USC Async1 Chip - 1.45 GHz 64-bit Asynchronous Adder

1       General Description

The USC Async1 chip was fabricated in March 2004 using TSMC 0.25 um process. It comprises two test circuits:

§       A 64-bit asynchronous prefix adder with input generator and output sampler blocks. All blocks were fabricated to show the high performance of the proposed STFB (Single-Track Full-Buffer) standard cell design, which provides low latency and fast cycle time.

§       A Sequential Decoder implemented with QDI (Quasi-Delay Insensitive) standard cells (this circuit is under test).

2       STFB template

In the STFB cell template, shown below, the Right Completion Detection circuit (RCD) allows the cell to work only if the output channel (R) is empty, and the State Completion Detector circuit (SCD) removes the input data from the input channel (L) when it is no longer needed.  

Basic internal diagram of a STFB cell.

Async1 die photo (bigger picture).

3       STFB standard cell library

The STFB cells were created using Cadence Custom IC Design and Synopsys tools (see publications below), and were made freely available through MOSIS Educational Program. The figure below illustrates the library creation process.

4       STFB block design

Using the STFB library, the circuit blocks described below were designed as follows:

5       INPUTGEN block

With the STFB library, the input pattern generator for the adder was designed as shown in the block diagram below. The INPUTGEN circuit allows 64-bit operands and 1-bit carries to be loaded to 9-stage rings that will continuously feed the adder inputs.

6       The 9-stage ring

The INPUTGEN block has 129 rings with 9-stages each (9 STFB cells in a loop) as show in the block diagram below. These rings can load up to 7 different bits that will be continuously duplicated and send to the adder inputs. All the 129 rings allow us to load the up to 7 sets of numbers we want to add (each set with 64-bit for each A operand, 64-bit for each B operand and 1-bit for each carry in).

7       ADDER64 block

The figure below represents an 8-bit prefix adder implemented with STFB cells. The thin arrows are dual-rail channels (2 wires), while the thick arrows are 1-of-3 channels (3 wires). The 64-bit version implemented in our design is an extension of the diagram below.

8       SAMPLER block

Due to the high-performance of the STFB cells, it is necessary to sample the output results in order to avoid slow down the adder. The circuit shown in the diagram below allows us to select the sample rate and also multiplex the result in order to output one byte at a time.

9       Circuit layout

The layout of the three blocks were automatically generated and placed side-by-side as shown below. The three blocks have 260k transistors in 3.3 mm². The power grid (not shown) on top of the blocks was designed to supply the high current required when running at full throughput (28 pads were allocated for power supply).

10  Test board

To test the chip, an interface board, shown below, has been designed which connects to an FPGA evaluation board. The FPGA is a XILINX XCS2100 Spartan II on a Xess XSA prototyping board.  The software utilized to program the FPGA was ISE V.6 and the Xess package. Once programmed, the FPGA loads the STFB INPUTGEN block with the operands, sets the sample rate in the SAMPLER block, and run the chip by acknowledging all requests as they come out of the chip.

11  Test setup

An oscilloscope (Tektronix TDS210) was used to check the byte and carry acknowledge signals. One multimeter was used to measure the temperature on top of the package (40^oC!!), while another displays the on-chip voltage (2.5V). The current (2.26A) was measured by the power supply (Agilent ES610A). A 24-charniel logic analyzer (Link Instruments LA-2124) was used to capture the waveforms, which allows checking the initialization and operation of the demonstration chip. For some clips of the chip performance, go to the bottom of this page.

12  Performance

The figure below is the acknowledgment signals for the carry and for the eight bytes outputted by the SAMPLER block. In this example, since acknowledge frequency was 313kHz and the sample rate was set to 1:3971, the internal adder throughput was 1.24GHz. This is an impressive performance when considering the technology (0.25 um) and the fact that the layout was automatically generated using a conventional back-end flow, which resulted in a simple, fast and efficient design process that can be easily understood by synchronous designers.

13  Test results

The STFB circuits, fabricated in the Async1 chip, worked perfectly, and generated the expected results (very close to the Nanosim simulations) reaching up to at 1.45 GHz (with cooling). The following table shows the results of our tests at room temperature with a fan over the chip under test.

STFB circuits at full speed at room temperature with fan.

The graphics on the right show the throughput and power dissipation of sample #3 and #4 running at full speed. Chip #3 was tested at room temperature with and without fan. Chip #4 was tested with air at -25^oC blowing on top of the device. Notice that, since the STFB circuits are asynchronous, the performance automatically adjusts according to the supply voltage and temperature. Also, notice that, with cooling, the STFB circuits reached the impressive mark of 1.45 GHz.

14  Conclusion

STFB templates were proposed for high-speed area-efficient asynchronous non-linear pipeline design. A freely available STFB standard cell library using TSMC 0.25 um technology was generated and posted with MOSIS Educational Program. A complete STFB design with 260k transistors was successfully implemented and tested reaching 1.45 GHz.

The templates have higher throughput than the fastest known QDI templates and have less timing assumptions and lower latency than the most aggressive GasP templates. Consequently, for systems that are latency-critical, STFB templates may yield a significant performance advantage. It also offers a small cycle time that allows the STFB circuits to operate at very high throughputs with small distances between consecutive data tokens, resulting in smaller and faster circuits than their QDI alternatives.

15  Publications

For more information please visit the following links:

- Thesis:
Single-Track Asynchronous Pipeline Template, Marcos Ferretti, Ph.D. Thesis, University of Southern California, Jun, 2004.

- Papers:

• Single-Track Asynchronous Pipeline Templates using 1-OF-N Encoding, M. Ferretti and P. A. Beerel, DATE'02, Paris, France, March 2002.
• High Performance Asynchronous Design Using Single-Track Full-Buffer Standard Cells, M. Ferretti and P.A. Beerel, IEEE Journal of Solid-State Circuits, Vol. 41, No. 6, pp. 1444-1454, June 2006.

• High Performance Asynchronous ASIC Back-End Design Flow Using Single-Track Full-Buffer Standard Cells, M. Ferretti, R.O. Ozdag and P.A. Beerel, 10th Symposium on Asynchronous Circuits ASYNC, Herssonissos, Greece, April 2004.

16  Acknowledgements

This research has been partially supported by NSF Grant CCR-0086036 and gifts from TRW, Fulcrum Microsystems and the MOSIS Educational Program. Thanks to Jay Moon for his valuable help with the CAD tools, to Sachit Chandra for his help with the design flow and Sunan Tugsinavisut for many helpful discussions.

Nanosim and Hspice are trademarks of Synopsys, Inc. (Mountain View, CA). Dracula, Verilog, Virtuoso, Envisia and Silicon Ensemble are trademarks of Cadence Design Systems, Inc. (San Jose, CA). All other trademarks are proprietary of their respective owners.