top500.org lists the world's fastest supercomputers. For the foruth successive time, only 7 Canadian computers were included in the most recent list. There is some good news though, since five of these machines are installed at universities, one us used for industrial applications (C.O.R.E. Digital) and the remaining computer is the CMC machine in Dorval. All of the academic machines are clusters, with gigabit ethernet being the predominant cluster interconnect due to its low cost, although the second machine at the Universitie de Sherbrooke uses Infiniband. For the second succesive time the blade cluster at UBC is top of the list.
Rmax is the maximum computational performance (in GFlops) for the standard LINPACK kernel. The Rmax/Rmax(10) ratio is the speed relative to the machine in 10th place on the same Top 500 list.
Perhaps not surprisingly, there has been strong resistance to the new HPC Challenge benchmark suite. Vendors, in particular IBM, are accutely aware that it tests both latency and bandwidth of the interconnect as much as the processors. As such, a strong bias is shown towards machines such as the Cray X-1, which is indicative of the fact that ultimately a higher percentage of peak performance can be achieved on these machines (but they simply cost more). To gauge machine interconnect performance, the Nhalf value (the size of problem for which the machine achieves half its peak speed) is a good indicator of the machine bandwidth - a small value of Nhalf is indicative of good machine balance (a value 1/10 of Nmax is good). Noticeably a number of IBM Power4 clusters, and Gigabit clusters omit this value.
Academic machines constitute 16.0% (lower than on the previous list) of the top500, while industry dominates with a massive 55% of the systems (an increase from 48.4%) - clearly HPC skills are marketable and becoming ever more so! Second largest share, 22%, goes to `Research' machines such as those at weather centres.
Noticeably, of the seven main regional computing consortia in Canada only WestGrid and RQCHP are listed in the top 500. While a new round of CFI funding has been approved, HPCVL, SHARCNET and AC3 are currently undergoing RFPs.
The last place performance value has moved from 624 Gflops to 851 Gflops (a 36% increase, as compared to a 54% increase between the previous two lists). This is the first evidence of a slow down in the rate of increase - perhaps funding limits are finally being reached. Nonetheless, the next list will only include machines with more than a teraflop of performance.
Despite falling to number 3, Earth Simulator continues to exhibit a large performance advantage on real codes compared to the other machines in the top500. Although Blue Gene/L now sits at the top of the list it is highly questionable how well applications will run on it. Notably the Sandia Red Storm machine was not installed in time for this update of the top500.
In terms of vendor breakdown, IBM is again the predominant manufacturer, with a 43.2% market share (a slight decrease). HP market share has grown 6% to 34%. over HP. After these two vendors there is again an enormous drop off to third place which is held by SGI with an 4% share, and further, their market share continues to fall. Dell has leaped over Sun to have a 2.8% market share, while NEC now has a 2.4% share. Linux Networx has jumped to 2.2%, while Cray has fallen from 2.6% to 1.8%. Hitachi follows these two vendors at 1.2%. Sun comes in ninth, at a meager 0.8%, which is clearly demonstrating the lack of competiveness of SPARC hardware. Self-made clusters would actually rank seventh in a list of vendors at 2%.
Last Updated: Nov 2004