News - China retakes supercomputer lead with the Tianhe-2

[DanceswithUnix]

... they're also x86 based which is awesome.

They added a big custom floating point unit to an old Pentium design. That's like strapping a rocket to a moped, you are welcome to ride that, I would rather walk

It would work if they back ported the Phi architecture to mainstream CPUs such that you could buy an i7 and run 8 threads of identical code to the Phi albeit at much reduced speed. It sounds like the Phi extensions will remain Phi only though, so calling it x86 is wrong and an insult to our intelligence.

Even Nvidia got this right, they made cheap consumer cards like GT430 run FP64 instructions at a much reduced rate compared to the top end GPGPU cards, but the ability is there.

Oh course Phi failed as a graphics card, maybe the Intel CPU guys can see why and refuse to be tarnished by the project.

[brasco]

They added a big custom floating point unit to an old Pentium design. That's like strapping a rocket to a moped, you are welcome to ride that, I would rather walk

Haha, I like the analogy! I've read that the extenstions are only for the specific Xeon CPU and Phi combo. Someone described it to me as - "the Xeon mothership CPU and it's PCIe Phi drones"

Would love to see it develop into a mainstream accelerator, although Nvidia and even OpenCL are ahead in terms of software support and the fact I can use a consumer GPU for GPGPU is working out well.

[djfluff]

Chinese hackers crack facebook's 128bit encryption in 5...4...3..

[ThermalDroid]

Wonder what would happen if they overclock it?

It burns a hole in the earth and ends up in the south Pacific somewhere!

[DanceswithUnix]

Haha, I like the analogy! I've read that the extenstions are only for the specific Xeon CPU and Phi combo. Someone described it to me as - "the Xeon mothership CPU and it's PCIe Phi drones"

Would love to see it develop into a mainstream accelerator, although Nvidia and even OpenCL are ahead in terms of software support and the fact I can use a consumer GPU for GPGPU is working out well.

The difference is that AMD and Nvidia have been solving this for decades and know how to feed that many cores.

This amused me:

http://www.fudzilla.com/home/item/31...e-too-powerful

Too powerful indeed! Supposedly Phi failed as a graphics engine because the architecture bottlenecked like crazy, guess it sucks for compute as well. Guess they tried to take their moped around a corner rather than just a straight line

[brasco]

Ahaha, that sounds ludicrous.

My main problem with simulations is if they can't saturate the GPU, the speed boost is neglegible over running it on the CPU - so it's a case of finding the sweetspot, it's never simple is it? I thought it would be the same for this, but assumed there'd be plenty of huge simulations to throw at it!

I'm still happy there's competition between the 3 companies in the GPGPU sector, good for the end users.

[BlackShade]

i reckon minecraft would blue-screen it haha jk

[p14b]

Got up yesterday checked the weather FINE
Played golf at 07.30 RAINED AT 10.00
All this TECH and they cant even get that mornings weather forecast right