GPU Computing Will Change Your Life
nVidia and Partners Point the Way at Industry Confab
By Roger L. Kay
Last week, nVidia held a scaled-down GPU (graphics processing unit) Conference in San Jose.
The conference was much smaller than the previous year’s nVision blowout, which was like a
block party for more than 5,000 people in downtown San Jose, occupying all available venues with
LAN parties for gamers, classes for developers, celebrities for the masses, and pitch rooms from
entrepreneurs. This year, the scene was quieter, with 1,500 focused attendees, many of them
developers, sitting in three types of forums, which occupied most of the second floor of the
Fairmont Hotel.
The three forums were, essentially, plenary sessions on general subjects, smaller academic
seminars on theoretical topics, and “emerging company” gatherings, where small firms currently
using nVidia technology could show their wares. The “why” of having these three types of activities
together, while not immediately apparent, becomes clearer when you realize that nVidia is trying to
communicate a fairly esoteric message through all of this showmanship: GPUs will change how
computing is done, enabling entirely new ways of doing things.
The academics possess proofs of concept; they were able to show what will be. The emerging
companies are running real applications; they demonstrated what is. And the plenary sessions
— in addition to being a forum for nVidia to tout its good works and how it is contributing to this
industry ferment — presented the broader architectural picture and highlighted industry
developments with key nVidia partners.
The primary thesis behind GPU computing is that, with a lot of small processors doing little jobs
in parallel, GPUs can get certain kinds of work done faster. That’s why nVidia called its early
products, designed for gamers, graphics accelerators. GPU computing is evolving toward
applying the same parallel concept to computing tasks of a more general nature. Some of these
tasks — such as underground mapping for oil and gas exploration — render visible output, and
so may be confused with pure graphics acceleration. Others — such as Monte Carlo simulations,
used by Bloomberg to price millions of collateralized debt obligations (CDOs) — have no
particular graphic output.
The CDO pricing model illustrates well how GPU computing changes how work gets done.
Bloomberg was able to drop its requirement for running this complex application from 1,000
servers down to 48 when it paired nVidia Tesla GPUs with the eight-core x86 CPUs in each
system. In addition to handing an ever more complex financial model, the hybrid GPU-CPU
architecture enabled the company to run the model in a couple of hours rather than overnight. By
accelerating this important task, Bloomberg was able to change its business model. Rather than
next day, it could offer on-demand results, for which it could charge a premium.
More importantly, some of this acceleration is allowing execution in real time of tasks that were
formerly iterative. Until now, you had to run, wait, look at results, change things, run again, and
repeat until done. Now, results are in real time or nearly so. A demo done by nVidia CEO Jen-
Hsun Huang of physical re-rendering of high-quality ray tracing took about eight seconds to
resolve into a new view when parameters were changed. With this type of responsiveness, an
artist rather than a technologist can run the tool.
A great example of this was shown by Richard Kerris, CTO of Industrial Light and Magic (ILM), in
one of the plenary sessions. ILM, which is George Lucas’s vehicle for various properties such as
Lucas Film Animation, Lucas Licensing, Lucas Online, Lucas Arts, and Skywalker Sound, is
adopting hybrid computing with GPUs to simulate complex animated scenes. Years ago, when
the firm did the simulation of the killer wave for Perfect Storm, the one that finally toppled George
Clooney’s fishing boat, it took so long to render that they were able to run it only once. The output,
made up of millions of fluid particles, looked, in its unvarnished form, like a tiny tsunami curling its
way across a pan of Elmer’s Glue. “That’s the one we had to use,” said Kerris of the single
instance.
To create the maelstrom across which the ships fought in the Pirates of the Caribbean II, ILM
engineers needed 20 hours to render a single frame. To run smoothly, video needs at least 30
frames per second, and this scene went on for several minutes. At that rate one minute of render
takes 36,000 hours or more than four years. Of course, ILM achieved some time savings by
running the render across multiple machines in the company’s render farm.
Kerris cited a more recent example in which ILM used hybrid computing with GPUs to reduce the
time per frame it took to render the fire scene from Harry Potter from 13 hours to 10 seconds. This
quantum leap in computing power will change the way animated movies are made. Artists can
see the results of their tweaks soon after making them and have a chance to go back and make
the work better.
Some of the emerging companies showed how they are using GPU computing to layer objects
onto video, create visual training materials, make more realistic online games, and drop the user
into a 3D virtual world by way of a captured and manipulated camera image. In real time. Such
fast execution has brought real time tools to market so that artists, analysts, and explorers can
use them, not just software gurus and engineers.
Although nVidia gets some bragging rights for creating this conference and highlighting these
ideas, there is nothing to say that other firms’ GPU computing effort won’t contribute to this
revolution. AMD also has graphics chips being used to accelerate computing tasks. Even Intel
has an effort underway to bring out its own discrete graphics chips, which may someday play a
role in GPU computing.
