CFX Vs 3-way SLI

ECH

Regular
[H] originally did a review on 6990+6970 (each at stock clocks of 830MHz/880MHz) vs 3 580 SLI (At 772MHz) here using a:
-MSI x58
-920 OC 3.60GHz
-6 Gigs of DDR3 1600 ram
-etc

The results at that time showed the AMD solution to beat the Nvidia solution. Folks complained and the benchmark test was done again but this time using:
-ASUS P8P67 WS Revolution (NF200 controller)
-2600K OC to 4.8GHz
-DDR3 Memory Kit DDR3 2000 CL 10
-etc

This time Nvidia shows a substantial improvement in scaling while AMD's solution does not. AMD's CFX is also showing a performance decrease over the previous PC setup using the same games/hardware, etc. It's not clear why a faster CPU/ram is exhibiting this behavior with AMD's CFX. Some in the forums are questioning:
-NF200 chip
-Drivers
-Game
-etc
But it's not really clear what's going on. Anyone have any thoughts why this is happening?
 
You'd have to do the same testing on a bunch of other chipsets, and use the same damn speed RAM to get a better picture.
 
Hmm, the MSI Eclipse x58 should offer x16/x16/x4 natively. I believe the Asus P8P67 WS Revolution does not offer this natively. If so, then perhaps that explains the bandwidth issue.
 
Last edited by a moderator:
Interesting.. normally we would assume the AMD setup to be more cpu limited.

Would have been nice to (also) see the 2600k at around stock clock, to see if we should point the fingers at NF200 (it smells a bit that the 6990+6970 has lower performance than in the 920 system in some games).. Also 3*6970 to have the same kind of PCIe setup would have been more useful.

Hmm, the MSI Eclipse x58 should offer x16/x16/x4 natively.

Really? Guess it would be running 8x8x8x8x with 3 cards.. otherwise that would also be part of the explanation (slowing the 3*580 down).
 
Interesting.. normally we would assume the AMD setup to be more cpu limited.

Would have been nice to (also) see the 2600k at around stock clock, to see if we should point the fingers at NF200 (it smells a bit that the 6990+6970 has lower performance than in the 920 system in some games).. Also 3*6970 to have the same kind of PCIe setup would have been more useful.



Really? Guess it would be running 8x8x8x8x with 3 cards.. otherwise that would also be part of the explanation (slowing the 3*580 down).

From what I recall P67 chipset boards don't offer two x16 pcie slots. The x58 boards do.
 
Last edited by a moderator:
But it's not really clear what's going on. Anyone have any thoughts why this is happening?

Maybe the Tri-Fire became CPU-limited relatively. nVidia driver eat a lot of cycles, if they contain AVX-optimized portions it's possible (maybe not likely, it's all presumption) that the nVidia driver eats its more OPS faster than the AMD driver eats its less but slower OPS.
Is there an indication if a single 580 goes faster from i7 to SandyBridge, in terms of utilization of CPU in driver code? I mean something like 30% CPU on i7 but 15% on SB? In a scenario like this 3x 580 may already fill up the CPU and eat from the games, while on SB 55% of SB is surely enough for a game?
If nVidias driver-code isn't a problem anyway this isn't an explanation, sure.
 
Maybe the Tri-Fire became CPU-limited relatively. nVidia driver eat a lot of cycles, if they contain AVX-optimized portions it's possible (maybe not likely, it's all presumption) that the nVidia driver eats its more OPS faster than the AMD driver eats its less but slower OPS.
Is there an indication if a single 580 goes faster from i7 to SandyBridge, in terms of utilization of CPU in driver code? I mean something like 30% CPU on i7 but 15% on SB? In a scenario like this 3x 580 may already fill up the CPU and eat from the games, while on SB 55% of SB is surely enough for a game?
If nVidias driver-code isn't a problem anyway this isn't an explanation, sure.
Why say that...Is it nv's driver-code put much concentration on using all power of GPU and when GPU power is enough ,CPU is limited?
One more question is why driver use FP not INT?
 
Back
Top