This paper describes a high-performance, low-power, mixed-signal Digital Signal Processor (DSP) and its application to a single-chip Vector Sum Excited Linear Prediction (VSELP) speech codec. The DSP consists of a 25MIPS, 24bit floating point core-DSP; 13bit oversampling ADC/DAC; 6 KW data ROM; and 3.5 KW data RAM. The total transistor counts of the DSP is 1.3 million and its chip size is 11.0 mm15.8 mm. Unique design techniques are used to reduce the power dissipation, such as the programmable machine cycle time control and the clock supply control scheme in the core-DSP, the address detection for on-chip data ROM/RAM, and the shared-hardware design for digital filters of ADC and DAC. As an application of the DSP, the VSELP speech codec, which is the standard speech codec for the North American and Japanese digital cellular telephone system, has been implemented in a single-chip. Owing ti the salient architecture design and the program optimization techniques, sufficient quality was obtained in the codec at performance of 16.4 MIPS with low-power dissipation of 490 mW.

A fuzzy inference processor which performs fuzzy inference with 12-bit resolution input at 200 kFLIPS (Fuzzy Logical Inference Per Second) has been developed. To keep the cost performance, not parallel processing hardware but processor type hardware is employed. Dedicated membership function generators, rule instructions and modified add/divide algorithm are adopted to attain the performance. The membership function generators calculate a membership function value in less than a half clock cycle. Rule instructions calculate the grade of a rule by one instruction. Antecedent processing and consequent processing are pipelined by the modified add/divide algorithm. As a result, total inference time is significantly reduced. For example, in the case of typical inference (about 20 rules with 2 to 4 inputs and 1 output), the total inference needs approximately 100 clock cycles. Furthermore by adding a mechanism to calculate the variance and maximum grade of the final membership function, it is enabled to evaluate the inference reliability. The chip, fabricated by 1 µm CMOS technology, contains 86k transistors in a 7.56.7 mm die size. The chip operates at more than 20 MHz clock frequency at 5 V.

This paper details a unique VLSI design scheme which employs self-timed circuits. A 32-bit 50-MFLOPS data-driven microprocessor has been designed using a self-timed clocking scheme. This high performance data-driven microprocessor with sophisticated functions has been designed by a combination of several kinds of self-timed components. All functional blocks in the microprocessor are driven by self-timed clocks. The microprocessor integrates 700,000 devices in a 14.65 mm14.65 mm die area using double polysilicon double metal 0.8 µm CMOS technology.

Built-in self-test (BIST) has been applied to test an analog to digital converter (ADC) and a digital to analog converter (DAC) embedded in a DSP-core ASIC. The eight performance characteristics of the ADC and the DAC designed in accordance with the ITU-T recommendations are measured using the BIST. Three of the eight characteristics - the attenuation/frequency distortion, the variation of gain with input level, and the signal-to-total distortion - have been evaluated and the measured results have shown good agreement with measured results by conventional tests. In the BIST operation, the DSP-core generates input stimulus and analyzes output response by control of the self-test program, The sizes of the self-test program and coefficient data are 822 words of the IROM and 384 words of the data ROM, respectively. This area overhead is less than 0.5% of total chip area. Test-time by the BIST is reduced to approximately 3.2 seconds, which is one-tenth that of conventional testing. The mixed-signal DSP-core ASIC is testable with only logic test equipment, and as a result, test-cost - that is test investment and test-time - is reduced compared with conventional test methods.

A low power consumption 16-bit fixed point Digital Signal Processor (DSP) has been developed to realize a half-rate CODEC for the Personal Digital Cellular (PDC) system. Dual datapath architecture has been employed to execute multiply-accumulate (MAC) operations with a high degree of efficiency. With this architecture. 86.3% of total MAC operations in the Pitch Synchronous Innovation Code Excited Linear Prediction (PSI-CELP) program are executed in parallel, so that total instruction cycles are reduced by 23.1%. The area overhead for the dual datapath architecture is only 3.0% of the total area. Furthermore, in order to reduce power consumption, circuit design techniques are also extensively applied to RAMs. ROMs, and clock circuits, which consume the great majority of power. By reducing the number of precharging bit lines, a power reduction of 49.8% is achieved in RAMs, and above 40% in ROMs. By applying gated clock to clock lines, a power reduction of 5.0% is achieved in the DSP that performs the PSI-CELP algorithm. The DSP is fabricated in 0.5 µm single-poly, double-metal CMOS technology. The PSI-CELP algorithm for the PDC half-rate CODEC can operate at 22.5 MHz instruction frequency and 1.6 V supply voltage. resulting in a low-power consumption of 28 mW.

A built-in self-test (BIST) based on a signature-analysis (one of the data compression techniques) has been implemented in a 24 bit floating point digital signal processor (DSP). By using only a single pair of linear feedback shift registers (LFSR's) and 253 words of instruction of the DSP, 95% of the functional blocks are self-tested. The number of test patterns is 35 million. It takes only 2.6 seconds for the test at fc26.7 MHz. The overhead of the BIST hardware is about 2.0% of the die size. By comparing the pass rate in a conventional function test to the BIST, nearly the same fault coverage is obtained. This result shows that the BIST is effective for VLSI processors, such as DSPs. By improving this method, manufacturing go/no-go tests without expensive test equipment will be possible.

This paper describes the 10-bit 50 MS/s pipelined CMOS A/D Converter using a "reference feed-forward architecture." In this architecture, reference voltage generated in a reference generator block and residual voltage from a DA/subtractor block are fed to the next stage. The reference generator block and DA/subtractor block are constructed using resistive-load, low-gain differential amplifiers. The high-gain, high-speed amplifiers consuming much power are not used. Therefore, the power consumption of this ADC is reduced. The gain matching of the reference voltage with the internal signal range is achieved through the introduction of the reference generator block having the same characteristics as a DA/subtractor block. Each offset voltage of the differential amplifier in the reference generator block and the DA/subtractor block is canceled by the offset cancellation technique, individually. In addition, the front-end sample/hold circuit is eliminated to reduce power consumption. Because of the introduction of high-speed comparators based on the source follower and latch circuit into the first stage A/D subconverter, analog bandwidth is not degraded. This ADC has been fabricated in double-polysilicon, double-metal, 0.5µm CMOS technology, and it operates at 50 MS/s with a 300-mW (Vdd=3.0 V) power consumption. The differential linearity error of less than +/-1 LSB is obtained.

This paper proposes a dual port color palette SRAM using a single bit line cell. Since the single bit line cell consists of fewer bit lines and transistors than standard dual port cells, it is able to reduce the area. However, the cell has had a problem in writing a high level. The port swap architecture solves the problem without any special mechanism such as a boot strap. In the architecture, each of two bit lines is assigned to the read/write MPU port and the read only pixel port, respectively. When writing a low level, the MPU port uses pre-assigned bit line. On the other hand, when writing a high level, the MPU port uses the bit line assigned to the pixel port by a swap operation. During the swapping, the pixel port continues the read operation by using the bit line assigned to the MPU port. A color palette using this architecture is fabricated with a 0. 5 µm CMOS process technology. The memory cell size reduces by up to 43% compared with standard dual port cells. The color palette is able to supply the pixel data at 300 MHz at the supply voltage of 3.3 V. This speed is enough to support the practical highest resolution monitors in the world.