© 2009 Endpoint Technologies Associates, Inc. All rights reserved.
The conference was much smaller than the previous year’s nVision blowout, which was like a
block party for more than 5,000 people in downtown San Jose, occupying all available venues with
LAN parties for gamers, classes for developers, celebrities for the masses, and pitch rooms from
entrepreneurs. This year, the scene was quieter, with 1,500 focused attendees, many of them
developers, sitting in three types of forums, which occupied most of the second floor of the
Fairmont Hotel.
The three forums were, essentially, plenary sessions on general subjects, smaller academic
seminars on theoretical topics, and “emerging company” gatherings, where small firms currently
using nVidia technology could show their wares. The “why” of having these three types of activities
together, while not immediately apparent, becomes clearer when you realize that nVidia is trying to
communicate a fairly esoteric message through all of this showmanship: GPUs will change how
computing is done, enabling entirely new ways of doing things.
The academics possess proofs of concept; they were able to show what will be. The emerging
companies are running real applications; they demonstrated what is. And the plenary sessions
— in addition to being a forum for nVidia to tout its good works and how it is contributing to this
industry ferment — presented the broader architectural picture and highlighted industry
developments with key nVidia partners.
The primary thesis behind GPU computing is that, with a lot of small processors doing little jobs
in parallel, GPUs can get certain kinds of work done faster. That’s why nVidia called its early
products, designed for gamers, graphics accelerators. GPU computing is evolving toward
applying the same parallel concept to computing tasks of a more general nature. Some of these
tasks — such as underground mapping for oil and gas exploration — render visible output, and
so may be confused with pure graphics acceleration. Others — such as Monte Carlo simulations,
used by Bloomberg to price millions of collateralized debt obligations (CDOs) — have no
particular graphic output.
The CDO pricing model illustrates well how GPU computing changes how work gets done.
Bloomberg was able to drop its requirement for running this complex application from 1,000
servers down to 48 when it paired nVidia Tesla GPUs with the eight-core x86 CPUs in each
system. In addition to handing an ever more complex financial model, the hybrid GPU-CPU
architecture enabled the company to run the model in a couple of hours rather than overnight. By
accelerating this important task, Bloomberg was able to change its business model. Rather than
next day, it could offer on-demand results, for which it could charge a premium.
More importantly, some of this acceleration is allowing execution in real time of tasks that were
formerly iterative. Until now, you had to run, wait, look at results, change things, run again, and
repeat until done. Now, results are in real time or nearly so. A demo done by nVidia CEO Jen-
Hsun Huang of physical re-rendering of high-quality ray tracing took about eight seconds to
resolve into a new view when parameters were changed. With this type of responsiveness, an
artist rather than a technologist can run the tool.
A great example of this was shown by Richard Kerris, CTO of Industrial Light and Magic (ILM), in
one of the plenary sessions. ILM, which is George Lucas’s vehicle for various properties such as
Lucas Film Animation, Lucas Licensing, Lucas Online, Lucas Arts, and Skywalker Sound, is
adopting hybrid computing with GPUs to simulate complex animated scenes. Years ago, when
the firm did the simulation of the killer wave for Perfect Storm, the one that finally toppled George
Clooney’s fishing boat, it took so long to render that they were able to run it only once. The output,
made up of millions of fluid particles, looked, in its unvarnished form, like a tiny tsunami curling its
way across a pan of Elmer’s Glue. “That’s the one we had to use,” said Kerris of the single
instance.
To create the maelstrom across which the ships fought in the Pirates of the Caribbean II, ILM
engineers needed 20 hours to render a single frame. To run smoothly, video needs at least 30
frames per second, and this scene went on for several minutes. At that rate one minute of render
takes 36,000 hours or more than four years. Of course, ILM achieved some time savings by
running the render across multiple machines in the company’s render farm.
Kerris cited a more recent example in which ILM used hybrid computing with GPUs to reduce the
time per frame it took to render the fire scene from Harry Potter from 13 hours to 10 seconds. This
quantum leap in computing power will change the way animated movies are made. Artists can
see the results of their tweaks soon after making them and have a chance to go back and make
the work better.
Some of the emerging companies showed how they are using GPU computing to layer objects
onto video, create visual training materials, make more realistic online games, and drop the user
into a 3D virtual world by way of a captured and manipulated camera image. In real time. Such
fast execution has brought real time tools to market so that artists, analysts, and explorers can
use them, not just software gurus and engineers.
Although nVidia gets some bragging rights for creating this conference and highlighting these
ideas, there is nothing to say that other firms’ GPU computing effort won’t contribute to this
revolution. AMD also has graphics chips being used to accelerate computing tasks. Even Intel
has an effort underway to bring out its own discrete graphics chips, which may someday play a
role in GPU computing.
© 2009 Endpoint Technologies Associates, Inc. All rights reserved.
Why Take Days When You Need Only Seconds?
 |
| |
 |