From the beginning of the project, and in accordance to the objectives of the COMPUSAPIEN project, the main results achieved can be divided in three different areas:
(1) The design of energy-minimal 3D many-core server architectures that enable energy-proportional computing by leveraging the use of heterogeneity, parallel computing and in-memory computing (IMC). In this respect, we have proposed novel IMC architectures by developing a novel in-cache computing accelerator, named BLADE. This accelerator has been integrated into ARM32 / 64 architectures. We have also incorporated High Bandwidth Memories (HBMs) into 3D many-core designs, and developed a novel simulation framework, named gem5-X, to evaluate new server architectures. Our server architectures were assessed for a wide range of novel HPC benchmarks, i.e. Convolutional Neural Networks, Genome Sequencing or video streaming transcoding, and Artificial Intelligence (AI) analytics.
(2) The design of an integrated power and microfluidic cooling delivery subsystem able to overcome the limits of current cooling strategies in 3D MPSoCs. By imitating the dual role of blood in the brain, the FCA technology is able at the same time to extract heat and generate power thanks to the temperature-driven electrochemical reaction of FCAs. In this respect, in COMPUSAPIEN we enhanced the PowerCool framework and integrated it with the 3D-ICE simulator to enable the exploration of FCAs and assess their benefits for a wide range of different architectures. Furthermore, we oroved that FCAs can provide enough power to sustain the caches of an HPC processor, or 50% of the power needed by a low-power accelerator layer in a 3D MPSoC. We also have analyzed the impact of FCAs on the power delivery network of 3D MPSoCs to reduce the voltage drop, the number of power TSVs for the 3D MPSoC, and to increase chip bandwidth with more signal TSVs.
(3) The system-level multi-objective management of 3D stack computing resources by trading-offs energy consumption, temperature and performance. Specifically, we have proposed joint cooling and workload management techniques at the server level, These techniques use as control knobs cooling, workload allocation (i.e. task mapping to cores and accelerators) and setting the frequency of cores and other hardware accelerators. To cope with the large dynamism of the problem and the huge design space, we have proposed reinforcement learning based techniques, both single-agent and multi-agent, to tackle the problem and improve all metrics jointly.
