Re: ATI intorduces the new HD 5870 today

In article <e8adnXPy97->,
William <> wrote:
>Highlights:
>
>The HD 5870 is capable of delivering up to 2.72 TeraFLOPS of computing
>power.

But still only a gigabyte of memory.

OK, the GPGPU-ish job I want to do (GNFS linear algebra: find a kernel
vector for a 30M*30M matrix with ~70 non-zero entries per row, a
simple matter of about a million multiplies of the 30M*30M matrix by a
30M*64-bit vector) is a somewhat esoteric one, but it would be really
handy to have 16G on the graphics card even if that had to be slower
than the blazing-in-all-senses RAM on current machines.

Or to be able to read over PCIe from main memory; I don't know if
ATI cards nowadays support this for buffers of >4G. I know the nVidia
cards are tied to 32-bit addressing to the point that a 16G Tesla
presents as four 3.75G cards ...

Tom

Thomas Womack wrote:
> In article <e8adnXPy97->,
> William <> wrote:
>> Highlights:
>>
>> The HD 5870 is capable of delivering up to 2.72 TeraFLOPS of computing
>> power.
>
> But still only a gigabyte of memory.
>
> OK, the GPGPU-ish job I want to do (GNFS linear algebra: find a kernel
> vector for a 30M*30M matrix with ~70 non-zero entries per row, a
> simple matter of about a million multiplies of the 30M*30M matrix by a
> 30M*64-bit vector) is a somewhat esoteric one, but it would be really
> handy to have 16G on the graphics card even if that had to be slower
> than the blazing-in-all-senses RAM on current machines.
>
> Or to be able to read over PCIe from main memory; I don't know if
> ATI cards nowadays support this for buffers of >4G. I know the nVidia
> cards are tied to 32-bit addressing to the point that a 16G Tesla
> presents as four 3.75G cards ...
>
> Tom

If you look for the swiss-army-knife of brute-force ask your next friendly
Deep-Thought agent. I suppose after the last job they run pretty idle now.

Ciao, pun intended
Niels