I think it was mentioned either on TR or by Anand. Sample B3D and I'll make sure your primitive / cycle figures are correct. Eventually
I think it was mentioned either on TR or by Anand. Sample B3D and I'll make sure your primitive / cycle figures are correct. Eventually
Yeah it was on anandtech, specifically mentioned here (given that it was on a slide you could have seen it elsewhere too but I think noone else said so in the text):I think it was mentioned either on TR or by Anand. Sample B3D and I'll make sure your primitive / cycle figures are correct. Eventually
http://www.anandtech.com/show/6976/...wering-xbox-one-playstation-4-kabini-temash/4
Though there actually seems to be a bug on that slide, it says 64 SP ops / clock, 8 DP adds / clock (making this 1/8 rate which looks plausible), but 4 DP Muls/FMAs per 16 clocks which would make this 1/256 rate, now that would be really glacially slow, for the whole graphic chip 1 dp mul every two clocks
.(Though in any case even with 1/16 mul/fma rate if you've got some opencl kernel consisting of mostly DP ops it should run faster on the cpu - after all for SP the cpu can do 4x4 mul+add per cycle whereas gpu is 128x fmas so considering the clock difference the gpu has more than twice the theoretical flops, but the cpu can do 4x2 DP adds and 4x1 DP muls per clock whereas the gpu would be limited to 8 fma / clock which is definitely less (as gpu runs at 1/3 the clock) unless you need the added precision from the FMAs.)
I'm actually wondering if that 1/4 prim rate is better or worse than what the top-end ARM SoCs have. If I go by some numbers anandtech has published in the past for glbench triangle test and assuming they get somewhere close to their theoretical peak it actually looks like they might also be 1/4 prim / clock but I have never seen official theoretical figures for these chips.
Last edited by a moderator:
Anandtech has a brief article on the 1H 2014 Kabini/Temash successors, Beema and Mullins.
http://www.anandtech.com/show/7514/amd-2014-mobile-apu-update-beema-and-mullins
I had previously expected these to be just rebrands of the 2013 dies benefiting from fabrication process maturity, but it turns out they are new SoCs and include new Puma CPU cores as a further evolutionary step over Jaguar and Bobcat.
AMD are claiming 2x the performance/watt of Kabini/Temash on the same 28nm process (albeit without any explanation of how this is achieved) and supply the following impressive performance numbers:
http://www.anandtech.com/show/7514/amd-2014-mobile-apu-update-beema-and-mullins
I had previously expected these to be just rebrands of the 2013 dies benefiting from fabrication process maturity, but it turns out they are new SoCs and include new Puma CPU cores as a further evolutionary step over Jaguar and Bobcat.
AMD are claiming 2x the performance/watt of Kabini/Temash on the same 28nm process (albeit without any explanation of how this is achieved) and supply the following impressive performance numbers:
My guess is better power management, or more specifically finally a really working Turbo. These "old" Kabini/Temash have a bit lacking turbo functionality (if they have any at all), so they only reach anywhere close to their TDP with both combined (multi) cpu and (especially) gpu load (cpu benchmarks alone never seem to be anywhere close to TDP). Note AMD uses a performance metric of some-benchmark / TDP, not some-benchmark / real power draw for the 2x claim, hence the actual energy efficiency improvement will probably be much lower. Also, it may not be the exact same 28nm process (going for HPM?), which could help somewhat too. In any case these chips might turn out to be quite nice.
Max turbo on some of the Haswell i7 ULV chips is over 3GHz: http://ark.intel.com/products/75114/Intel-Core-i7-4650U-Processor-4M-Cache-up-to-3_30-GHz
I am pissed at AMD I have been waiting months for their low power stuff to be available to consumers as I wanted to rebuild my car PC, but AMD will never actually get it to the market. (Of course I haven't checked in a couple weeks now but seriously it was months after "release")
How am I supposed to blame OEMs that AMD doesn't have any in the retail channel? Are you talking motherboard makers? That is the only thing I can think of. When I look for kabini all I get is premade netbook kind of things
http://www.newegg.com/Product/Produ...=BESTMATCH&Description=kabini&N=-1&isNodeId=1
Anyway I should not be grumpy, but I really though August at the latest they would be available. AMD still has their older integrated models out that I could buy, but nothing of the newer ones. I think Intel will be taking my business quite soon on that front. I was just holding out since AMD had better graphics, but I really probably don't need to.
There are things like these out as well
http://www.newegg.ca/Product/Product.aspx?Item=N82E16813135363
http://www.newegg.ca/Product/Product.aspx?Item=N82E16813135364
There is FS1b, which is a pin-socket for Kabini. The latestet roadmap has this version on it.
It was wished by vendors, since soldered versions are a higher inventory risk.
Blazkowicz
Legend
Good, but still lacks availability, that's only one mobo vendor on one continent. At $145 it's expensive too, you could pretty much get some Haswell instead.
Yeah it has been frustrating. I would think they might want to get in on the rasberry pi sort of market with a setup like this. But obviously that is not the case. Intel has been doing so a little.
Car PC?
I'm guessing all the Kabini chips are BGA, so he's right it's down to the OEMs to build products around them. If they don't believe they can make money off them then they won't, simple.
Blazkowicz
Legend
There is a socketed Kabini, it appears on the current AMD roadmap and it is FS1b µPGA. I suppose it's about convenience for a few OEMs, and would be of little use to consumers like MXM graphics cards are.