|
Cybernetic Vermin posted:I think you are vastly underestimating the number of extremely pessimistic global locks in software today. There are bound to be a lot of software that go from being just "multi-threaded" to basically scaling linearly in some operations with cheap memory transactions. edit: you're basically arguing that there are a lot of applications that are inherently parallel, CPU bound, and extremely impacted by locking overhead to some shared object. those are the only cases where transactional memory could theoretically make a performance difference. what applications would those be? also, please note that Haswell TSX won't solve those problems due to the inability for any app that relies on a monolithic global lock to fit all of its shared state in L1. the transactional memory people have been tooting this horn for ten years; if it were actually as amazing as all the academic papers claim, Rock probably wouldn't have killed Sun Professor Science fucked around with this message at Feb 23, 2013 around 23:36 |
| # ? Feb 23, 2013 23:24 |
|
|
|
| # ? May 20, 2013 06:29 |
|
|
Professor Science posted:you're basically arguing that there are a lot of applications that are inherently parallel, CPU bound, and extremely impacted by locking overhead to some shared object. those are the only cases where transactional memory could theoretically make a performance difference. what applications would those be? also, please note that Haswell TSX won't solve those problems due to the inability for any app that relies on a monolithic global lock to fit all of its shared state in L1. Right, that is the claim I am making (assuming that you are saying "locking overhead" not to mean overhead but rather actual lock contention). Hard to have any real statistical foundation for this, but I have worked on a fair bit of software myself where a lot of time is spent waiting on a lock on some huge collection where each thread will touch only a tiny random subset of the data held. It is not that this situation occurs because finer-grained locking is that hard, but it is just one of those things one knows is very hard to reason with perfect certainty about, so locks tend to be very pessimistic to be sure they are covering sufficiently much. To some part the way to view this is that a perfectly successful transactional memory implementation will give you the performance of the finest-grained locking possible with the level of bug-resistance (and effort) of coarse locking. Your phrasing of my statement is a bit disingenuous though, of course I am talking about CPU-bound software, since I have no idea what you expect Intel or AMD to put in their CPUs to improve the situation for software that is not CPU-bound. It should also be clear that any discussion is in terms of successfully eliminating false lock dependencies, since no parallel processing technology will help when there are real dependencies. If you find a huge step towards perfect extraction of parallelism to be amazing then fine. I have no idea how you imagine that transactional memory on Rock would have saved Sun, since they very notably didn't actually manage to make a Rock CPU, a very notably recurring phenomena when it came to Sun hardware promises (on the other hand they very successfully wasted money on unprofitable stuff like Java and OpenOffice).
|
| # ? Feb 24, 2013 00:47 |
|
|
Crysis 3 does in fact favor AMD CPU's, according to this test from yesterday. This is kinda cool, perhaps it means I should go for hyperthreading when I upgrade to a Skylake/AMD-equivalent chip later if these trends continue?
|
| # ? Feb 25, 2013 19:15 |
|
|
Keep in mind that that's a rather odd mix of overclocked and non-overclocked processors, I wouldn't read too much into it until we get repeatable, correctly tested results from reputable English-language sites.
|
| # ? Feb 25, 2013 19:33 |
|
|
So what's new at AMD?
|
| # ? May 17, 2013 01:56 |
|
|
Waiting for Kabini and HSA to blow us away, I guess.
|
| # ? May 17, 2013 03:03 |
|
|
The 8970M is out (7970M with boost clocking). Richland and Kaveri are starting to sneak out. Richland is just showing up now in gaming notebooks, a higher-clocked A10 in the same power envelope. Kaveri is coming next June and it'll be the next APU revision with HSA junk. Yawn.
|
| # ? May 17, 2013 03:14 |
|
|
PS4 and Next-Box news are the only things really - and anyone who yawns at the thought of HSA - wtf?
|
| # ? May 17, 2013 16:05 |
|
|
Richland is literally just a higher-clocked Trinity. Kaveri might be nice, what with GCN and HSA and an on-chip ARM core, but we know practically nothing about it compared to Haswell, which is coming out at about the same time. Except that it's Gen 3 Bulldozer. Big deal? Who knows?
|
| # ? May 17, 2013 16:17 |
|
|
https://twitter.com/anandshimpi/sta...518239830454272 Apparently we may see NDA lift on Jaguar next week. Brazos was exciting and popular. Brazos 2.0 was not so much. Here comes Tiny APU For Baby Computers Gen 3.
|
| # ? May 19, 2013 12:05 |
|
|
|
| # ? May 20, 2013 06:29 |
|
|
There are benchmarks of Temash out there already. http://www.notebookcheck.net/Review...sh.92264.0.html It compares pretty well to Atom. No sign of Kabini yet, though.
|
| # ? May 19, 2013 12:40 |
|
