STILES@CC.USU.EDU (Dyke Stiles) (09/19/89)
Several people mentioned that they did not receive the earlier
comparison of the Transputer to other machines on the Flux
Corrected Transport fluid dynamics code, so I am posting it
again - with figures for the Motorola 88000 added.
Results are given for both one-D and two-D versions. The one-
D version has longer vectors, giving the better vector
machines an edge.
The one-D results show the machine, execution time in seconds,
relative speed ratio (assuming that 1 T800 = 1.0), approximate
cost in $1,000 units, and relative speed/cost ratio (assuming
8 T800s = 1.0 this time). The number of points in the one-D
problem was 2048, and 100 time steps were calculated.
All machines ran the same Fortran code. The T800s were on
Definicon boards, running at 20 MHz. The 88k system was on
Tektronix's development board, also running at 25 MHz.
Speed Approx. Rel. Cost.
Machine # Proc. Time Ratio Cost. Effect.
===========================================================
Ardent Titan 4 18.2 64.4 206 0.515
Alliant FX/8 8 20.6 56.9 1000 0.094
SG 240 GTX 4 43.4 27.0 263 0.169
T800 8 148.5 7.9 13 1.0
VAX 8650 1 293.3 4.0 300 0.021
SUN 4/280 1 306.9 4.0 78 0.081
T800 1 1172.3 1.0 -- -----
The SG 240 GTX is a Silicon Graphics machine.
The two-D problem was 32 x 32 points, again with 100 time
steps. NCube times are taken from Gustafson's article in the
SIAM Scientific & Statistical Computation Journal, July, 1988.
We assume he was running the same code, since ours also came
from Sandia.
Speed Approx. Rel. Cost.
Machine # Proc. Time Ratio Cost. Effect.
===========================================================
Alliant FX/8 8 4.7 101.4 1000 0.171
Ardent Titan 4 7.3 65.3 206 0.535
T800 8 62.2 7.7 13 1.0
88000 1 109.3 4.4 15 0.495
VAX 8650 1 111.8 4.3 300 0.024
NCube 16 118.3 4.0 30 0.225
SUN 4/280 1 132.0 3.6 78 0.078
T800 1 476.4 1.0 -- -----
The cost for the T800 system assumes $1.5k per processor plus
$1k for a cheap PC clone host. The cost for the 88k assumes
$10k for the processor board plus another $5k for some type of
VME host (a used Sun?? - I'm not really familiar with prices
here). This month's (Sept.) Electronics on p. 21 describes a
company that provides an 88k VME board with one processor for
about $10k.
The 88k results are courtesy of Jean Michel Favre of
Tektronix.
NCube has indicated the next release of their processor will
be four times faster than the current one.
Dyke Stiles
Electrical Engineering
Utah State Univ.
Stiles@cc.usu.edu
Stiles@usu.bitnetbraner@TPOT.TN.CORNELL.EDU (Moshe Braner) (09/19/89)
The T800 is starting to lag behind some in speed, but its cost effectiveness is actually improving. The cost of an 8-node system should drop below $5000 in the next few months given the dropping RAM prices and the recent drastic cut in T800 pricing. The main cost factor would be your RAM requirements. That $1000 price for the PC host is a bit unfair if you need more I/O capacity, etc. You may want a $3000 386-based host. But that's still relatively cheap. But what will happen when the m88000 and i860 drop in price, as is bound to happen some day? - Moshe Braner
hjm@cernvax.UUCP (Hubert Matthews) (09/20/89)
In article <8909191352.AA03741@tpot.tn.cornell.edu> braner@TPOT.TN.CORNELL.EDU (Moshe Braner) writes: >The main cost factor would be your RAM requirements. That has always been the case, and will continue to be the case for some time. CPU power is relatively cheap; large memories are not. >That $1000 price for the PC host is a bit unfair if you need more I/O >capacity, etc. You may want a $3000 386-based host. But that's >still relatively cheap. And the I/O capability doesn't scale as the CPU power does. Compare the ratio of CPU power to I/O bandwidth for an XT and a wizzo 386 box and you'll see what I mean. Too may people seem to look purely at CPU speed and not enough people seem to look at overall system performance. Why is the Logical Systems C compiler for the transputer a cross-compiler? Because passing the compiler and the source files over to the transputer is slower than using the PC to do the compilation locally. >But what will happen when the m88000 and i860 drop in price, as is >bound to happen some day? Prices for these new boards will probably be dictated by RAM prices, just like other boards. If you are hanging one of these boards on a PC, then you *need* a lot of memory to compensate for the host's poor I/O speed. (Try running a virtual memory system from a RISC board via a PC disk!). Look at the numbers for a typical RISC machine: $400 for the CPU, $300 for the caches (quite a variable cost), and $3000 for 16 MB of RAM. That's just raw chip costs, without counting board space and power consumption. Follow-ups have been redirected to comp.arch, as this discussion seems to be more general than just transputers. -- Hubert Matthews ...helping make the world a quote-free zone... hjm@cernvax.cern.ch hjm@vxomeg.decnet.cern.ch ...!mcvax!cernvax!hjm