3 ms·
I can't seeing this be effective in any other situation. My understanding is that the bandwidth to and from the GPU is the big bottleneck. Once its on the card,
by skolor 17y ago
I can't seeing this be effective in any other situation. My understanding is that the bandwidth to and from the GPU is the big bottleneck. Once its on the card, the processing is fast, but getting things into the GPU memory is the bottleneck.
If, on the other hand, they have some sort of complex heuristics that is heavily process intensive, this may be very successful. As it stands, I don't think (although IANASecurity Expert) a definition based solution will work do to the memory limitations.
- Xichekolas 17y agoI don't think the transfer to GPU memory is so much a bottleneck as it involves some latency. The PCIe bus provides a ton of bandwidth, it just takes longer to get stuff there from main memory than it does to get stuff to the CPU from main memory. So, wasn't trying to pick nits, just trying to say that if it's a latency issue, the penalty is roughly the same for small amounts of data as it is for big amounts, and if you have a huge amount of data that takes non-trivial amount of processing, the latency penalty is, as you said, worth it.
- cschneid 17y agoIf I recall correctly, AGP, the older graphics interface was very lopsided with respect to bandwidth (very fast to send data to the card, very slow to get data back). As far as I understand, the newer PCI Express (newer == several years old now), has more balanced (and fast) bandwidth.