Article: Parallelism at HotPar 2010
By: Michael S (already5chosen.delete@this.yahoo.com), August 9, 2010 4:35 am
Room: Moderated Discussions
AM (myname4rwt@jee-male.com) on 8/9/10 wrote:
---------------------------
>Michael S (already5chosen@yahoo.com) on 8/6/10 wrote:
>---------------------------
>>AM (myname4rwt@jee-male.com) on 8/6/10 wrote:
>>---------------------------
>>>none (none@none.com) on 8/5/10 wrote:
>>>---------------------------
>>>>I picked the one which claims x300 speedup.
>>>>
>>>>http://www.opticsinfobase.org/oe/abstract.cfm?uri=oe-17-22-20178
>>>>
>>>>"In comparison,we compiled the CPU implementation, tMCimg,
>>>>using the “-O3” option with the gcc compiler and double-
>>>>precision computation on a single core of an Intel 64bit Xeon
>>>>processor of 1.86GHz."
>>>>
>>>>So single core, low frequency, double precision, out-of-the
>>>>box compilation. Enough said.
>>>
>>>They got these results using a card available (new) for $80+ these days http://www.google.com/products?q=9800+GT&scoring=p
>>
>>>
>>>> (8800GT no longer avail, 9800GT uses the same G92 chip).
>>>
>>>Apparently, $80 won't buy you even the cheapest new Core 2 Duo, let alone a 1.86 GHz Xeon, which can cost as little as $188 (E5502) or as much as $3157 (L7555) http://www.intc.com/priceList.cfm
>>
>>That's stupid argument. Video card can't do anything on its own.
>
>But neither can any CPU. Right, you can't run a system w/o a CPU, so to make the
>argument valid we should adjust for the price of the cheapest CPU. But even then,
>the $45 delta (cheapest Celeron) is just enough to shift to the cheapest Core 2
>Duo. Keep in mind that component costs (PCB, RAM, all the supporting circuitry)
>work against the GPU, and do so to a larger extent when low-cost hw like said 9800 GT is considered.
>
The point is, your "apparently, $80 won't buy you even the cheapest new Core 2 Duo, let alone a 1.86 GHz Xeon" is bogus, because $70 above platform's absolute minimum do buy you 2x2.93 GHz + no-nonsense cache, i.e. 3.15 times more raw FLOPs than what was used as "CPU" reference. Add to that double precision. Then add almost certainly non-used SIMD on CPU side. Then add less than stellar compiler (although gcc is, probably better than its reputation). And suddenly we are back into 10x gain range which, everybody here agree, is sometimes possible. All that still on dirt chip CPU hardware rather than $1000 hexacore.
---------------------------
>Michael S (already5chosen@yahoo.com) on 8/6/10 wrote:
>---------------------------
>>AM (myname4rwt@jee-male.com) on 8/6/10 wrote:
>>---------------------------
>>>none (none@none.com) on 8/5/10 wrote:
>>>---------------------------
>>>>I picked the one which claims x300 speedup.
>>>>
>>>>http://www.opticsinfobase.org/oe/abstract.cfm?uri=oe-17-22-20178
>>>>
>>>>"In comparison,we compiled the CPU implementation, tMCimg,
>>>>using the “-O3” option with the gcc compiler and double-
>>>>precision computation on a single core of an Intel 64bit Xeon
>>>>processor of 1.86GHz."
>>>>
>>>>So single core, low frequency, double precision, out-of-the
>>>>box compilation. Enough said.
>>>
>>>They got these results using a card available (new) for $80+ these days http://www.google.com/products?q=9800+GT&scoring=p
>>
>>>
>>>> (8800GT no longer avail, 9800GT uses the same G92 chip).
>>>
>>>Apparently, $80 won't buy you even the cheapest new Core 2 Duo, let alone a 1.86 GHz Xeon, which can cost as little as $188 (E5502) or as much as $3157 (L7555) http://www.intc.com/priceList.cfm
>>
>>That's stupid argument. Video card can't do anything on its own.
>
>But neither can any CPU. Right, you can't run a system w/o a CPU, so to make the
>argument valid we should adjust for the price of the cheapest CPU. But even then,
>the $45 delta (cheapest Celeron) is just enough to shift to the cheapest Core 2
>Duo. Keep in mind that component costs (PCB, RAM, all the supporting circuitry)
>work against the GPU, and do so to a larger extent when low-cost hw like said 9800 GT is considered.
>
The point is, your "apparently, $80 won't buy you even the cheapest new Core 2 Duo, let alone a 1.86 GHz Xeon" is bogus, because $70 above platform's absolute minimum do buy you 2x2.93 GHz + no-nonsense cache, i.e. 3.15 times more raw FLOPs than what was used as "CPU" reference. Add to that double precision. Then add almost certainly non-used SIMD on CPU side. Then add less than stellar compiler (although gcc is, probably better than its reputation). And suddenly we are back into 10x gain range which, everybody here agree, is sometimes possible. All that still on dirt chip CPU hardware rather than $1000 hexacore.