cancel
Showing results for 
Search instead for 
Did you mean: 

Archives Discussions

WTrei
Journeyman III

Integer Performance of Southern Islands Graphic Cards

Hi folks,

I am currently planning to buy new hardware for our research team. Currently we are massively using integer multiplication in our projects, because we are trying to implement several number theoretic algorithms to GPU's.

Currently, our hardware cluster mainly consists of HD 5800 type graphic cards, since they even offer a higher throughput than HD 6900 ones (because beside the integer mult we can do some additions in parallel within the same stream processor while 6900 series cards lack this feature due to their new architecture. Thus I want to ask how integer Ops are performed on HD7700 and HD 7900 type cards, especially can every one of the 64 SIMD cores proceed one integer mult per clock or have multiple of this cores have to work together or does one of this need more than one clock?

This architectural information would be a great help planning future hardware buyings and implentational decisions.

Thanks in advance!

WTrei

0 Likes
1 Solution

No, performance of 32-bit integer multiply is not affected by performance of double-precision arithmetic.  I can't verify it at the moment, but I have the documentation in front of me   32-bit integer multiplies are performed at 1/4 the floating point rate on all SI GPUs.

View solution in original post

7 Replies
nou
Exemplar

24 BIT INT MUL/MULADD/LOGICAL/SPECIAL @ full SP rates

– Heavy use for Integer thread group address calculation

– 32-bit Integer MUL/MULADD @ DPFP Mul/FMA rate

from http://developer.amd.com/afds/assets/presentations/2620_final.pdf page 44

that mean 1/4 rate on 7900 and 1/16 on 7700

WTrei
Journeyman III

Uhh - that is not the answer I hoped to see - 1/4 Sp performance on HD 7900 series is ok, but only 1/16 on HD 7700 series is quite low. Does that mean that integer multiplication performance on 7770 is only roughly 1/3 of my older 5770? We hoped to have at least 1/4th of the SP performance even on HD 7700 series so we can buy some for having cheap testing platforms with low power consumptions, but this was... no go

Maybe I should hope for the 7800 series then ...

0 Likes

with new GCN architecture all GPU will get DP which rate can be scaled from 1/16 to 1/2. so my assumption is that 7800 will get 1/8 rate and maybe some FirePro will get full 1/2 rate (just speculation).

0 Likes
WTrei
Journeyman III

I just saw even the 7800 series cards have only 1/16th of SP-performance for DP. Can any member of AMD confirm that this also means that 32 bit integer multiplication is cut by that factor? If this is true the new 300$ HD7870 cards will run my programs slower (max the half performance) than my old 100$ 5770 card

I am realy disapointed by the new architecture regarding integer performance

0 Likes

No, performance of 32-bit integer multiply is not affected by performance of double-precision arithmetic.  I can't verify it at the moment, but I have the documentation in front of me   32-bit integer multiplies are performed at 1/4 the floating point rate on all SI GPUs.

Integer maths runs at full speed (ADD/SUB/CMP/SHL etc.) on Southern Islands with the exception of int MUL as Jeff_golds pointed out. This was exactly the same as on Cypress (only 24bit Int MUL was full speed, but 32bit int MUL ran at 1/4 speed - identical to the Nvidia gt200 series as well) so it will not be slower than your old card.

0 Likes

In response to nou, HD7800 parts have 1/16 rate DP performance like HD7700 parts.

0 Likes