The interconnections between the FPGA modules are done with interconnect cables connecting regular FPGA I/Os. Multi-Gigabit-Transceiver (MGT) connections between FPGAs are realized via extension boards with optical links (Gigabit Ethernet and QSFP+).
A PCIe or Ethernet link interfaces the proFPGA system to the host and enables high-speed data exchange between GNs and HNs.
Processing cores in HPC systems
In HPC most processing tasks are done by high-performance CPUs that are optimized mostly for serial processing of sequential data. GPUs have a massive parallel architecture with thousands of smaller, more efficient computing units that have been designed to perform multiple tasks simultaneously. GPU accelerated co-processing, in which compute-intensive functions are outsourced to the GPU and the rest of the sequential codes is processed by the CPU, offers a much higher performance resulting in a much shorter execution time, but it can't be efficiently applied to all algorithms.
FPGAs offer a very high I/O bandwidth and allow a programmable, fine-grained and flexible parallelization. Because parallel computing models are becoming increasing- ly important, FPGAs have gained their share on the HPC market, while their maturity and capabilities are increasing at the same time. FPGAs are still comparably difficult to program, though they're moving closer to procedural languages and OpenCL.
Role and advantage of FPGAs in MANGO
The big advantage of FPGAs is that they are versatile. In the MANGO project they are used for two purposes – as a compute platform and as an emulation platform.
FPGAs as a compute platform are used for accelerating the processing of bulk data. In this case the FPGA is tightly coupled to the system as a tile (or a collection of tiles). In MANGO this approach will be validated through dedicated accelerator designs.
When used as an emulation platform the focus is on architectural exploration, both at system- and compute unit level. Emulated systems will be general-purpose, software-programmable compute units. The MANGO-approach mainly targets functional aspects and allows validation of the software stack. Physical performance will be inferred from performance counters, performance improvements will be measured in relative terms.
Summarizing, Philipp Ampletzer, Director Sales & Business Development at PRO DESIGN Electronic GmbH, says: “MANGO shows the importance of FPGAs for today’s and tomorrow’s HPC systems. FPGAs enable designers do explore different many core architectures and are at the same time indispensable as hardware accelerators for processing huge amounts of data. By supporting and subsidizing promising projects like MANGO the European Union acts against the American supremacy in the HPC field. Even if it’s still a long way to go, Europe is doing well in catching up.”
* Philipp Ampletzer is Director Sales & Business Development at PRO DESIGN Electronic GmbH in Bruckmühl, Germany.