A 1024-core 70 GFLOP/W Floating Point Manycore Microprocessor

This paper describes the implementation of a software programmable floating point multicore architecture scalable to thousands of cores on a single die. A 1024 core implementation at 28nm occupies less 128mm and has a simulated energy efficiency of 70 GFLOP/Watt with a peak performance of 1.4 TFLOP. The aggressive claims are supported by a 65nm silicon proven 16-core version of the same design with measured efficiency of 35 GFLOPS/Watt. Architecture Overview The multi-core architecture proposed in this work was designed to accelerate signal processing kernels requiring floating point math, such as large FFTs and matrix inversions. Examples of embedded applications requiring floating point math include: synthetic aperture radar, ultra sound, cellular antenna beam forming, and graphics processing. Figure 1 shows a block diagram of the architecture, with 16 processor tiles arranged as a 4 x 4 array. Figure 1: Multi-core Floating Point Accelerator Architecture The tiles are connected through a 2D mesh network. Each processor tile contains a full routing cross-bar, a custom dual issue floating point RISC CPU, a DMA engine, and 32KB of multi-bank SRAM. All cores are ANSI-C programmable and share a single unified 32 bit flat address map. The processor cores can be programmed and run completely independently of each other or can work together to solve larger problems. An important architectural decision was the replacement of the traditional power hungry cache hierarchy with a distributed flat memory model that offers a total memory bandwidth of 32GB/s per processor core. The high bandwidth memory architecture and the flat unprotected 32 bit memory map lets up to 4096 cores communicate with each other directly with zero startup communication cost. Network-On-Chip The performance of a Network-On-Chip depends on a number of different factors such as: network topology, routing algorithms, packet strategy, buffer sizes, flow control, and quality of service support [6]. The proposed mesh NOC, shown in Figure 2, takes advantage of spatial locality and an abundance of short point-to-point on chip wires to send a source address, destination address, and data in parallel on every clock cycle. Figure 2: Network-On-Chip Architecture The address inefficiency overhead of sending an address on every transaction was compensated for by a significantly simpler NoC router design and smaller FIFOs. On write transactions 32GB of data can flow into and out of each routing node on every clock cycle. The mesh throughput is balanced with the load/store throughput of the core, allowing the processor core to store data from its register file directly into adjacent cores memory without stalling the CPU pipeline. Round robin arbitration at each crossbar node ensures fairness in bandwidth allocation and together with the single cycle transaction design guarantees that the NoC is free of deadlocks. The effectiveness of the Network-On-Chip was tested in implementing multicore versions of 1024 point FFT and variable size matrix multiplication routine (SGEMM). Chip Implementation A chip product based on the above proposed architecture was implemented in a 65nm triple-Vt high speed CMOS process, and contains 40 million transistors, staggered pad-ring wire bonding, and is packaged in a 324 ball 15x15mm BGA package. Figure 3 shows the silicon evaluation platform for the 65nm chip. The evaluation platform hooks up to a standard GNU debugger based tool chain running on a Linux distribution. The chip daughter card is located at the right side of the picture with two external power supplies, a 1V core supply and a 2.5V IO supply. Figure 3: 65nm Silicon Evaluation Platform Comparison to State of the Art Table 1 compares this work to previously published work, demonstrating the merits of this multicore architecture approach for low power floating point applications. Table2 demonstrates the transistor efficiency advantage of this NoC compared to previous publications. [2] [3] [4] [5] This work Process (nm) 65 90 45 45 65 Frequency (MHz) 3130 85