AMD - Piledriver chitchat

[scaryjim]

New semi-custom ARM chip in the works, I see - wonder who that's for....

[watercooled]

I don't recall hearing anything about it TBH. You?

[scaryjim]

Nope, new one on me. Could just be a case of a server manufacturer slapping their name on an existing roadmap product - Samsung did that with the A6-1450 in one of their lifebook range (it was marketed as a Samsung quad core CPU, but based on the specs (and Radeon graphics ) was pretty much an A6-1450 - not sure how much customisation Samsung actually did!). Or it could me a mobile application processor with Radeon graphics, although I'm not convinced that GCN will go low enough power to make it worth pairing with an ARM core. Android gaming console maybe?

I think I'm veering towards the former, tbh: ARM server chip seems the most likely.

[DanceswithUnix]

I think I'm veering towards the former, tbh: ARM server chip seems the most likely.

That was my expectation, since they are already doing a server chip and seem to be having a hard time in tablets vs Intel giving chips away.

[HalloweenJack]

AMD on samsungs 14nm process? they`ve just licenced it at glofo

[scaryjim]

AMD on samsungs 14nm process? they`ve just licenced it at glofo

If it was an AMD chip (i.e. being released with an AMD name) it wouldn't be classed under semi-custom. It might essentially be an AMD SKU with a different name on it, but that different name is kind of important in business terms. It could be a genuine semi-custom like the console chips: using AMD IP but in a combination that no-one else is using, but even so the only places it would really make sense (given AMDs complete lack of presence in the mobile segment) would be in an Android (or other linux derivative) console, or in a server environment. And who's going to release a linux console with AMD graphics....

None of which, of course, has any impact on what process it might be on - but frankly that's not what interests me. AMD will inevitably migrate to 20nm and 14/16nm in time. it's what the specs and performance of those future chips is that interests me....

[watercooled]

http://hexus.net/gaming/news/hardwar...gaming-device/

An obvious guess would be the next Wii, but who knows, maybe we'll see another competitor? Or maybe it will be a mobile device?

ARM A15 vs A17* at same clocks, same OS.
http://browser.primatelabs.com/geekbench3/1215304
http://browser.primatelabs.com/geekbench3/1457583

*Although ARM have done away with the A12 designation, as mentioned here, this chip still technically identifies itself as A12. I don't know how much this matters, but it's worth noting IMO.

Anyway, aside from the memory subsystem which could be down to memory chosen for the platform etc, the A17 does impressively well vs the A15! I know it's just the one benchmark etc, but it seems hard to find any comparisons like this.

[CAT-THE-FIFTH]

Interesting test of performance and power consumption of an overclocked FX8350 and FX6300:

http://www.techspot.com/article/932-...r-performance/

[watercooled]

Some more random CPU info, this time comparing ARM A7 and A53 at similar clocks.

Moto G with quad A7: http://browser.primatelabs.com/geekbench3/1560053
Galaxy Ace 4 with quad A53: http://browser.primatelabs.com/geekbench3/1556541

I work out roughly a 41% uplift in per-core performance. This is more or less in line with ARM's charts which seem to show somewhere between 40 and 50 percent over A7 on the same node (they're both 28nm LP).

However it's also worth bearing in mind that the A53 is running in 32 bit mode here and performance may improve further when its 64 bit capability is used.

Interestingly, this is also approaching the performance of Krait 200: http://browser.primatelabs.com/geekbench3/1558700

Of course, there's far more to a mobile SoC than CPU core performance (e.g. the GPU is still fairly low end), but it's still interesting IMO - a budget SoC is nearing the CPU performance of a flagship SoC from 2 years ago.

I've just noticed something about the Broadwell-U announcement; the smaller HD 5500 die seems to match the specs of the Core M/Broadwell-Y die, 24EUs, 1.3B transistors, 82 mm^2.

I wonder if it's the same (binned) die in different packaging etc? The die shots look pretty much identical too.

