According to the latest rumors, the Nvidia GeForce RTX 3090 and GeForce RTX 3080 would hide a transverse coprocessor on the other side of the PCB, making sense to use a fan on the front of the graphics card that could take advantage of airflow to reduce Their temperature. As indicated, this coprocessor would be focused on improving the performance of RayTracing.
Specifically, this coprocessor would serve to shelter all the dedicated hardware related to RayTracing, which should save costs instead of integrating everything in the same silicon.
This is one of the hypotheses, and it is that the second hypothesis happens to be able to be an AIC that works autonomously, similar to the old PhysX cards for physics acceleration. At least the rumors make some sense since it must be remembered that it was indicated that the RayTracing’s performance would be multiplied by 4x in the Ampere graphics compared to the current Turing ones.
As if that weren’t enough, Nvidia patented a cross-coprocessor design in 2018 (image below), adding credibility to either story. The leaks are intensifying as the launch date approaches, presumably in September, so more information will be available in the coming weeks.
“More specifically, an example of non-limiting technology, in this case, concerns a hardware-based traverse coprocessor that efficiently traverses an acceleration data structure, for example for RayTracing in real-time.
In more detail, the non-limiting technology example provides dedicated hardware to accelerate RayTracing. In non-limiting additions, a hardware coprocessor (hereafter referred to as “cross-coprocessor” or, in some additions, “tree cross-unit” or “UAT”) speeds up certain processes that support interactive ray tracing, including tests of the intersection of volumes with ray boundaries, positive ray intersection tests, and “instance” ray transformations, “says the Nvidia patent.