Alpha microprocessors have been performance leaders since their introduction in 1992. The first generation 21064 and the later 21164 raised expectations for the newest generation-performance leadership was again a goal of the 21264 design team. Benchmark scores of 30+ SPECint95 and 58+ SPECfp95 offer convincing evidence thus far that the 21264 achieves this goal and will continue to set a high performance standard. A unique combination of high clock speeds and advanced microarchitectural techniques, including many forms of out-of-order and speculative execution, provide exceptional core computational performance in the 21264. The processor also features a high-bandwidth memory system that can quickly deliver data values to the execution core, providing robust performance for a wide range of applications, including those without cache locality. The advanced performance levels are attained while maintaining an installed application base. All Alpha generations are upward-compatible. Database, real-time visual computing, data mining, medical imaging, scientific/technical, and many other applications can utilize the outstanding performance available with the 21264.

http://home.eng.iastate.edu/~zzhang/cpre581/reading/imicro99-kessler-21264.pdf

The Alpha 21264 microprocessor

Increasing levels of microprocessor power dissipation call for new approaches at the architectural level that save energy by better matching of on-chip resources to application requirements. Selective cache ways provides the ability to disable a subset of the ways in a set associative cache during periods of modest cache activity, while the full cache may remain operational for more cache-intensive periods. Because this approach leverages the subarray partitioning that is already present for performance reasons, only minor changes to a conventional cache are required, and therefore, full-speed cache operation can be maintained. Furthermore, the tradeoff between performance and energy is flexible, and can be dynamically tailored to meet changing application and machine environmental conditions. We show that trading off a small performance degradation for energy savings can produce a significant reduction in cache energy dissipation using this approach.

Selective cache ways: on-demand cache resource allocation

This paper explores a novel way to incorporate hardware-programmable resources into a processor microarchitecture to improve the performance of general-purpose applications. Through a coupling of compile-time analysis routines and hardware synthesis tools, we automatically configure a given set of the hardware-programmable functional units (PFUs) and thus augment the base instruction set architecture so that it better meets the instruction set needs of each application. We refer to this new class of general-purpose computers as PRogrammable Instruction Set Computers (PRISC). Although similar in concept, the PRISC approach differs from dynamically programmable microcode because in PRISC we define entirely-new primitive datapath operations. In this paper, we concentrate on the microarchitectural design of the simplest form of PRISC—a RISC microprocessor with a single PFU that only evaluates combinational functions. We briefly discuss the operating system and the programming language compilation techniques that are needed to successfully build PRISC and, we present performance results from a proof-of-concept study. With the inclusion of a single 32-bit-wide PFU whose hardware cost is less than that of a 1 kilobyte SRAM, our study shows a 22% improvement in processor performance on the SPECint92 benchmarks.

/pdf/a-high-performance-microarchitecture-with-hardware-udjxip5i6e.pdf

A high-performance microarchitecture with hardware-programmable functional units

This paper describes a new design methodology to analyzethe on-chip power supply noise for high-performance microprocessors.Based on an integrated package-level andchip-level power bus model, and a simulated switching circuitmodel for each functional block, this methodology offersthe most complete and accurate analysis of Vdd distributionfor the entire chip. The analysis results not only providedesigners with the inductive ΔI noise and the resistive IRdrop data at the same time, but also allow designers to easilyidentify the hot spots on the chip and ΔV across the chip.Global and local optimization such as buffer sizing, powerbus sizing, and on-chip decoupling capacitor placement canthen be conducted to maximize the circuit performance andminimize the noise.

/pdf/power-supply-noise-analysis-methodology-for-deep-submicron-4ynt8clso1.pdf

Power supply noise analysis methodology for deep-submicron VLSI chip design

The 21264 is the third generation Alpha microprocessor from Compaq Computer (formerly Digital Equipment) Corporation. This microprocessor achieves the industry-leading performance levels of 30+ Specint95 and 50+ Specfp95. In addition to the aggressive 600 MHz cycle time in a 0.35 /spl mu/m CMOS process, there are also many architectural features that enable the outstanding performance level of the 21264. This paper discusses many of these architectural techniques, which include an out-of-order and speculative execution pipeline coupled with a high-performance memory system.

