Re: Anybody Home? (DS20 needs CPU+RAM)

Hi !

gl@xxxxxxxxxxxx wrote:
3) a friend has a 3CPU AS4100 (big beast, to say the least).
the agregated memory bandwidth (despite older memory technology)
is larger than a single-CPU EV6.
Mine is a 4x466 4MB L2 each
But when I misured memory bandwidth with hdparm, I only found 140MB/s or so.
Maybe I'm recalling wrong, maybe hdparm is not the best test for memory
bandwidth, but the P4 3000 I'm writing from does 1174.47MB/s
I will test the AS4100 bandwidth again tonight.
Could you tell me how is the performance of XP900 memory measured by
hdparm?

hdparm is not a tool designed to measure the memory bandwidth,
but rather, the OS's ability to move data around...
and eventually to/from disk :-)

A lot of parameters come into the equation,
and if your program's "kernel" (main loop) fits data and code
in the L1 or L2, it flies.

The multi-CPU ALPHA boxes have a huge (comfortable for each CPU)
memory bandwidth (well, looking back at the 1999 era),
IIRC 64 bits*200MHz/CPU (1,6GB/s peak) for EV6.
For a DS20, it's 3,2GB/s peak for a dual CPU system,
all that made of 256-bit banks of 83MHz of SDRAM.
According to http://seb0099.free.fr/Kannar%20PC/tableau.JPG
the same (theoretical) bandwidth is available today on a desktop,
with much faster RAM and a 128-bit bus.

I use and like the Alphas because i can do "heavy" works
that are painful to optimise for x86. 32 64-bit registers
are a blessing for some of my applications. x86-64 does not
even address this issue (there are 15 usable registers).
I thought that two of the 32 registers were special purpose, zero and
null, if I'm not wrong. In this case they are 30, but still double the x86
registers!
sure, registers are critical !
the more there are registers, the less we access memory (in theory)

So you are developing on alpha for business customers?
What kind of applications?

i would have loved using Alphas for Computational Fluids Dynamics
(CFD), because i have developped in the past an algorithm
that uses mostly binary operations (XOR, AND, OR, SHL/ROR...)

i have finally optimised it to death for MMX (P55) but it was like a cedar tree in the @$$.

Today i'd like to play with 3D 'voxels', still using bits and bytes,
treated in large chunks. 64-bit pointers become a must-have for
spaces larger than 2K*2K*2K bytes.

I always wanted to start some development with alpha assembly, but never
found the time for.
*sigh*
me too but at least i have exciting projects ATM
(i'm an independent electronician and work with artists and designers)

Last week I printed the whole "Alpha Assembly language guide" to start
studying. Do you know some good tutorial or book available from the
internet that could help?
no idea right now, but in 2000, i bought an alpha assembly book
(it's in some archival boxes now). Maybe i could sell it (for a small fee)
but to me, it is too "basic", explaining "traditional RISC stuffs".
I'm too l33t for this :-)

I guess i once collected alpha asm pages from the net,
i may bundle a .tgz if i can boot the hosting computer.

However i'm a "low level guy" and just programming apps on this platform
is not enough for me. If you don't see what i mean, look at http://f-cpu.org


ciao!
gl
yg
.

Relevant Pages

  • Re: Anybody Home? (DS20 needs CPU+RAM)
    ... 1,5) the EV6 has a separate L2 bus, so it's different. ... the agregated memory bandwidth ... even address this issue (there are 15 usable registers). ...
    (comp.os.linux.alpha)
  • Re: DRAM Speed vs Bus Speed
    ... Note that, the memory interface bus, is not 100% efficient. ... the FSB and memory performance. ... if you had dual channel memory at PC2-4800, ... That allows a doubling of memory bandwidth, ...
    (microsoft.public.windowsxp.hardware)
  • Re: DRAM Speed vs Bus Speed
    ... Note that, the memory interface bus, is not 100% efficient. ... the FSB and memory performance. ... if you had dual channel memory at PC2-4800, ... That allows a doubling of memory bandwidth, ...
    (microsoft.public.windowsxp.hardware)
  • Re: Strange memory problem
    ... is get a copy of memtest86+ from ... memory bandwidth indicators on the upper left hand ... memtest86+ should report a faster memory bandwidth on ... but CPUZ got the wrong memory speed, ...
    (microsoft.public.windowsxp.hardware)
  • Re: Cool (but useless?) expose feature
    ... >>> Trick to show off the Mac's graphic capabilities. ... > the real IO killer on the AMD chips is the ondie memory controller, ... > 6.4GB/s of memory bandwidth that's UTTERLY seperate from the I/O bandwidth ...
    (comp.sys.mac.advocacy)
Loading