1 |
On Wed, 2006-01-25 at 12:24 -0700, Duncan wrote: |
2 |
> B Vance posted <1138205784.11529.1.camel@×××××××××××.ShadowRealm>, |
3 |
> excerpted below, on Wed, 25 Jan 2006 11:16:24 -0500: |
4 |
> |
5 |
> >> virtual cpus? |
6 |
> >> |
7 |
> >> intel has hyperthreading, if you mean that, not amd. |
8 |
> >> |
9 |
> >> the rule of thumb is cores+1 |
10 |
> >> |
11 |
> >> but it is totally personal taste, what you set for jX. |
12 |
> >> 1,2,3,4,5 just try it and choose the number where the result is most to your |
13 |
> >> liking. |
14 |
> > |
15 |
> > Cool, thanks. Must of been thinking of something else that AMD had |
16 |
> > licenced from Intel for core CPU use. Could have sworn that AMD had a |
17 |
> > version of Hyperthreading. |
18 |
> |
19 |
> Maybe you are thinking about (the unrelated) HyperTransport? That's an |
20 |
> all-AMD thing, the reason they seem to have such good performance compared |
21 |
> to Intel, which is stuck with legacy front-side-bus architecture for the |
22 |
> moment, with it throwing a /serious/ wrench into their upgrade plans ATM. |
23 |
> |
24 |
> As for Hyperthreading, AMD hasn't used it and really doesn't need or want |
25 |
> to. Hyperthreading works its magic by working around the extremely deep |
26 |
> pipeline that Intel had used, in ordered to get their clock rates so high. |
27 |
> Such a deep pipeline forces a *HUGE* penalty when there's a branch |
28 |
> miss-prediction and the CPU has to throw away several cycles worth of work |
29 |
> and let the bottom end of the pipeline idle while it loads the correct |
30 |
> instructions from the top of the pipeline again. Hyperthreading is a |
31 |
> rather neat solution that avoids the severe penalty by having a second |
32 |
> task ready and waiting to execute while the first one gets loaded into the |
33 |
> pipeline once again. |
34 |
> |
35 |
> AMD has a somewhat shorter pipeline, and uses some other technologies to |
36 |
> improve the branch prediction and lessen the severity of mis-prediction |
37 |
> hits, both. It isn't /that/ much shorter a pipeline, the branch |
38 |
> prediction isn't /that/ much better, and the miss-prediction penalty not |
39 |
> /that/ much less, but put all three together, and add in the effect of the |
40 |
> front-side-bus vs hypertransport on memory latency, and the individual |
41 |
> single-digit improvements begin to add-up to something rather bigger. |
42 |
> Hyperthreading only yields an approximate 30% increase in performance, |
43 |
> best-case, anyway, compared to a theoretical 100% increase if it were a |
44 |
> real CPU. Thus, AMD doesn't have to cover as much ground with their |
45 |
> individual "minor" improvements as one might think, and because |
46 |
> hyperthreading has certain downsides (effectively halving the available |
47 |
> cache per thread, and certain thread-cache-access security issues where |
48 |
> the hyperthread has access to the memory of the other thread), they |
49 |
> believe they do better without hyperthreading. |
50 |
> |
51 |
> I believe AMD does better without hyperthreading as well, from a technical |
52 |
> viewpoint, altho Intel has the PR game won with hyperthreading, as it's |
53 |
> simply not possible to explain in two sentences why AMD doesn't need |
54 |
> hyperthreading, as opposed to the relatively simple concept Intel has to |
55 |
> get across, of pretending there are two CPUs so the second task can get |
56 |
> some work done when the first task is waiting for something. |
57 |
> |
58 |
> -- |
59 |
> Duncan - List replies preferred. No HTML msgs. |
60 |
> "Every nonfree program has a lord, a master -- |
61 |
> and if you use the program, he is your master." Richard Stallman in |
62 |
> http://www.linuxdevcenter.com/pub/a/linux/2004/12/22/rms_interview.html |
63 |
> |
64 |
> |
65 |
Thank you for explaining that Duncan. Guess it's time for me to start |
66 |
reading a bit more then I have lately. |
67 |
|
68 |
B. Vance |
69 |
|
70 |
-- |
71 |
gentoo-amd64@g.o mailing list |