More than an FPU, less than a GPU

I think that there is an opportunity to enhance, or augment, the processing units in our current PCs.

Augmenting processors is not a new idea. Intel supplied numeric coprocessors for its 8086, 80286, and 80386 processors. These coprocessors performed numeric computations that were not natively available on the main processor. The main processor could perform the calculation, but the coprocessors were designed for numeric functions and performed the work much faster.

(Intel did not invent this idea. Coprocessors were available on minicomputers before PCs existed, and on mainframes before minicomputers existed.)

A common augmentation to processors is the GPU. Today's GPUs are in a combination of video adapter and numeric processor. They drive a video display, and they also perform graphic-oriented processing. They often use more power than the CPU, and perform more computations than the CPU, so one could argue that the GPU is the main processor, and the CPU is an auxiliary processor that handles input and output for the GPU.

Let's consider a different augmentation to our PC. We can keep the CPU, memory, and input-output devices, but let's replace the GPU with a different processor. Instead of a graphics-oriented processor that drives a video display, let's imagine a numeric-oriented processor.

This numeric-oriented processor (let's call it a NOP*) is different from the regular GPU. A GPU is designed, in part, to display images, and our NOP does not need to do that. The NOP waits for a request from the CPU, acts on it, and provides a result. No display is involved. The request that we send to the NOP is a program, but to avoid confusion with our main program, let's call it a 'subprogram'.

A GPU does the same, but also displays images, and therefore has to operate in real-time. Therefore, we can relax the constraints on our NOP. It has to be fast (faster at computations than the CPU) but it does not have to be as fast as a GPU.

And as a NOP does not connect to a display, it does not need to have a port for the display, nor does it need to be positioned in a PC slot for external access. It can be completely in the PC, much like memory modules are attached to the motherboard.

Also, our NOP can contain multiple processors. A GPU contains a single processor with multiple cores. Our NOP could have a single processor with multiple cores, or it could have multiple processors, each with multiple cores.

A program with multiple threads runs on a single processor with multiple cores, so a NOP with one processor can one run 'subprogram' that we assign to it. A NOP with multiple processors (each with multiple cores) could run multiple 'subprograms' at a time. A simple NOP could have one processor, and a high-end NOP could have multiple processors.

Such a device is quite possible with our current technology. The question is, what do we do with it?

The use of a GPU is obvious: driving a video card.

The use of a NOP is... less obvious.

It's less obvious because a NOP is a form of computing that is different from our usual form of computing. Our typical program is a linear thing, a collection of instructions that are processed in sequence, starting with the first and ending with the last. (Our programs can have loops and conditional branches, but the idea is still linear.)

Our NOP is capable of performing multiple calculations at the same time (multiple cores) and performing those calculations quickly. To use a NOP, one must have a set of calculations that can run in parallel (not dependent on each other) and with a large set of data. That's a different type of computing, and we're not accustomed to thinking in those terms. That's why the use of a NOP is not obvious.

There are several specialized applications that could use a NOP. We could analyze new designs for automobiles or airplanes, simulate protein folding, explore new drugs, and analyze buildings for structural integrity. But these are all specialized applications.

To justify a NOP as part of a regular PC, we need an application for the average person. Just as the spreadsheet was the compelling application for early PCs, and GPS and online maps were the compelling app for cell phones, we need a compelling app for a NOP.

I don't know what such an application be, but I know something of what it would look like. A NOP, like a GPU, has a processor with many cores. Today's GPUs can have in excess of 6000 cores (which exceed today's CPUs that have a puny dozen or two). But cores don't automatically make your program run faster. Cores allow a program with multiple threads to run faster.

Therefore, a program that takes advantage of a NOP would use multiple threads -- lots of them. It would perform numerical processing (of course) and it would operate on lots of data. The CPU would act as a dispatcher, sending "jobs" to the NOP and coordinating the results.

If our PCs had NOPs built in, then creative users would make programs to use them. But our PCs don't have NOPs built in, and NOPs cannot be added, either. The best we can do is use a GPU to perform some calculations. Many PCs do have GPUs, and some are used for numeric-oriented programming, but the compelling application has not emerged.

The problem is not one of technology but one of vision and imagination.

- - - - -

* I am quite aware that the term NOP is used, in assembly language, for the "no operation" instruction, an instruction that literally does nothing.