Author: Robert Hyatt
Date: 20:45:39 11/02/02
Go up one level in this thread
On November 02, 2002 at 18:08:49, Gian-Carlo Pascutto wrote: >On November 02, 2002 at 17:49:43, Robert Hyatt wrote: > >>I think the point being that Vincent's post was pure nonsense. SMT _does_ >>provide a speed-up. Not a slow-down as Vincent claimed... > >I'm quite surprised it does, in fact. I wouldn't have expected >parrallel_efficiency * smt_speedup to have equalled something >more than 2. > >How come SMT is efficient, at all? I have a basic understanding >what it is but I don't get why it'd work in a chessprogram at all. > >-- >GCP The cpu can run two threads. When, for example, one thread blocks on a memory read, which takes hundreds of clock cycles, the other thread can do useful work. It's exactly the same principle as the multi-tasking operating system uses to make multiple processes run more efficiently by overlapping the I/O of one process with computation by another. Just done at a micro-level so that the I/O becomes a memory read (or a device control read or whatever)... A single "thread" (SMT type thread) blocks so often waiting on memory reads, zillions of clock cycles are lost inside the processor while waiting. This eliminates some of that wasted cpu power.. A logical development, that has been talked about for years by the computer architecture folks, but it took breaking the 50M transistors/chip barrier to have enough circuitry to do this, since the CPU now basically has an internal "process scheduler"...
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.