Intel Core i7 Review Thread

[JumpingJack]

your numbers are done with a SINGLE 8800GTX judging from those screens....

btw, 200MHz | 400 MHz | Increase in performance
average: 55.5 68.1 22.7 %
Snow: 64 64.1 0.1 %
Cave: 80.1 85.7 6.99 %

and this for a single 8800GTX is quite impressive if you ask me

what do you mean by sock puppet? this is my only XS account dont be silly...

btw, as you can see on my anand link, 2 GPU's scale quite the same on qx9770 and x58, not a lot of difference, but 3 GPU's act really weird
the FSB is just my guess, if your guess is more computational power because of nehalem, fine, then i just search some 5.5ghz qx9770 screens with 3-way sli...

A) Average is real time, the final numbers are updated when the scene finishes .... the Snow and Cave only have real meaning. 'Average' will change depending on when I press screen shot. So your first number is bogus... download Lost Planet demo to see what I mean. (Yes, it was an 8800GTX, 4870X2's were installed later)

B) Oddly, Anand's tech explanation counters your argument completely, did you comprehend what you actually linked?? In multiple GPU setups nVidia utilizes the broadcast feature of the PCIe, meaning all 3 cards gets the command set in one broadcast:

Broadcast technology allows only one message to be sent by the CPU where it is then received, replicated, and broadcasted to all GPUs, eliminating the need for multiple, near-identical transfers over the FSB.

What ever traffic the CPU will send over the bus will be the same for one card as for 3 ... this is typical in broadcast networks where each bus agent is sent the same data. Also, if FSB BW is so critical, why does not cutting it in half have closer to 50% impact? You did not answer that question.

Texture data is stored on card, and the only time bus transfers are used is when the texture required is not cached on the card. This is why memory keeps going up on video cards. When a texture is required for a new scene or object that is not in memory you will know -- that game stutters horribly (FSB, HT or even QPI) there is no bus currently available that matches the BW of VRAM to GPU...

Texture bandwidth is consumed any time a texture fetch request goes out to memory. Although modern GPUs have texture caches designed to minimize extraneous memory requests, they obviously still occur and consume a fair amount of memory bandwidth. Modifying texture formats can be trickier than modifying frame-buffer formats as we did when inspecting the ROP; instead, we recommend changing the effective texture size by using a large amount of positive mipmap level-of-detail (LOD) bias. This makes texture fetches access very coarse levels of the mipmap pyramid, which effectively reduces the texture size. If this modification causes performance to improve significantly,you are bound by texture bandwidth.

Texture bandwidth is also a function of GPU memory clock.

http://http.download.nvidia.com/deve...erformance.pdf

EDIT: Ok, so no puppet -- I wondered because you level of understanding of the concepts is on the order of Gosh's.