Project Crayon: A Truly General-Purpose Graphics Processing Unit

General purpose programming on graphics processing units (GPGPU) is a growing field. GPGPU relies on high parallelism in an application domain to exploit the massive parallelism of a GPU. However, graphics rendering is an embarrassingly parallel problem, one where finding parallelism is extremely easy. Some stages of the graphics pipeline can be parallelised to the pixel level – a level of parallelism not shared by many real-world problems.

Project Crayon reconceptualizes the GPU as an array of very simple CPUs design on RISC principles. each with minimal extra features to support graphics processing. The Intel Larrabee architecture was an attempt at implementing a similar concept, but with a starting point of a Pentium pipeline, which is not competitive with simple RISC pipelines of the same era.

Why Crayon? Some of the ideas are inspired by Seymour Cray’s vector machines. I would like to call the idea a Cray on a Chip (CrayOn) but the name Cray is trademarked. As a compromise, a crayon is a graphics rendering tool of a sort – so why not Project Crayon? It does after all draw on good ideas…

To avoid confusion, the project logo is clearly a line of crayons, not a trademarked family of supercomputer companies.

Stages of the project are:

Implement a simple graphics pipeline using a conventional instruction set
Identify the performance bottlenecks that can be addressed by special instructions.
Identify a minimal subset of additional features to address the bottlenecks:
- Start from general-purpose forms of additional parallelism like vector instructions.
- Add as small a number as possible of less accessible forms of parallelism.
Evaluate the approximate chip areas needed for the resulting design.
Work out how many cores the new design can have, allowing for a generous amount of on-chip static RAM, to create a reasonable GPU

A standard multicore design	Proposed Crayon design
Conceptual design of Crayon, composed of a scaled up multicore design with a network on chip interconnect.

References

Philip Machanick, How General-Purpose can a GPU be? [Viewpoint]. South African Computer Journal no. 57, 113–117. http://dx.doi.org/10.18489/sacj.v0i57.347
Philip Machanick, Project CrayOn: Back to the future for a more General-Purpose GPU?, Proc. 2nd Workshop on Pioneering Processor Paradigms, Vienna, 25 Austria 2018 [online].