On another subject, I saw wccf posted an article claiming Maxwell wouldn't make it to the HPC space, in which case you have to wonder what's the motivation behind a huge-die GM200? I mean it would explain why we've just seen the GK210 refresh for HPC, but where does that leave GM? I'm just speculating of course, but the supposed lack of FP64 hardware would rule out another Titan, and would the relatively small high-end consumer market provide enough sales for it?

The GTX 900 naming seems to rule it out arriving too soon (and if lack of FP64 is true, I doubt Titan could be used), except for maybe 980Ti. Considering we're possibly looking at a huge, near-reticle-size die, even on a very mature node, defects would likely still mean a lot of part-disabled dies to sell, and there's not a lot of room in 900 for two more parts. Maybe 990?

Well I guess we'll see. The bits of rumours we're getting just don't seem to fit together very well. But then that's rumours for you...

[CAT-THE-FIFTH]

Carrizo reference notebook pictured:

http://www.tomshardware.com/news/amd...apu,28345.html

[watercooled]

It seems Anandtech have an Intel sponsorship now.

[CAT-THE-FIFTH]

It looks like a $105 A8-7650K is being released:

http://www.gamersnexus.net/news-pc/1...pu-105-dollars

Some AMD benchmarks are on the webpage in the link too.

It looks like the GTX960 will have GTX760/R9 280/R9 285 class performance but with lower power consumption. Whats the bet the tech press will hype the release up per the Nvidia review guide??

[watercooled]

Choo choo...

Well it would be disappointing if it didn't achieve something like that TBH. For the past weeks/months, Nvidia have had nothing in the mid-high end, it jumps straight from the 760 to the 970. You either can't buy anything between, or they're selling for stupid-prices.

I can understand discontinuing GK110 on the consumer market with it being such a massive die which was probably borderline unprofitable especially if they had to drop the price to fit in to the current market. But GK104 too? They seemingly discontinued it without having anything to replace it; I wonder if that has anything to do with the ~50% disabled mobile GM204 GPUs?

[CAT-THE-FIFTH]

Choo choo...

Well it would be disappointing if it didn't achieve something like that TBH. For the past weeks/months, Nvidia have had nothing in the mid-high end, it jumps straight from the 760 to the 970. You either can't buy anything between, or they're selling for stupid-prices.

I can understand discontinuing GK110 on the consumer market with it being such a massive die which was probably borderline unprofitable especially if they had to drop the price to fit in to the current market. But GK104 too? They seemingly discontinued it without having anything to replace it; I wonder if that has anything to do with the ~50% disabled mobile GM204 GPUs?

Don't get me wrong - lower power consumption is not a bad thing,but I expect a massive amount of hype for similar performance we had for the last year or so.
Both the GM107 and probably the GM206 seemed to be more mobile orientated where they make the most sense or even for OEM systems,but for DIY desktop they are rather meh,but the tech press seems to follow the Nvidia review guides and overhype them,when they really are not improving the most important metric for a budget card - performance.

Look at any of the sub £150 cards in the last two years - you could run them fine off a 400W to 500W PSU.

Even look at the R7 265 and the GTX750TI - even with the former you are lucky to see 200W consumed at the wall with an Intel CPU.

Edit!!

The thing is though I have a GTX660 which I bought nearly two years ago,and in 27 months we are likely to see only a 40% improvement in speed??

Add the fact the GTX960 is likely to be current for a year or so,that means that 40% improvement is spread over three years.

[DanceswithUnix]

Anandtech review the FX8320E, and manage a 50% overclock

http://www.anandtech.com/show/8864/a...-95w-vishera/2

[CAT-THE-FIFTH]

AMD Carrizo OpenCL benchmark leaked:

http://wccftech.com/amd-carrizo-grap...#ixzz3PDmNjDoM

BGA version might come to desktops:

http://wccftech.com/kaveri-refresh-l...#ixzz3PDmYBnm5

It appears Carrizo will be made using a modified 28NM process and it appears to use much denser wiring:

https://translate.google.co.uk/trans...27_677727.html

So not only are there are design changes,but there are also deep physical changes in circuit design for Carrizo over Kaveri.