Migration of Cell Broadband Engine from 65nm SOI to 45nm SOI

Abstract

We describe the challenges of migrating the Cell Broadband Engine TM (Cell BE) [1-2] design from a 65nm SOI [3] to a 45nm twin-well CMOS technology on SOI with low-κ dielectric (κ = 2.4) and 10 copper metal layers [4]. The technology offers dual-gateoxide thicknesses of 1.16nm and 2.5nm for 1.0V and 1.5V nominal power supply, respectively. Thicker oxide devices are used in analog circuits. To guarantee the proper operation of existing gaming software, the exact cycle-by-cycle machine behavior, including operating frequency, must be preserved. We set the focus of design migration to four goals: 1) automated design migration where possible, 2) 30% power reduction, 3) 30% area reduction, and 4) design for manufacturability (DFM) improvement. With the design rules across technologies being relatively compatible, we take advantage of automated migration for the bulk of Cell BE circuit blocks. Circuits are manually fine tuned for timing, noise tolerance, and design robustness after the initial automatic migration. We take a different approach with memory and analog circuits. Analog circuits do not scale well due to the required area for decoupling capacitance. The I/O area, especially the area for C4 bumps, dictates chip dimensions since the same number of I/O signals is required and the C4 pitch does not scale from the previous technology.Since digital circuits occupy the bulk of chip area, it is crucial to migrate them effectively. The original digital circuits in 65nm consist of 3 types of components: parameterized cells, common leaf cells (flip-flop and local clock buffer), and custom cells. The migration of parameterized cells is done through software. A tool called Migration Assistant Shape Handler (MASH) [5] is applied to the common leaf cells first. MASH first shrinks the shapes according to the scale factor between technologies. Second, MASH corrects as many design rule violations in 45nm as possible with minimum layout perturbations and the remaining violations are repaired manually. The pin locations for these cells are fixed and scaled only in size. Metal blockage changes are minimized to reduce the effect on higher design levels. Hierarchical migration is performed in 2 phases. The first phase includes placing scaled leaf cells at scaled coordinates and scaling any remaining shapes. The second phase applies MASH to remove design rule violations. Using this design migration methodology, we shrink the chip size by 34% with respect to 65nm. Figure 4.3.1 shows dimensions of Cell BE and its major partitions in 3 technologies.We take advantage of the automated approach as much as possible by applying it to smaller memory array blocks and then tuning circuits manually. As the SRAM cell size (0.404mm 2 ) shrinks in 45nm from 65nm (0.7mm 2 ), we address the cell stability concern due to process variability [6] by using a separate array power supply (V CS ). Lowering the main power supply (V DD ) is critical for reducing the chip power consumption. However, we cannot lower V CS by the same amount as V DD due to SRAM...

A 45nm 1.3GHz 16.7 double-precision GFLOPS/W RISC-V processor with vector accelerators

“…Conservatively assuming that a double-precision floatingpoint matrix-multiplication kernel can achieve peak floatingpoint performance of 204.8 GFLOPS, Blue Gene/Q achieves 6.9 GFLOPS/W-55% of our energy efficiency at 0.8 V. The Cell processor [10] optimized for double-precision achieves 108.8 GFLOPS while dissipating 75 W in 65 nm SOI [11]. The single-precision optimized Cell processor reported a 40% power reduction from 65 nm to 45 nm [12]. Conservatively assuming peak floating-point performance and half of the power going to the cores, the Cell processor would achieve 4.8 GFLOPS/W at 0.8 V in 45 nm SOI-38% of our energy efficiency at the same voltage.…”

Software Defined Radio: Challenges and Opportunities

“…The peak processing power is quoted at an impressive > 256 GFlops [75], however the corresponding power consumption is not stated, in a chart comparison with other processors [67] it is located at approximately 50W. Although lower-power versions have been released [76], [77], these presumably still have a somewhat high power consumption when considering battery powered applications.…”

Nanometer MOSFET Effects on the Minimum-Energy Point of Sub-45nm Subthreshold Logic---Mitigation at Technology and Circuit Levels

“…Partially depleted SOI has been known for years as a very interesting technology for high-performance applications thanks to strong reduction of parasitic capacitances [Takahashi et al 2008]. Fully Depleted (FD) SOI technology is another flavor of SOI technology which features an ultra-thin active Silicon film [Fenouillet-Beranger 2007;Weber 2008], as shown in Figure 9.…”