The Alpha 21264 microprocessor architecture

A RISC (reduced-instruction-set computer)-style microprocessor operating up to 200 MHz, implements a 64-b architecture that provides huge linear address space without bottlenecks that would impede highly concurrent implementations. Fully pipelined and capable of issuing two instructions per clock cycle, this implementation can execute up to 400 M operations per second. The chip includes an 8-kB I-cache, an 8-kB D-cache, and two associated translation buffers, a four-entry 32-B/entry write buffer, a pipelined 64-b integer execution unit with 32-entry register file, and a pipelined floating-point unit with an additional 32 registers. The pin interface includes integral support for an external secondary cache. The package is a 431-pin PGA with 140 pins dedicated to VDD/VSS. The chip is fabricated in 0.75- mu m n-well CMOS with three layers of metallization. The die measures 16.8*13.9 mm/sup 2/ and contains 1.68 M transistors. Power dissipation is 30 W from a 3.3-V supply at 200 MHz. >

http://www.hpl.hp.com/hpjournal/dtj/vol4num4/vol4num4art2.pdf

A 200-MHz 64-bit Dual-Issue CMOS Microprocessor.

A circuit is provided for using the input operands of a floating point addition or subtraction operation to detect the leading one or zero bit position in parallel with the arithmetic operation. This allows the alignment to be performed on the available result in the next cycle of the floating point operation and results in a significant performance advantage. The leading I/O detection is decoupled from the adder that is computing the result in parallel, eliminating the need for special circuitry to compute a carry-dependent adjustment signal. The single-bit fraction overflow that can result from leading I/O misprediction is corrected with existing circuitry during a later stage of computation.

Leading one/zero bit detector for floating point operation

A floating point multiply of two n-bit operands creams a 2n-bit result, but ordinarily only n-bit precision is needed, so rounding is performed. Some rounding algorithms require the knowledge of the presence of any "1" in the n-2 low-order bits of the 2n-bit result. The presence of such a "1", indicates the so-called "sticky bit" is set. The sticky bit is calculated in a path separate from the multiply operation, so the n-2 least significant sums need not be calculated. This saves time and circuitry in an array multiplier, for example. In an example method, the difference between n and the number of trailing zeros, "x", in one of the n-bit operands is detected, by transposing the operand and detecting the leading one. The other operand is right-shifted by a number of bit positions equal to this difference. A sticky bit is generated if any logic "1's" are in the low-order n-x-2 bits fight shifted out of the second operand.

Method and apparatus for controlling a rounding operation in a floating point multiplier circuit

A description is given of a uniformly pipelined, 50-MHz, 64-b floating-point arithmetic processor implemented in a 1.5- mu m (drawn) CMOS technology which performs single- and double-precision floating-point operations and integer multiplication as defined by a superminicomputer architecture standard. The chip is composed of an interface section and a five-segment execution core. The core insists of a divider, bypassed in all instruction except division, and four fully pipelined stages that are uniformly utilized in the execution of all instructions. The performance is summarized. First pass silicon has been functionally verified at 50 MHz with a set of over one million vectors. >

A 50 MHz uniformly pipelined 64 b floating-point arithmetic processor

A four-chip custom VLSI implementation of a 32-b computer comprised of a CPU, a secondary cache controller, a floating-point accelerator, and a clock generator is described. It operates at a cycle time of 28 ns and is compatible with an existing computer architecture. The chip set is fabricated in a 1.5- mu m n-well, double-layer-metal CMOS process and includes over 650000 transistors. The CPU is a six-level pipeline engine built around three semiautonomous pipes. These provide simultaneous instruction prefetch and decode, specifier decode and execution, memory management, and I/O access. The CPU averages nine cycles/instruction on typical benchmarks. Chip functionality is verified through test vectors at the pins, with stuck-at fault coverage greater than 95%, and complete control store and cache tests. Summaries of process characteristics and physical specifications are presented. >

Sridhar Samudrala

Papers

A 200-MHz 64-bit Dual-Issue CMOS Microprocessor.

Leading one/zero bit detector for floating point operation

Method and apparatus for controlling a rounding operation in a floating point multiplier circuit

A 50 MHz uniformly pipelined 64 b floating-point arithmetic processor

CMOS implementation of a 32 b computer