AMD - Piledriver chitchat

[watercooled]

I think it was Semiaccurate claiming AMD have had some form of stacked memory ready to go for a while now. Presumably it was just a case of waiting for something like price or availability to improve if so.

IIRC AMD were one of the developers behind HBM memory, so it makes complete sense that if Nvidia are planning to use it, AMD will be too.

Edit: I'm not sure this was the article I had in mind, but it shows what is supposedly an AMD test chip from 2011 with some form of die stacking: http://semiaccurate.com/2011/10/27/a...-gpu-pictured/

[scaryjim]

Well, the die-stacking slide deck is actually targeted at SoC development, *not* GPU development, if you read the full deck. I'm not saying that we won't see stacked-die HBM on GPUs as the first implementation of that, but it's certainly not where the slide deck was focussed - it was looking at die stacking as the future of integration (i.e. instead of everything being integrated in an SoC, you'll stack multiple IP blocks on top of an interpose to create 3D SoCs). If AMD could stack 4 CPU cores, some GPU CUs, a memory crossbar, and 4GB of HBM on a single interposer, it'd make a pretty compelling mobile SoC, after all And it ties in very nicely with their aim of increasing revenue from the semi-custom side of the business.

Given how much PCB space is currently dedicated to memory chips and traces, moving towards some form of stacked high bandwidth memory for GPUs would kind of make sense, particularly for low and mid range graphics - a single 1GB HBM cube @ 1.2GT/s would have around twice the bandwidth of the 128bit 4.5GT/s GDDR5 found on a 7750/7770. So for an R7 grade card, 1GB or 2GB of HBM would have more than sufficient bandwidth.

[CAT-THE-FIFTH]

AMD sets new benchmark record with their new server:

http://www.zdnet.com/amd-seamicro-sm...ds-7000029395/

[DanceswithUnix]

Charlie on AMD's ARM developments: http://semiaccurate.com/2014/05/14/a...-x86-strategy/

What I want to know is how much of the K12 chip will be shared with Jaguar architecturally.

[DanceswithUnix]

Phoronix is a strange site. They benchmark things in a sort of directionless mess, but if you dig around some interesting ideas pop out.

Am1 benchmarks: http://www.phoronix.com/scan.php?pag...c_jaguar&num=4
Tegra K1 benchmarks: http://www.phoronix.com/scan.php?pag...cc48_tk1&num=4

Looking at results like MP3 encoding and Apache pages served per second, they are closer than I was expecting.

These are Cortex A15 and Jaguar cores all running just over 2GHz. The AMD box has 3GB of ram vs the 2GB on the Nvidia board.

Nvidia need that on an ITX board at comparable price to AM1.

[CAT-THE-FIFTH]

Phoronix is a strange site. They benchmark things in a sort of directionless mess, but if you dig around some interesting ideas pop out.

Am1 benchmarks: http://www.phoronix.com/scan.php?pag...c_jaguar&num=4
Tegra K1 benchmarks: http://www.phoronix.com/scan.php?pag...cc48_tk1&num=4

Looking at results like MP3 encoding and Apache pages served per second, they are closer than I was expecting.

These are Cortex A15 and Jaguar cores all running just over 2GHz. The AMD box has 3GB of ram vs the 2GB on the Nvidia board.

Nvidia need that on an ITX board at comparable price to AM1.

The K1 is not exactly that low power(the SOC needs active cooling) and also has a 15% clockspeed advantage too,and by the time K1 actually enters production,Mullins and Beema will enter the fray.

Plus AMD has its server orientated ARM based SOC already,with 8 64 bit ARM cores,which will be faster.

[DanceswithUnix]

The K1 is not exactly that low power(the SOC needs active cooling) and also has a 15% clockspeed advantage too,and by the time K1 actually enters production,Mullins and Beema will enter the fray.

Plus AMD has its server orientated ARM based SOC already,with 8 64 bit ARM cores,which will be faster.

Clock speed means nothing across architecture, proper system power consumption tests would be valid though but probably not forthcoming. ISTR the A5350 has a fan on top too, so not too different there.

Some of the SciMark tests early on in those articles the AMD part takes a sound thrashing, which again I wouldn't expect from just an A15. I guess the ram on the Nvidia system is better suited to those tasks? Investigation would be nice, but I wouldn't expect anything that interesting from Phoronix

But yes, 64bit chips are on the horizon and should leave the A15 for dead. I thought those chips were going to be OK, but now I wonder just how fast they really are. The AM1 parts are already working in 64bit, so they have nothing more to give.

