Condensing an overload of parallel computing ingredients into a single architecture recipe

A Highly-Efficient and Tightly-Connected Many-Core Overlay Architecture

“…The BM is at the 3rd level in the memory hierarchy of the DRAGON architecture, the first being the Register File, the second being the LM and the last being the GM. In [3]- [5], the BM was implemented as a single dual-ported block that is composed of 16 Ultra RAMs. The inputs and outputs of the first port were shared by all PEs.…”

“…This datum can come from any BM bank, thus all these banks are accessible to all PEs and the original BM size in previous implementation can be emulated by combining storage capacity of all 16 BM banks. Moreover, in [3]- [5], since the BM data input was shared by all the 16 PEs in the cluster, storing data to BM was done in a sequential manner. A sampler was used to register the PE data outputs and to serialize them to the BM input port.…”

“…Finally, The BMC acts as a half duplex DMA that manages bulk data transfers from BM to LM following a call to an LDBM instruction. Similarly to previous implementations in [3]- [5], the same data management logic had been preserved, except, now, the BMC manages 16 separate channels, one for every BM bank/PE pair.…”

“…In the current implementation, a 2D Mesh topology had been adopted to simplify the design in order to allow it to fit into a packaged RTL kernel as described earlier. Nevertheless, we have shown in [4], [5] how the base architecture could implement a 3D Torus or even achieves a 4D Torus interconnection as presented in [3]. The limitation here comes from the number of physical SLLs (Super Long Lines) that has to cross the SLRs (Super Logic Regions) in the FPGA [18], in order to transfer the data from the HBM banks to their corresponding BCs.…”

“…DRAGON was first introduced in [3] and is based upon our cumulative work in [4], [5] where the design was partially implemented on FPGA (accelerator part only). Back then, the controller part was not yet finalized and its continually changing behavior had only been emulated through a software environment.…”