CPU vs. GPU

What is the primary and overridding difference between the CPU and the GPU is a nice consumer computer? It is NOT the speeds at which vectors can be processed, it is NOT the ability or lack of ability to processes double precision numbers.

The most important factor is who makes the most money from that PC you bought. Under the current configuration a nice graphics card can account for 50% of the cost of a computer and the GPU often costs more than the CPU.

This situation does not set well with Intel and AMD. The latter has made a move to change this by purchasing ATI and working on a new computer design that combines the capabilities of the CPU and the GPU and brings more computer revenues to the compined company (AMD+ATI).

Intel, on the other hand, is changing the paradigm from inside the company and inside the CPU. Their Larrabee project is looking to perovide a multi-core chip that includes cores that can handle the graphics that have traditionally been owned by the GPU.

So what is Nvidia doing to defend their very profitable turf? It appears that they are pursuing high-applications with their Tesla product that uses multiple GPUs to handle high-compute problems. Given that the consumer desktop is where all of the money is, you would expect them to be doing their own innovation in the consumer space. That may include multi-core GPU, multi-chip cards, combined CPU/GPU architectures ... Or something entirely different.

THe situation where the GPU pulls in a significant share or the PC price is equivalent to a serious threat to Microsoft's ownership of the O/S and Office productivity tools. The power of Intel HAS to rise up to reclaim these revenues. There will be a new architecture for consumer grade computers in CPU/GPU specifically because of the current revenue share ... Intel will make it happen. The real question is why has it taken so long?

Nvidia Tesla Supercomputer

Nvidia has released a “supercomputer” in three small form factors: (1) a PCI card that fits inside a desktop computer, (2) a deskside box that connects to a PC, and (3) a 1U server that fits in a rack. These machines are based on GPU graphics chips and the smallest provides 128 computing cores on a single card. These are vector processors which are very good at performing the same mathematical operation on large volumes of data – such as processing photo images or radar returns, computing fluid flow, or calculating line-of-site. Unfortunately, they are not the most efficient at processing logical code like that found in simulators. However, as the cost point for these machines comes down, they may brute force their way into being a useful solution for us.

Nvidia Tesla website

Server-side Rendering - Sun and Nvidia

Sun has been working with Nvidia to create a capability to do both the computation and any associated rendering on the server side. Then they stream the screen image to the client device. This is significant switch from what we do with the DIS and HLA federations now. But Sun’s goal is to make it possible to experience rich 3D scenes on lightweight client devices because all of the rendering is done on the server. They are also working on a capability to use new graphic chips to render “Pixar quality” images in real-time for display in CAVE environments.

If Sun is successful then it is an indication that network bandwidth is becoming plentiful enough that we can change the model we have used for decades of creating very small data packets and doing all of the scene generation on the client side. This is valuable for customers who do not want to have to hold a powerful graphic machine in their hand (like a cellphone). Instead, customers will be able to see rich 3D worlds on very minimal computing clients, e.g. something that is capable of playing MP3 movies today. This brings down a significant commercial barrier. Even the cheapest cellphones and pocket PCs would be able to play a rich 3D game because the game would really be running and rendering on the server. It is hard to imagine a world in which bandwidth is that plentiful for the consumer. Probably it would be rolled out to industrial customers for limited applications and private bandwidth first. Their product name for this is TurboVNC.

On the military side, we would be able to tap into any scene that anyone in the training event is seeing. We could see it on a regular cellphone (if/when it becomes available through a cellular network) or a wireless pocket PC.

Teraflop Computing in Your Palm

DARPA is looking to build embedded computers out of commercial graphics chips. Their target applications are small, hand-held devices like sensors (imagery, chem/bio, motion, acoustic, etc.). They would prefer to do the signal processing on the sensor device rather than downloading it to a processing center. Downloading requires lots of bandwidth and power to send the info. If the processing is on the sensor, then the transmitted information can be simplified to something like [ID, Target Type, Location, Velocity ... and other parameters]. This can go in short transmission bursts. If the processing is on the sensor, then the computer/chip has to be small, low cost, and low power. Reusing commercial chips is the best approach to get low cost because commercial customers will amortize the development costs. That is what led them to Nvidia chips.

The project looks forward to a day when the processing on the sensor is equal to a current supercomputer (Teraflops). Imagine a hand-held digital camera like something that you can get at Best Buy - but with 500Mpixels and a computer inside. Once the picture is taken the camera can process the image, extract the people in the picture, search a local database and identify the names of the people in the picture, their general location, and the season of the year. All of this could become metadata so that when the picture is posted to the web (like Flikr) the metadata explains what is in the picture and everyone on the net can search for pictures with specific characteristics. If the sensor also has a cellular net connection, the pictures can be uploaded in real-time.

What does this have to do with simulation and training?
IF such a computer in a handheld device existed, then you could run WARSIM or OneSAF on your own Palm Pilot-sized device. You could also hook up to everyone else who is running a simulation on their Palm Pilot, share scenarios, collaboratively train. It would be to modern simulation centers what Wikipedia has been to the Encyclopedia Britannica. It would allow the masses to create and run their own exercises. Like Wikipedia vs. Britannica, the sim center staff is going to immediately criticize this plan because it lacks the "experts". But if you have followed what has happened on Wikipedia you will have noticed that "the masses" have a lot of expertise and are quick to share it when a medium like Wikipedia is available.