Multithreaded CPU & GPU
- Thread starter vblh
- Start date
Aren't GPUs already VERY multi-threaded, in a sense? Doesn't get more parallel than that nowadays.
If you mean multi-core GPUs, there are some topics on the subject on the 3D Hardware forum.
If you mean multi-core GPUs, there are some topics on the subject on the 3D Hardware forum.
vblh said:
Nothing is certain yet.
multithreading doesn't apply to GPUs by the way.
The advantage of multithreading is that at some point it's becomes a cheaper way to increase performance.
Multithreading has disadvantages: harder to program; synchronization of threads is a bother; additional overhead lowers performance.
If Intel had the tech to double clock frequency of their P4s without requiring extreme cooling or expenses, they would have done, instead of going dualcore. But they can't.
To put it bluntly: multithreading is bad, but we are poor.
Add to above: cache thrashing issues. The more threads you got competing over the same space, the bigger the mess. Threads can keep ejecting each other out of the cache, causing LOWER overall performance (as witnessed in the original Northwood P4 for example). If nextbox has 1MB of shared cache space for six threads, it's going to be really interesting to see how that'll work out!Sandwich said:
Well, if the patents are accurate, then it has a few nice tricks to reduce cache contention issues - special streaming load/store instructions, that can bypass the cache entirely for data that doesn't need to be cached, cache-locking for when you absolutely want a certain thread to have data cached, etc. As long as people take advantage of those capabillities, then cache thrashing shouldn't be too much of an issue.
Northwood was a very straightforward smt design, which did the bare minimum to support smt, with very little attention paid to how the multithreading would affect cache performance (it was also intel's first attempt at smt, i believe, so those issues are understandable). By all reports, the xenon cpu has had much more attention given to the multithreading aspects, and cache architecture (and ibm have a ton of experience dealing with multicore/smt cpus, with shared cache architectures, in fields where maximum performance is hugely important).
Northwood was a very straightforward smt design, which did the bare minimum to support smt, with very little attention paid to how the multithreading would affect cache performance (it was also intel's first attempt at smt, i believe, so those issues are understandable). By all reports, the xenon cpu has had much more attention given to the multithreading aspects, and cache architecture (and ibm have a ton of experience dealing with multicore/smt cpus, with shared cache architectures, in fields where maximum performance is hugely important).
