“book” — 2005/9/30 — 15:44 — page 6 — #18
✐
✐
✐
✐
✐
✐
✐
✐
6
CHAPTER 1. INTRODUCTION
A noteworthy point is that the concept of real time is often not a very useful term in scientific comput-
ing. For instance, in chemistry, simulated time is the order of pico seconds but the computation takes days
to complete. The main idea of the simulations is often to see a chemical process in slow motion in order
to observe what is happening. This is not doable in a laboratory. In contrast, astro and sea simulations are
too slow and are therefore simulated at higher rates. In sea-flow simulations a couple of hundred years
are simulated per day. In conclusion, turn-around time in scientific computing is often of the order of 24
hours to 30 days [Vinter, 2005].
Looking at physics-based animation from a computer graphical perspective gives a slightly different
picture.
Fluid phenomena:
Smoke simulation for lar ge-scale phenomena [Rasmussen et al., 2003] using half a million parti-
cles takes 2–10 secs. of simulation time per frame, and rendering time takes 5–10 minutes per
frame (2.2 GHz Pentium 4). In [McNamara et al., 2003], key frame control of smoke simulations
takes between 2–24 hours to perform on a 2 GHz Pentium 4. Target-driven smoke simulation
[Fattal et al., 2004] on a 2.4 GHz Pentium 4 in 2D on a 256
2
grid takes 50 secs. for a 1 second
animation in 3D on a 128
3
grid it takes 35 minutes.
Particle systems:
In [Feldman et al., 2003] suspended particle explosions are animated. The simulation time ranges
from 4–6 secs. per frame on a 3GHz Pentium 4. With render times included, a simulated second is
in order of minutes. Animation of viscoelastic fluids [Goktekin et al., 2004] using a 40
3
grid runs
at half an hour per second of animation on a 3 GHz Pentium 4. Animation of tw o-way interaction
between rigid objects and fluid [Carlson et al., 2004] using a 64 ×68 ×84 grid on a 3GHz Pentium
4 with 1 GB RAM takes 401 simulation steps for one second of animation with average CPU time
per simulation step of 27.5 secs.
Cloth animation:
Robust cloth simulation without penetration [Bridson et al., 2002] of a 150 × 150 nodes piece of
cloth runs at two minutes per frame on a 1.2 GHz Pentium3. Untangling cloth [Baraff et al., 2003b]
for a cloth model with 14 K vertices yields additional cost of 0.5 secs simulation time per frame
on a 2 GHz Pentium 4. Changing mesh topology [Molino et al., 2004] during simulation for 1 K
triangles runs at 5–20 minutes per frame, 380 K tetrahedra runs at 20 minutes per frame. Stacked
rigid body simulation [Guendelman et al., 2003] with simulations of 500–1000 objects takes 3–7
minutes per frame on average.
Computer Games
Game consoles such as Play Station 2 (www.playstation.com) has only 32 MB RAM and 6.2
GFLOPS and common home PCs are for the most part inferior to the computers used by computer
graphics researchers . This has called for some creative thinking in computer game development to
reach the performance requirements. Furthermore, there must be time set aside for other tasks in
such applications as computer games. For instance, in Hitman from IO-Interactive, only 5–10% of
the frame time is used for physics simulation [Jakobsen, 2005]. The future may change these hard-
ware limitations. For instance, recent GPUs do not suffer from latency problems and promise 60
我的内容管理
展开
