From: Alejandro Dubrovsky (s328940@student.uq.edu.au)
Date: Fri Apr 12 2002 - 07:55:34 MDT
On Fri, 2002-04-12 at 22:28, Eugen Leitl wrote:
>
> In terms of access latency, it was ballpark 120 ns then, and it's ballpark
> 50 ns now (the costs for an access might have actually increased, because
> of elaborate mechanisms in the CPU, the memory interface, and within the
> memory itself). That's a factor of two, and since we've got broder buses
> now (a factor of eight), all I see is a factor of 16 for nonstreaming
> memory access. Which is obviously not 100. It would be actually
> interesting to write a NONSTREAM benchmark, and to run it on a number of
> systems, vintage and modern. I think the results would be surprisingly
> dismal, because so much hardware must be engaged before you can get at the
> contents of a remote memory location.
yes, but what kind of problem would generate close to zero contiguous
access? most problems have some locality, and in that case you'll get
cache hits, and cache access takes a couple of clock-cycles (3-4 ns).
Note also that since cache sizes nowadays are almost as big as main
memory was a decade and a half ago, for problems that would have fit in
main memory then, you'll probably get quite a large memory bandwith
gain, even in problems requiring random access patterns. The picture is
even more favourable for modern cpus if you consider that while memory
access was a relatively fast 120ns back then, the CPU couldn't handle it
that fast. eg the 8088 would take 5 to 12 clockcycles to access memory
which at 4Mhz translates to at least 1.2 microsecs, and that, as you
mentioned, for only 8 bits. So in the nonstream benchmark, i'd guess
that an Athlon 1.4 Ghz would be at least 100 times faster than an
8088@4Mhz.
Alejandro
This archive was generated by hypermail 2.1.5 : Sat Nov 02 2002 - 09:13:26 MST