Re: Athlon 64 memory latency numbers

Thomas Womack <[EMAIL PROTECTED]> writes:
[Athlon 64]
>And it doesn't seem to have anything like the trouble that I recall
>Intel chips having with TLB misses when you start randomly accessing
>meaningful-sized blocks of memory.

Well, looking at the desktop P4 results below, the problem does not
look too bad on the Pentium 4, either.

>I'm starting to regret my assumption that laptops are all the same ...
>these are figures in cycles (2.8GHz cycles) from the HP laptop I
>picked up on short notice in July, they're large enough that I
>don't quite believe them even given that the laptop has one-channel
>DDR and a low-end Radeon chipset. Almost-microsecond memory latency
>seems absurd for a this-year UKP1300 machine, even if it is running
>on batteries.
>
>bytes          cycles
>128            4.43
>256            4.38
>512            4.43
>1024           4.39
>2048           4.41
>4096           4.34
>8192           4.35
>16384          55.8
>32768          47.6
>65536          55.8
>131072         57.5
>262144         55.0
>524288         1028.3
>1048576                1854.3
>2097152                1806.4
>4194304                1787.9
>8388608                1761.7
>16777216       2015.9
>33554432       1898.5
>67108864       2134.4

Ouch.  I find them hard to believe, too; the L1 and L2 numbers are too
large; the main memory numbers, too, but I can imagine explanations
for that (sending the RAM into some low-power state right after every
access, then taking longer to wake up).  Could it be that the CPU was
running in low-power mode at a reduced clock frequency (because it was
running on batteries, or because it was running too hot), but you used
the maximum clock rate for computing these numbers?

Here are numbers from a 2.26GHz Pentium 4 with an i845E chipset:

    bytes     cycles
     1024       2.16
     2048       2.11
     4096       2.44
     8192       3.39
    16384      20.29
    32768      19.51
    65536      19.05
   131072      19.42
   262144      21.94
   524288     237.87
  1048576     369.78
  2097152     370.27
  4194304     365.31
  8388608     363.74
 16777216     363.68
 33554432     365.82
 67108864     373.40
134217728     383.23

- anton
-- 
M. Anton Ertl                    Some things have to be seen to be believed
[EMAIL PROTECTED] Most things have to be believed to be seen
http://www.complang.tuwien.ac.at/anton/home.html