Design of Special Function Units in Modern Microprocessors
MetadataShow full item record
Today’s computing systems demand high performance for applications such as cloud computing, web-based search engines, network applications, and social media tasks. Such software applications involve an extensive use of hashing and arithmetic operations in their computation. In this thesis, we explore the use of new special function units (SFUs) for modern microprocessors, to accelerate such workloads. First, we design an SFU for hashing. Hashing can reduce the complexity of search and lookup from O(p) to O(p/n), where n bins are used and p items are being processed. In modern microprocessors, hashing is done in software. In our work, we propose a novel hardware hash unit design for use in modern microprocessors. Since the hash unit is designed at the hardware level, several advantages are obtained by our approach. First, a hardware-based hash unit executes a single hash instruction to perform a hash operation. In a software-based hashing in modern microprocessors, a hash operation is compiled into multiple instructions, thereby degrading performance. Second, software-based hashing stores hash data in a DRAM (also, hash operation entries can be stored in one of the cache levels). In a hardware-based hash unit, hash data is stored in a dedicated memory module (a hardware hash table), which improves performance. Third, today’s operating systems execute multiple applications (processes) in parallel, which entail high memory utilization. Hence the operating systems require many context switching between different processes, which results in many cache misses. In a hardware-based hash unit, the cache misses is reduced significantly using the dedicated memory module (hash table). These advantages all reduce the power consumption and increase the overall system performance significantly with a minimal increase in the microprocessor’s die area. We evaluate our hardware-based hash unit and compare its performance with software-based hashing. We start by evaluating our design approach at the micro-architecture level in terms of system performance. After that, we design our approach at the circuit level design to obtain the area overhead. Also, we analyze our design’s power and delay for each hash operation. These results are compared with a traditional hashing implementation. Then, we present an FPGA-based coprocessor for hash unit acceleration, applied to a virus checking application. Second, we present an SFU to speed up arithmetic operations. We call this arithmetic SFU a programmable arithmetic unit (PAU). In modern microprocessors, applications that require heavy arithmetic computations are done in software. To improve the performance for such computations, we present a programmable arithmetic unit (PAU), a partially reconfigurable methodology for arithmetic applications. The PAU consists of a set of IP blocks connected to a reconfigurable FPGA controller via a fast mesh-based interconnect. The IP blocks in the PAU can be any IP block such as adders, subtractors, multipliers, comparators and sign extension units. The PAU can have one or more copies of the same IP block (for example, 5 adders and 7 multipliers). The FPGA controller is an on-chip FPGA-based reconfigurable control fabric. The FPGA controller enables different arithmetic applications to be embedded on the PAU. The FPGA controller is programmed for different applications. The reconfigurable logic is based on a LUT-based design like a traditional FPGA. The FPGA controller and the IP blocks in the PAU communicate via a high speed ring data fabric. In our work, we use the PAU as an SFU in modern microprocessors. We compare the performance of different hardware-based arithmetic applications in the PAU with software-based implementations in modern microprocessors.
Subjectspecial function units
programmable arithmetic unit
Fairouz, Abbas A E A (2019). Design of Special Function Units in Modern Microprocessors. Doctoral dissertation, Texas A&M University. Available electronically from