Imagination, designer of mobile graphics IPs has announced its new B-series GPUs with a multi-chiplet based design. This is the first GPU line with an MCM outlay in the consumer market. Unlike the usual method of “pushing” the workloads to GPUs, these products work by “pulling” the workloads to the graphics processors. Imagination calls this a “decentralized” design.
One of the GPU chiplets is the “primary” or “master” GPU with the firmware and work scheduler. It divides the workload into tiles which are then pulled by the various GPU chiplets (slaves). Similar to their earlier designs, this rendering technique is tiled. Unlike AFR or SFR where latency penalties or temporal methods handicap performance, the work here is distributed to the different GPUs in the form of multiple tiles from one frame.
Another core difference between traditional and this multi-GPU approach is that it’s implemented on a hardware level. The OS and developers see it as a single GPU. In comparison, SLI and XFX work by allowing the driver/API to distribute the workload between the various GPUs. The explicit multi-GPU mode used by DX12 also lets the game engine allocate the workload on a lower level.
The basic design of the GPUs is unchanged with a TPU (not to confuse with NVIDIA’s TPUs), ALUs, geometry, pixel, and raster block. The HSR is the Hidden Surface Removal block that takes care of culling.
The BXT 32-1024 is the flagship product with up to 4 GPU chiplets, promising up to 6 TFLOPs of single-precision performance. The low-power BXE lineup is different from BXT and BXM in that it’s for the smallest form factor devices. It comes with a two-die solution. However, unlike the flagship products, the slave die lacks a firmware chip and a geometry block, relying on the master GPU’s units.
