FPGAs or ASICs? There is a long-running debate on this. FPGAs are extremely flexible while ASICs offer top efficiency but inflexible. We believe that FPGAs and ASICs are better together, to offer both flexible and efficient solutions. We propose single-package heterogeneous 2.5D integration of FPGAs and ASICs, using Intel's Embedded Multi-Die Interconnect Bridge (EMIB). Since the ASICs are separate chips from the FPGA, this approach (1) does not change FPGA fabric, allowing re-use of existing ecosystems (FPGA chips/packaging/boards/software/etc), and (2) allows freedom in the ASIC design (area/freq/process/etc unconstrained by FPGA fabric). This approach is more effective than developing traditional stand-alone ASICs. Intel® Stratix® 10 FPGAs already have EMIBs, enabling single-package integration with other chips, or "tiles". We propose leveraging them to mix-and-match any domain-specific ASICs with Stratix10 FPGAs.

FPGAs or ASICs? FPGAs are extremely flexible while ASICs offer top efficiency. We believe that FPGAs and ASICs are better together, to offer flexibility and efficiency. We propose single-package heterogeneous 2.5D integration of FPGAs and ASICs, using Intel's Embedded Multi-Die Interconnect Bridge (EMIB). Since the ASICs are separate chips from the FPGA, this approach (1) does not change FPGA fabric, allowing re-use of existing ecosystems (FPGA chips, packaging, boards, software), and (2) allows freedom in ASIC design (area/freq/process/etc unconstrained by FPGA fabric). Intel® Stratix® 10 FPGAs already have EMIBs, enabling single-package integration with other chips, or “tiles”. We propose leveraging them to mix-and-match any domain-specific ASICs with Stratix10 FPGAs. In particular, this work focuses on deep learning (DL) domain, which demands efficient tensor (matrix/vector) operations. We propose TensorTile ASICs for Stratix10 FPGAs to provide ASIC-level tensor performance, while relying on FPGA's flexibility for application-specific operations (e.g., Winograd). Our evaluation shows: (1) a small TensorTile offer much better tensor throughput than a large Stratix102800 FPGA; (2) FPGAs and TensorTiles mix-and-match provide scalable solutions (e.g., ~69 peak INT8 TOPs with 1×TensorTile+small Stratix10-400 FPGA, to ~194 peak FP16 TOPs with 6×TensorTiles+large Stratix10-2800); (3) AlexNet performance (performance/Watt) of Intel's DL FPGA design improved by 4× (3.3×) when enhanced with 2×TensorTiles.

We present an OpenCL compilation framework to generate high-performance hardware for FPGAs. For an OpenCL application comprising a host program and a set of kernels, it compiles the host program, generates Verilog HDL for each kernel, compiles the circuit using Altera Complete Design Suite 12.0, and downloads the compiled design onto an FPGA.We can then run the application by executing the host program on a Windows(tm)-based machine, which communicates with kernels on an FPGA using a PCIe interface. We implement four applications on an Altera Stratix IV and present the throughput and area results for each application. We show that we can achieve a clock frequency in excess of 160MHz on our benchmarks, and that OpenCL computing paradigm is a viable design entry method for high-performance computing applications on FPGAs.

Some of today's TM systems implement the two-phase-locking (2PL) algorithm which aborts transactions every time a conflict occurs. 2PL is a simple algorithm that provides fast transactional operations. However, it limits concurrency in benchmarks with high contention because it increases the rate of aborts. We propose the use of a more relaxed concurrency control algorithm to provide better concurrency. This algorithm is based on the conflict-serializability (CS) model. Unlike 2PL, it allows some transactions to commit successfully even when they make conflicting accesses. We implement this algorithm in a STM system and evaluate its performance on 16 cores using standard benchmarks. Our evaluation shows that the algorithm improves the performance of applications with long transactions and high abort rates. Throughput is improved by up to 2.99 times despite the overheads of testing for CS at runtime. These improvements come with little additional implementation complexity and require no changes to the transactional programming model. We also propose an adaptive approach that switches between 2PL and CS to mitigate the overhead in applications that have low abort rates.

Today's transactional memory systems implement the two-phase-locking (2PL) algorithm which aborts transactions every time a conflict happens. 2PL is a simple algorithm that provides fast transactional operations. However, it limits concurrency in applications with high contention by increasing the rate of aborts. More relaxed algorithms that can commit conflicting transactions have recently been shown to provide better concurrency both in software and hardware. However, existing approaches for implementing such algorithms increase latencies of transactional operations, require complex hardware support and alter standard cache coherence protocols. In this paper, we discuss how a relaxed concurrency control algorithm can be efficiently implemented in hardware. More specifically, we use a technique which approximates conflict-serializability and implement it in hardware on top a base hardware transactional memory system that provides support for isolation and conflict detection. Our novel hardware scheme is based on recording conflicts as they occur, instead of aborting transactions. Transactions serialize at commit time according to these conflicts by sending broadcast messages. Our evaluation of this hardware scheme using a simulator and standard benchmarks shows that it captures the benefits of conflict-serializability. Applications with long transactions and high contention benefit the most, abort rates are reduced up to 7.2 times and the performance is improved up to 66%. We argue that this improvement comes with little additional hardware complexity and requires no changes to the transactional programming model.