Given by Peter Kogge Notre Dame at JNAC (PetaFlops) Presentation on August 28,1996. Foils prepared August 24 1996
Outside Index
Summary of Material
This was part of a set of PetaFlop (JNAC) Presentations to group of Federal Program Managers |
JNAC = Joint National Advanced Computing Initiative |
This uses Moore's Law Projections of Technology for Logic and Memory |
and uses Bodega Bay Application Analysis to cost memory for a "realistic" machine |
Outside Index Summary of Material
Slope of Each: approx 2.7X every 3 years |
Assumes: |
* 100% efficient utilization |
* SIA Growth in Clock Rate |
* Growth in instruction issue parallelism |
In 2004 technology (for 2007 machine) |
160,000 such processors gives 1 PF peak. |
Problem: each requires caches, memory, I/O! |
$1 Billion |
$10 Million |
$ are chip costs along (packaging, cooling, "batteries" extra)
|
"% of Petaflops Apps" refers to Bodega Bay Applications
|
Real systems probably require some multiple of this memory for next app preload, I/O buffers, etc. |
PETABYTE |
TERABYTE |
15 min |
1 Day |
1 Year |
10 TB Primary Memory supports almost all apps whose run time is under 1 day + over half of those that run > 1 day. |
For 2007 conventional machine, using 2004 technology and a mere 10 TB primary memory
|
At $1,000 per uP (cache, cooling, ...), this is approx
|