Didn't realise the Nvidia board is make by Zotac: http://www.maplin.co.uk/p/zotac-jets...oper-kit-a30ny

Way too expensive though. OK, I change my mind. AMD, get an A57 SOC out there in an AM1 socket. Quad core is fine, I don't need the full 8 cores. Ta.

[CAT-THE-FIFTH]

It depends on how well optimised those applications are for thing like X86 or the ARM instruction set. I don't know TBH.

BTW,the first UE4 game is benchmarked:

http://gamegpu.ru/images/remote/http...est-d_proz.jpg



http://gamegpu.ru/images/remote/http...st-d_intel.jpg



http://gamegpu.ru/images/remote/http...test-d_amd.jpg



The game seems to look crap and is overly demanding on GPUs too:

http://www.dsogaming.com/pc-performa...ance-analysis/

[DanceswithUnix]

It depends on how well optimised those applications are for thing like X86 or the ARM instruction set. I don't know TBH.

I'm looking for home server tasks, under Linux, so instruction set agnostic. Hardest thing a server would do at my house is run Minecraft which my current FM1 box does with 1 core pretty much always at lowest 800MHz speed it can do.

Looks like that Nvidia board draws 12W for the entire board under full load, not bad for the people that made my phone CPU that kept my hands nice and warm all winter

[CAT-THE-FIFTH]

Extensive CPU test of Thief under DX11 and Mantle done by a forum member over on Anandtech forums:

http://forums.anandtech.com/showpost...&postcount=136

Edit!!

This might indicate a new AMD chipset in 2015:

http://www.kitguru.net/components/ha...2015-chipsets/

[scaryjim]

Extensive CPU test of Thief under DX11 and Mantle done by a forum member over on Anandtech forums:

Interesting to see the big jumps in minimum frame rates there, rather than in averages. I'd be interested to see the per frame timings and variance, as the suggestion from those numbers is that you get smoother, rather than faster, performance.

This might indicate a new AMD chipset in 2015

tbh I'd be surprised if we didn't get new chipset next year. A88X was actually announced in August 2013, and released in Feb this year, so come 2015 it's going to be at least a year old. The real question is whether we're going to get a new FX CPU platform or if it's just going to be a refresh of the APUs. SATA Express sounds very much like an FCH/Southbridge addition, so it could be as uninspiring as a new FCH but remaining on FM2+. Alternatively, it could be that 2015 will be when AMD reunify the performance and mainstream platforms. There's no good reason they can't drop 8+ core CPUs with lots of L3 cache (plus on-die MC and PCIe) into the same socket as 4-core APUs with 512+ shader GPUs. Silicon space shouldn't be too dissimilar, after all, and hopefully within a year we'll have mainstream DDR4.

Actually, there's a retro-loving part of me that really hopes DDR4 will bring back low-cost ATX motherboards with 3 RAM slots, since the point to point nature of DDR4 would mean there'd be genuine benefit to have 3 slots over 2...

[CAT-THE-FIFTH]

AMD is releasing something on June the 4th:

https://twitter.com/killyourfm/statu...00615726678017

Oh ****, AMD is dropping bombs on this press call. (Embargo drops on June 4th but I'm pretty excited about what I'm seeing)

Probably not as exciting for gaming enthusiasts.

[DanceswithUnix]

AMD is releasing something on June the 4th:

https://twitter.com/killyourfm/statu...00615726678017

Another laptop or tablet spec chip perhaps?

[DanceswithUnix]

Supposedly a leak of AMD laptop chip specs:

http://www.anandtech.com/show/8053/amd-kaveri-specs

19W is just 2W more than my E350 netbook CPU. An A10-7300 looks like a very nice upgrade indeed!

[scaryjim]

We had this a few weeks ago didn't we? I definitely remember commenting on how the A10-7300 is essentially the same spec as the A10-4600M in the laptop I'm currently typing on.

And I've just given away my desktop because this laptop provides all the gaming performance I currently need...

EDIT: yeah, just about 20 comments back: http://forums.hexus.net/cpus/241925-...ml#post3271941

[DanceswithUnix]

SA say that RockChip will be shipping an Atom based SOC, and sounds like it will have non Intel graphics.

http://semiaccurate.com/2014/05/27/i...inese-tablets/