• We should now be fully online following an overnight outage. Apologies for any inconvenience, we do not expect there to be any further issues.

Has anyone ever tested SYNCHRONOUS memory performance for GPUs - asynch was bad for CPU's, but is it for GPU's?

MadRat

Lifer
Oct 14, 1999
11,999
307
126
When Intel first released the P!!!-EB (Slot 1, 512K external L2, 133fsb) we had to endure the tenous prospects of either overclocking the BX chipset (synchronous memory) or the VIA Apollo Pro 133A (asynchronous), which we all know the BX was better for performance. Most graphics processors run asynchronous between their core and their memory. I was wondering if anyone has ever tested the efficiency of using synchronous timings on graphic cards versus asynchronous timings.
 

DXM

Senior member
Jul 26, 2003
264
0
0
The first generation Radeons performed better when run synchronously but the difference was less than 5%.
 

beatle

Diamond Member
Apr 2, 2001
5,661
5
81
Since there is no FSB on video cards, there's nothing to run synch or asynch with the memory. If you mean running the GPU and memory at the same speed, you'd be putting a serious hurt on the performance (unless you can figure out a way to increase your GPU's clock 100% or more. Memory usually runs more than double the speed of the GPU on higher end cards.
 

Matthias99

Diamond Member
Oct 7, 2003
8,808
0
0
Memory usually runs more than double the speed of the GPU on higher end cards.

This is not true; I have a RADEON 9800Pro, and the default clocks are 380/680(DDR), or, in terms of actual clock rate, 380/340.

GeForceFX 5900Ultras run at 450/850, or 450/425 if you look at the actual physical clocks.

A CPU/GPU can't pull data from the memory more than once per one of its clock cycles (or once per FSB clock cycle in the case of a CPU with an FSB), so it makes very little difference to have the memory running faster than the GPU/FSB; it'll just be sitting idle during the extra clock ticks.

I haven't tried underclocking my GPU's core to match its memory, but I can't imagine that it would work better (overall) at 365/365 than at 435/365, or at 340/340 compared to 380/340. It may be more efficient per clock during memory-intensive operations, but the extra 10% or so of clock speed likely makes up for that by running faster during other operations.