By: Vincent Diepeveen (diep.delete@this.xs4all.nl), August 14, 2012 3:04 am
Room: Moderated Discussions
Gabriele Svelto (gabriele.svelto.delete@this.gmail.com) on August 13, 2012 9:16 am wrote:
> jp (asdasdf.delete@this.gmail.com) on August 13, 2012 1:54 am wrote:
> > When
>
> > measuring the actual power consumption of the GT240 we found that purely
> compute
> > bound applications didn't manage to max out the power consumption
> (50-60 % of
> > max power while reaching near peak GFLOP/s numbers). It was
> rather the bandwidth
> > bound applications that seemed to put more strain on
> the memory controllers that
> > were consuming the most power.
>
> That's not
> surprising considering that both ROPs and texture units can burn a significant
> amount of power in graphics workloads (they are both quite logic-intensive). I
> would assume that a pure computational load is unable to reach peak
> power-consumption, if you look at stress applications like OCCT those load the
> ROPs and TUs as well as the ALUs with a balanced workload in order to reach peak
> consumption.
Who told you to believe this fairy tale?
You get really close to max power consumption if you have something that is number crunching and doing lots of multiplications during crunching.
Good example you can download for free.
http://www.mersenneforum.org/showthread.php?p=253351
the program mfaktc is posted there. Compile it and run it in CUDA.
It has 2 modes. Benchmark mode and operation.
It's hardly eating memory bandwidth, everything really runs in the compute cores.
Even in benchmark mode where it is just basically multiplying zero's and where the branch prediction (that doesn't exist) takes everywhere the same branch (so skipping code hardly happens) there it already seems to use a lot of power; whereas it's very 'interesting' this already eats that much power.
Please note: this is basically integer codes of 32 bits.
> jp (asdasdf.delete@this.gmail.com) on August 13, 2012 1:54 am wrote:
> > When
>
> > measuring the actual power consumption of the GT240 we found that purely
> compute
> > bound applications didn't manage to max out the power consumption
> (50-60 % of
> > max power while reaching near peak GFLOP/s numbers). It was
> rather the bandwidth
> > bound applications that seemed to put more strain on
> the memory controllers that
> > were consuming the most power.
>
> That's not
> surprising considering that both ROPs and texture units can burn a significant
> amount of power in graphics workloads (they are both quite logic-intensive). I
> would assume that a pure computational load is unable to reach peak
> power-consumption, if you look at stress applications like OCCT those load the
> ROPs and TUs as well as the ALUs with a balanced workload in order to reach peak
> consumption.
Who told you to believe this fairy tale?
You get really close to max power consumption if you have something that is number crunching and doing lots of multiplications during crunching.
Good example you can download for free.
http://www.mersenneforum.org/showthread.php?p=253351
the program mfaktc is posted there. Compile it and run it in CUDA.
It has 2 modes. Benchmark mode and operation.
It's hardly eating memory bandwidth, everything really runs in the compute cores.
Even in benchmark mode where it is just basically multiplying zero's and where the branch prediction (that doesn't exist) takes everywhere the same branch (so skipping code hardly happens) there it already seems to use a lot of power; whereas it's very 'interesting' this already eats that much power.
Please note: this is basically integer codes of 32 bits.
