Parallella, Raspberry Pi, FPGA & All That Stuff

Hi,

as you can see here the custom build of the eah client (out-of-place FFT) for the parallella is about 20% more efficient than the official client (NEON and in-place FFT).

Thank you,

The RAC of my Raspi2B hasn't stabilized yet, but if I calculate the theoretical peak RAC from it's average runtime of ca 49ksec per BRP4 unit for 62.5 credits (and doing 4 of them in parallel), the result is ca 440 credits / day.

I have no equipment to measure the power consumption of the Raspi2 with any meaningful accuracy, but this test looks credible http://www.cnx-software.com/2015/02/12/raspberry-pi-banana-pi-and-odroid-c1-boards-power-consumption/, where ca 2.3 W under load was measured. So we would get ca

190 RAC / W
for the Raspberry Pi 2. Not that bad at all. The Raspi B "1" would probably be more like 30 RAC/W (!).

(a more handy unit would probably be "credits per kWh or credits per Joule ;-) , but that's just a constant conversion factor calculation).

People who are currently participating in the Beta test of the new BRP6 GPU app have reported both higher power consumption and lower runtime, but I strongly suspect that the efficiency of the new app in terms of "RAC/W" will improve quite a bit, giving the GPU even more of a lead.

Caveat: Your i5-4690K and 750Ti combo crunch work units rom a different search (by now probably mostly BRP6) than the Raspberry PI2 (BRP4), so this is not exactly an apples-to-apples comparison. We at E@H always try to scale credit across different searches so that it reflects the different run time on the same hardware as accurately as possible, but in the end it's just an approximation).

HB

Yup, and note that the CPU of the Raspi2 isn't even a very recent design. The SoC of the new Samsung Galaxy S6 for example should be both fast and efficient.

HB

Wow, that's really impressive! (I just read the pdf overview)

This chip would be truely revolutionary, I think even more so than the Parallela. It probably s*cks at single threaded branchy code, but for anything more regular it could even outperform GPUs.

As promising as they look now I'm positive they can actually tape out the hardware. Ultimately the solution will fall and stand with software, as anything else than x86 does. It's a good sign they explicitly mention this and are working on seamless compiler integration. I have no idea how well this turns out. But time is actually working for them: the slower the manufacturing processes improve, the more time the software guys have to optimize and to make better use of the current hardware.

I disagree with their introduction, though, where they say caches and especially hardware managed ones were the worst thing happening to computation. This concept of hiding hardware complexity from software is actually what makes new hardware perform well with old code. One can argue that today's software ecosystem is more open and flexible than it ever was in the last 20 years or so, but before that I'd rather have an Athlon 64 X2 than a Playstation 3 Cell CPU.

Edit: actually, now that I think a bit more about it.. the concept beyond this CPU is related to what nVidia did with Maxwell: gain power efficiency by replacing complex scheduling logic with software (if it's sure that the software handles the job well enough - otherwise things break).

MrS