AMD - Piledriver chitchat

[CAT-THE-FIFTH]

http://translate.googleusercontent.c...jWIRNWFXFl62Zw

It seems there is the possibility of Kaveri using GDDR5M SODIMMs.

[kompukare]

Interesting but I can't help but think that cost and availability wise this might be almost as bad as RDRAM (well aside from potentially more manufacturer being able to provide them if the take off I guess).

Still a hybrid design (4GB soldered plus 1-2 SODIMM slots) for a laptop might make sense. Should perform better than Intel Iris and if the Kaveri core manages to close the IPC gap it could be good choice.

Still think AMD's driver team should invest big time in getting Enduro sorted out though: they've lost so much mobile marketshare to Nvidia it's crazy. Despite the perception (mostly caused by Tahiti), GCN offers similar performance/watt to Kepler: certainly on the chart which TPU's W1zzard produces only GT650/TI is ahead of Pitcairn.

[kalniel]

This is following on from the console work I presume?

[scaryjim]

The roadmap does imply a new 2P/4P platform Opteron in early 2014

... looking at the server roadmap it looks like they're pushing the Berlin APU platform into the single-socket market ...

This. According to the roadmap the high end server market will be serviced by many-core Piledriver designs, while the low-end server market will be shunted over to a server-qualified APU based on Kaveri. There's nothing on the roadmap between 4 and 12 cores, which seems like a suspicious gap to me - perhaps they're waiting to see how well Kaveri performs before they decide whether to make CPU-only versions? Given the alledged modular nature of the current design, if Steamroller plays well with consumer workloads they might be able to add an 8-core FM2+ part without GPU fairly quickly and consolidate back onto one socket in the consumer space. But I'm not holding my breath, and there's nothing on the roadmap to suggest that CPU-only steamroller parts with more than 4 cores are in the offing...

[CAT-THE-FIFTH]

This. According to the roadmap the high end server market will be serviced by many-core Piledriver designs, while the low-end server market will be shunted over to a server-qualified APU based on Kaveri. There's nothing on the roadmap between 4 and 12 cores, which seems like a suspicious gap to me - perhaps they're waiting to see how well Kaveri performs before they decide whether to make CPU-only versions? Given the alledged modular nature of the current design, if Steamroller plays well with consumer workloads they might be able to add an 8-core FM2+ part without GPU fairly quickly and consolidate back onto one socket in the consumer space. But I'm not holding my breath, and there's nothing on the roadmap to suggest that CPU-only steamroller parts with more than 4 cores are in the offing...

If you look at some of the earlier AMD documents there was noise about a 3 module Kaveri design. It could be that later AMD APUs move up to 3 modules/6 threads when they move onto 20NM. Even with 3 modules and a later core,I still suspect the GPU will be a large percentage of the die.

[CAT-THE-FIFTH]

A6 5200 reviewed:

http://translate.google.co.uk/transl...00-kabini.html

Interesting results with a GTX680!

[watercooled]

Shows how single-thread bottlenecked modern games are.

[CAT-THE-FIFTH]

Shows how single-thread bottlenecked modern games are.

BTW,check out the nerd rage on a certain forum since AMD has increased market share this quarter.

[scaryjim]

hmmm, I find the integrated graphics performance interesting, tbh. This is an APU with half the CPU performance of A10-6800k, one third of the shaders @ < three-quarters of the clock speed, single channel memory, so only half the memory bandwidth, and it performs at about 40% of the A10 in graphically intensive workloads? I think we can take from that that either GCN is very efficient compared to VLIW4, or that Richland is horribly crippled by memory bandwidth (in which case Kaveri is going to *need* GDDR5, IMNSHO). I suspect a bit of both, tbh

EDIT to clarify: the A10-6800k should have at least 4x the graphical grunt of the A6-5200, ignoring architecture improvements. Add that to the much higher CPU performance, and it should be a lot more than 2.5x as fast in games. To be as close as it is bandwidth has to be a significant factor. Come on AMD, sort it out - your aging dual channel DDR3 memory is cripling your hardware!!!

[CAT-THE-FIFTH]

hmmm, I find the integrated graphics performance interesting, tbh. This is an APU with half the CPU performance of A10-6800k, one third of the shaders @ < three-quarters of the clock speed, single channel memory, so only half the memory bandwidth, and it performs at about 40% of the A10 in graphically intensive workloads? I think we can take from that that either GCN is very efficient compared to VLIW4, or that Richland is horribly crippled by memory bandwidth (in which case Kaveri is going to *need* GDDR5, IMNSHO). I suspect a bit of both, tbh

EDIT to clarify: the A10-6800k should have at least 4x the graphical grunt of the A6-5200, ignoring architecture improvements. Add that to the much higher CPU performance, and it should be a lot more than 2.5x as fast in games. To be as close as it is bandwidth has to be a significant factor. Come on AMD, sort it out - your aging dual channel DDR3 memory is cripling your hardware!!!

AFAIK,the Intel memory controllers seem to be more efficient,so I wonder if Kaveri will have a reworked memory controller??

[scaryjim]

I've always found it fascinating that Intel have such a huge lead in synthetic benchmark tests - I was never sure whether it was a genuine limit to the bandwidth available to the AMD memory controller or just some clever caching and stuff that gave Intel such a huge lead. Looking back at Hexus reviews (using 2133MT/s DDR3) AIDA shows Intel hitting over 20GB/s available bandwidth with AMD barely scraping half that. If that's a true picture rather than an artifact of testing, the bandwidth available to the APU graphics is more like a 64bit card's, rather than a 128bit. A more efficient memory controller could make a huge difference if that's the case...

[watercooled]

At least the read bandwidth shown by Hexus seems to be on the low side vs other tests, e.g. http://www.hitechlegion.com/reviews/...owall=&start=3

[scaryjim]

Interesting. Sandra in that link shows the same results that Hexus get with AIDA64 (AMD in low teens, Intel around 20s), while their AIDA64 results put AMD right up there. Very odd. I think the whole AMD memory bandwidth thing could keep running. Until they sort out GDDR5 on Kaveri, of course

[CAT-THE-FIFTH]

AMD graphics shipments increased:

http://www.techpowerup.com/188980/am...-research.html

[watercooled]

Back on the subject of VCE, I wonder if there are any screen capture/recording solutions using hardware encode? The next gen consoles appear to be doing something along those lines for game recording, presumably using VCE.

Nvidia are using something similar for Shield, and Apple are using the Intel equivalent for Airplay mirroring.

[CAT-THE-FIFTH]

It looks like AMD discrete card shipments also increased:

http://www.digitimes.com/news/a20130816PR205.html

The figures exlude IGPs and laptop cards.