Microstuttering tests on ATI Radeon HD 4870 X2

[Hornet331]

No but some are very I/O hungry and the problem here is GRID. That game seems to scale well using threading and it is also intensive in graphics. And what we are discussing here is situations that could slow the game temporarily.
If the problem was in the GPU then it would be more logical that it showed up I more games. New games that will be out and if they are heavy on I/O, is using more threads will probably confirm or prove it false.

any source for that claims?

i just played a bit with the system monitor in xp, which allowes you to log I/O for a certain app. Crysis has avg ~100 I/O per second and ~1,7mb/s transferred (peak of ~50mb/s and 6000 I/O while loading) (level was onslaught, which has quite some action going on. )

i dont know for grid, but i doubt it will be significantly higher.

[gosh]

any source for that claims?

i just played a bit with the system monitor in xp, which allowes you to log I/O for a certain app. Crysis has avg ~100 I/O per second and ~1,7mb/s transferred (peak of ~50mb/s and 6000 I/O while loading) (level was onslaught, which has quite some action going on. )

I don’t think that you only have ~100 I/O requests per second to the GPU playing Crysis You have A LOT MORE.
Also it isn’t the size that is important; it is the number of requests to memory and GPU that is important.
If you fetch 1000 bytes (not much) but you get one byte on each request. If you have a cache miss for each request and say that it takes about 250 clocks to get it. Then you have 1000 * 250 = 250 000 clocks getting 1000 bytes. Compare this with finding data in the L2 cache (15 clocks). 1000 * 15 = 15 000 clocks. If the Front Side Bus is working with another thread or cache is used for another application the hit rate will go down and as you can see in this example the speed is will go down.
If one thread has all these misses in the cache and slows down and there are another thread that is dependent on work to be ready then that thread will also be slower

Now it will not be that ineffective in games but normally it is small requests and it isn’t the size that matters, it is the latency for each request. OC the FSB will improve performance for the whole computer on Intel because latency goes down.

Crysis is almost one single threaded game and the problem with scaling may not show up there. If just one thread manages the I/O and/or it also will be the main memory user then I don’t think it will be a problem.
If you are using a lot of threads and all threads are sending or getting information than complex situations could appear. Not often but in a world that executes ~2 000 000 000 operations each second it could be often if we think about it. Also if there are other applications in the computer that is working the game will have to share resources and that slows it down.

On AMD the situation is different because it has Hypertransport and that manages I/O alone, it doesn’t need to compete with memory traffic. Also AMD Phenom can read and write data at the same time.