AMD - Piledriver chitchat

[CAT-THE-FIFTH]

Saw this on Anandtech forums(thanks Abwx):

http://translate.googleusercontent.c...aKFDiz-VUIBnBA

Tags: AMD , these 2014 ; Elpida ; GDDR5M ; JEDEC ; SK Hynix ;
Published on 11/01/2014 04:09 by Damien Triolet
Envoyer Imprimer
Share on facebookShare on twitter

This summer, we told you about the existence of a type of memory that had been more than discreet and who presented himself as particularly well suited to power APU: the GDDR5M . It should ultimately never see the day.

As a reminder, GDDR5M is a derivative of GDDR5 memory, the latter being primarily intended for GPU (but also operated on the PS4). If the basic technology is the same, the GDDR5M is optimized to allow a format "memory bar" without too explode costs, including through 4Gb chips interfaced 16-bit, 32-bit cons for classical GDDR5 . Enough to allow the establishment of memory modules SO-DIMM 2GB and 4GB


During CES, we had the opportunity to talk with Joe Macri which carries the dual role of Chairman of the committee in charge JEDEC DRAM and Chief Technology Officer in charge of public platforms for AMD. Of course we wanted to take the opportunity to find out what had happened to this GDDR5M.

Details of future A Series APU is still under embargo, and the eventual abandonment of the support of a technology originally intended which are unlikely to be formally discussed, we have of course not attempt to delve deeper by combining Kaveri the GDDR5M. We believe more than likely that AMD had planned to use for Kaveri but this remains speculation on our part.

What is not speculation by cons is that GDDR5M is well and truly dead and buried. Joe Macri has explained that he was a "wonderful memory" which presented extremely well at the start. Very promising, it would allow to save time compared to the arrival of the memory Wide I/O2 HBM or HMC, and two industrialists had decided to develop and produce.

This is the central point of the destiny of this GDDR5M as in the DRAM industry, it is necessary that at least two sources are confirmed for a technology can evolve into a commercial existence. Otherwise the risks are too high in availability or pricing. This is what killed in the bud development GDDR5M.

The two industries that were examined were Hynix and Elpida above. However, there are almost two years, Elpida had announced its bankruptcy and placed under protection by the authorities of Japanese financial regulation. They have ended all new projects which GDDR5M. Remained then was Hynix, and unfortunately no other DRAM manufacturer, Micron, for example, was willing to take over and start the adventure.

If our speculations are correct, you easily imagine that bankruptcy Elpida and could have important implications for AMD and Kaveri.

That sucks.

[watercooled]

The cooler mounting system on the Kabini boards looks unusual, two holes rather than four. http://www.techspot.com/news/55296-m...ming-soon.html

Also, I wonder if it will be bare die or come with a heatspreader?

[CAT-THE-FIFTH]

This chap has got an A10 7850K and is comparing it against an A10 5800K:

http://translate.google.com/translat...p%3Ft%3D421821

Looking at scores,I estimate the A10 7850K to match the A10 6800K if using a discrete card,although in some games the A10 7850K is much faster it seems.

An A10 7700 SSD test:

http://forums.hardwarezone.com.sg/st...l-4536563.html

[kompukare]

Well those A10-7850K tests seem to show a very big gain in cache write and copy speeds but how that will translate into real life gains is hard to say. Cinebench R15 and Fritz Chess are hardly meaningful benches so until the NDA lifts it is hard to say. Pity AMD are so resource starved that they can't afford to make a 4 module version even it would only be on the market for a few months.

If they really went to all the trouble of including a GDDR5M controller they must be a bit upset that Elpida went bankrupt although I guess the PS4 GDDR5 controller might have been very similar so it might not have as much wasted effort as it at first appears to have been.

[watercooled]

Fritz Chess is an integer workload - here it looks like Kaveri is matching overclocked Trinity.

[kompukare]

Yes the only problem there is that its vs Trinity not Richland. A quick google shows something like this:
http://www.ocfreaks.com/forums/f18/f...richland-2799/

So that gives:
Kaveri A10-7850K 7676
Richland A10-6800K 7299
Trinity A10-5800K 6808 / 6629
Llano A8-3870K 7394
So, that's 5% faster than Richland (or finally faster than Llano's Stars architecture). So not that good. GF's 28nm probably takes a lot of the blame.

Still think the brute-force approach of using so much die for 512 GCN cores is not that great. While Intel's brute-force approach used in Iris Pro seems very wasteful (transistor & die space compared to GCN or Kepler looks poor but the Intel-is-God crowd on certain forums never want to acknowledge that), it's obvious that without some new ideas, APUs are massively bandwidth starved (hence why they wanted to go for GDDR5M if that is true).

[watercooled]

Richland is essentially tweaked Trinity with a 300MHz clock bump in this case.

You're comparing results from different set-ups; was different memory used, another OS, a different version of the program?

You can see yourself that the 5800k is scoring lower on the benchmark CAT posted, by ~200 points. We only have one reference point so it's far from ideal, but add a similar percentage to Kaveri and we have something like:

Kaveri 7900
Richland 7300

Edit: Or using Trinity as a reference point, we can work out relative performance with the information given in the two articles:
Trinity = 100%
Richland = 107%
Kaveri = 116%

Also bearing in mind that the 7850k has a lower clock speed than either the 5800k or 6800k. It also fits in with the 20% IPC claim.

[CAT-THE-FIFTH]

The compute benchmarks are up and the performance gains are huge.

Edit!!

I suspect that the HD4600(and probably even the HD5200) will be destroyed in overall compute performance by the Kaveri IGP:

http://www.tomshardware.com/reviews/...ew,3521-3.html
http://www.anandtech.com/show/6993/i...50hq-tested/17

I expect we will soon hear the cries,of "Broadwell will be 10X faster" on some forums.

[ik9000]

The compute benchmarks are up

where? the links don't seem to include the A10-7850 etc and dated June 2013? Or have I missed something/link got edited out etc?

[watercooled]

From post 2355: http://foro.noticias3d.com/vbulletin...d.php?t=421821 (it keeps getting updated).

[Alistair]

7850K in stock at Dabs - they had 50. Now its 49 Also found the 7700K on there too

Getting it & an Asrock FM2A88X-ITX tomorrow

[CAT-THE-FIFTH]

Looks like the Intel propanganda is in full swing on a certain forum we all love.

7850K in stock at Dabs - they had 50. Now its 49 Also found the 7700K on there too

Getting it & an Asrock FM2A88X-ITX tomorrow

It seems to the price it was rumoured to be!!

However,for an ITX system it does sound like a nice setup.

[Alistair]

Its for my HTPC. I have been running an i5 setup with an additional graphics card but am keen to try this. I already have 2x8GB Patriot Viper 3 Low Profile 2133 MHz. Initially going to put it in my Silverstone GD05 but if its up to the task. Will look to downsize to either their ML series or perhaps a Streacom case with external PSU. No gaming use, purely for media playback (madvr) & general use.

[CAT-THE-FIFTH]

Its for my HTPC. I have been running an i5 setup with an additional graphics card but am keen to try this. I already have 2x8GB Patriot Viper 3 Low Profile 2133 MHz. Initially going to put it in my Silverstone GD05 but if its up to the task. Will look to downsize to either their ML series or perhaps a Streacom case with external PSU. No gaming use, purely for media playback (madvr) & general use.

A little user review would be nice!!

[Alistair]

A little user review would be nice!!

Will do

[Terbinator]

The Anandtech thread is painful.