim]

Confused as to why they've decided to go for 512 shaders when they don't have a 512 shader part in the current generation of discreet GPUs. That's almost like saying they expect people to crossfire (well, dual graphics) it with a previous gen card. Given engineering samples have been out for a long time now they can't have decided this last minute to counter GT3e, but then why not make sure you have a 512 shader discrete part to couple with it?

So yeah, very confused

[CAT-THE-FIFTH]

They might release some more cards. See the "new" R9 270 which is being released just under the R9 290X. They could release a R7 260 non-X or an R7 250X as a die salvaged Bonaire based part with 512 shaders,when Kaveri launches.

If not I think AMD should drop the Hybrid Crossfire feature if they cannot get it to work properly like in the past.

On a side not seems Anandtech forums is fully hissy fit mod and the mods are infracting anyone who is not an Intel fanboi.

[Terbinator]

Confused as to why they've decided to go for 512 shaders when they don't have a 512 shader part in the current generation of discreet GPUs. That's almost like saying they expect people to crossfire (well, dual graphics) it with a previous gen card. Given engineering samples have been out for a long time now they can't have decided this last minute to counter GT3e, but then why not make sure you have a 512 shader discrete part to couple with it?

So yeah, very confused

This could literally be a new part as it's GCN1.1 and the only other 1.1 part is the 7790 AFAIK.