Teraflop
This post's title is not a reference to anything around me (for the record).
Instead, it refers to a performance measure, for the non-computing scientists among the readers. More specifically, a teraflop (or Teraflop, or teraFLOP, or TeraFLOPS, or whatever) means one trillion (that is, 10^12) floating point operations per second. In any case, it is almost a lulz number.
The important part is how fast we got there. In 1999, Apple released the G4 PowerMac, which was touted as "the world's fastest desktop computer ever" and which broke the gigaflop barrier for home computers, at 1-4 Gflops. Less than a decade later, I am sitting in my lab here at Hopkins and am poring over hardware specifications.
The task is to order a single workstation PC for some ultrasound projects in our group. Many of those projects involve lots and lots of correlation computations - elastography or strain imaging is full of them; for 128x2000 ultrasound RF arrays and "small enough" windows, this amounts to a huge number of correlations. So we moved beyond quad-cores, dual-processors, hyperthreading, clustering, and so on - instead the focus is on GPGPU programming (effectively CUDA), or programming a high-performance supercomputation device. Nvidia's Tesla cards are one example of this species, one of which will probably end up in the workstation-class PC, completely dwarfing the computation power of the CPUs inside. A single Tesla C1060, for example, has a performance of nearly one Teraflop, at a price point for a complete system not much higher than $5000. Even a fully-specced 4-card system with almost four Teraflops under your desk clocks in at well below $10k, including dual-Xeons, 16GB of RAM, and friggin' huge monitors.
This is a speedup factor of 1000 within one decade. In this light, that older post sounds very nostalgic.
Instead, it refers to a performance measure, for the non-computing scientists among the readers. More specifically, a teraflop (or Teraflop, or teraFLOP, or TeraFLOPS, or whatever) means one trillion (that is, 10^12) floating point operations per second. In any case, it is almost a lulz number.
The important part is how fast we got there. In 1999, Apple released the G4 PowerMac, which was touted as "the world's fastest desktop computer ever" and which broke the gigaflop barrier for home computers, at 1-4 Gflops. Less than a decade later, I am sitting in my lab here at Hopkins and am poring over hardware specifications.
The task is to order a single workstation PC for some ultrasound projects in our group. Many of those projects involve lots and lots of correlation computations - elastography or strain imaging is full of them; for 128x2000 ultrasound RF arrays and "small enough" windows, this amounts to a huge number of correlations. So we moved beyond quad-cores, dual-processors, hyperthreading, clustering, and so on - instead the focus is on GPGPU programming (effectively CUDA), or programming a high-performance supercomputation device. Nvidia's Tesla cards are one example of this species, one of which will probably end up in the workstation-class PC, completely dwarfing the computation power of the CPUs inside. A single Tesla C1060, for example, has a performance of nearly one Teraflop, at a price point for a complete system not much higher than $5000. Even a fully-specced 4-card system with almost four Teraflops under your desk clocks in at well below $10k, including dual-Xeons, 16GB of RAM, and friggin' huge monitors.
This is a speedup factor of 1000 within one decade. In this light, that older post sounds very nostalgic.
posted by PJS at
Saturday, March 07, 2009
0 Comments:
Post a Comment
<< Home