800 TFLOPS chip for ray tracing

Print Friendly, PDF & Email

Courtesy of Multicoreinfo.com,

Tops Systems Corp of Japan, a venture involved in multicore technology, together with Toyota Motor Corp and Nihon Unisys Ltd, both of Japan, is developing a dedicated integrated circuit (IC) for ray tracing*, an image rendering method used in 3D computer graphics (3D CG) processing. A total of 73 heterogeneous cores designed specifically for ray tracing operations will be single-chipped, and nine of these chips interconnected (see Fig). With high-definition (HD) resolution at 1920 x 1080 pixels, the target processing speed is 800 tera floating point operations per second (TFLOPS).

Researchers have figured out how to resolve the basic issues involved in the overall system architecture and application parallelism analysis, and will begin detail design and implementation as an application-specific IC (ASIC) shortly. They plan to fabricate the chips using 45nm manufacturing technology, and expect the chips to operate at 750MHz, integrating 130 million gates into a 17mm square footprint.

The primary application here is the generation of visual imagery. Why a ray tracing chip when GPUS are kicking butt and taking names? A different kind of graphical experience. GPUs operate on the assumption that the scenes being rendered have been polygonalized. Ray tracing can give a more accurate rendering of the scene, and create more realistic imagery. Of course, GPUs were created for the graphics industry too, and we co-opted them pretty successfully. Maybe we’ll be able to make use of these, too.

Full article here.


  1. […] Multicore IC for Realtime RayTracing via insideHPC.com. AKPC_IDS += “5839,”; Share this […]


  1. As exciting as the title is it is a bit misleading! It’s actually the total system performance that is aiming to be 800 TFLOPS (presumably sufficient to meet their visualisation/performance requirements).

    Each on of the nine dedicated devices, is specified to produce 88 TFLOPS and each is comprised of numerous cooperating clusters of very specialised cores (with limited amounts of local memory). Think of this as the ray-tracing equivalent of the MD-GRAPE machines!

    I struggle to see the economic case for this system. Is it only going to be used by Toyota’s Motor division or more likely the they hope to find interested users elsewhere to help defray the NRE’s involved in making such a specialised processor. If so they’d better have an interesting software interface to go with it. I’ve seen several HW accelerated systems for ray tracing over the years (ArtVPS being the most notable) but we also have Caustic Graphics doing something related today.

  2. App-specific it may be, but that’s still near a PF ina desk-side unit for 1kW. The text & figure suggests fixed point rather than DP floating point. Of course, ray tracing lends itself to tasks other than optical image generation too – e.g. non specular reflections, electromagnetic wavelengths other than light, target scattering, etc.