Alpha 21264A Microprocessor — Die-Shrunk EV67 for Higher Clock Speeds
The Alpha 21264A, known as EV67, took the proven out-of-order 21264 core and moved it to a smaller manufacturing process. The result was a faster, cooler and cheaper-to-build chip that powered the later generation of Alpha Processor, Inc. boards and modules.
Why a die shrink
The original Alpha 21264 was a large, power-hungry chip. Built in a 0.35-micron process, it pushed the limits of what that technology could do, and its die size kept manufacturing costs high. For a company like API, whose entire strategy depended on bringing Alpha down to mainstream price points, a cheaper and faster version of the same core was exactly what was needed.
A die shrink is the most conservative way to get there. Instead of redesigning the microarchitecture, the engineers moved the existing design to a finer process with smaller transistors and shorter wires. The core logic, pipeline and software behaviour stay essentially the same, so existing binaries, compilers and operating systems run without change, while the physical improvements deliver higher clock frequencies, lower supply voltage and a smaller die.
Architecture
Architecturally the EV67 is a 21264. It fetches four instructions per cycle, renames registers onto a large physical pool, issues up to six instructions per cycle out of order, and splits its integer execution units into two clusters with duplicated register files. The on-chip caches remain 64 KB for instructions and 64 KB for data, both two-way set associative, and the chip still uses an external backside L2 cache and the EV6 point-to-point system bus.
That continuity was intentional. The EV6 bus interface meant the 21264A could slot into existing chipset designs, including the Tsunami-class logic used by the UP2000 family, with firmware updates and module changes rather than completely new boards. For integrators it meant a straightforward upgrade path.
Instruction set additions
The EV67 did bring one visible change to software: support for the Alpha count extension, often called CIX. It added instructions for counting population (the number of set bits), leading zeros and trailing zeros in a register. These are simple operations but are expensive to emulate in software, and they are valuable in cryptography, compression, bitboard game engines, database bitmap indexes and certain scientific codes. Software could detect the extension at run time and use it where available.
Physical improvements
Moving to 0.25 micron shortened the critical paths through the pipeline, which is where most of the clock speed gain came from. The lower supply voltage cut dynamic power, so a 21264A running faster than an original 21264 could still produce less heat. The smaller die also improved yields per wafer, which mattered directly to API's pricing. On API's Slot B modules these gains translated into cooler, cheaper cartridges that could run at higher speed grades without exotic cooling.
Main characteristics
| Parameter | 21264 (EV6) | 21264A (EV67) |
|---|---|---|
| Process | 0.35-micron CMOS | 0.25-micron CMOS |
| Transistors | Roughly 15 million | Roughly 15 million (same core) |
| Clock speeds | Roughly 450-667 MHz | Roughly 600-833 MHz across speed grades |
| Die size | Large, around 300 mm² | Substantially smaller |
| Core voltage / power | Higher | Lower voltage, less heat per MHz |
| Fetch / issue | 4 / up to 6 | 4 / up to 6 |
| L1 cache | 64 KB I + 64 KB D | 64 KB I + 64 KB D |
| L2 cache | External backside | External backside, typically 2-8 MB |
| System bus | EV6 | EV6 |
| ISA extensions | BWX, MVI, FIX | BWX, MVI, FIX, CIX |
Clock speeds varied across manufacturers and speed bins, and published figures from the period differ. The overall pattern is clear, though: the 21264A extended the practical range of the core well past what the original 0.35-micron part could reach, while bringing power down to more manageable levels. Samsung was a major manufacturer, and IBM also fabricated Alpha processors for Compaq around this time.
Target workloads
The 21264A served the same markets as its predecessor, but the higher clocks and lower power opened some new doors:
- Dense rackmount servers — lower heat output made it more practical to pack multiple processors into 1U to 3U enclosures such as the PowerRAC Chassis 320.
- Linux compute clusters — more floating-point throughput per node and per watt.
- Technical workstations — faster turnaround on simulation, CAD and rendering.
- Internet infrastructure — 64-bit servers for ISPs and hosting companies.
Operating system and software support
Because the 21264A is binary compatible with the 21264, operating system support carried over directly. On API boards, Linux was the primary platform, with the Alpha ports of Red Hat, SuSE and Debian all able to run on EV67 systems. Kernel and compiler support for CIX followed as toolchains were updated. Compaq's Alpha compilers and math libraries for Linux remained the preferred choice for performance-critical scientific work.
On Compaq systems the EV67 ran Tru64 UNIX and OpenVMS. Windows NT for Alpha had been cancelled by the time the 21264A reached volume, so NT support was essentially irrelevant for EV67 systems, and Windows 2000 never shipped for Alpha. NetBSD and OpenBSD continued to support Alpha platforms built with the chip.
Position in the API line
The 21264A powered API's later boards, including the UP2000+ and the UP1100, and the company's dense server designs. It sat above the original 21264 in API's processor range and was the part most integrators chose for new designs from around 2000 onward. Later Alpha development moved to the EV68 on an even finer process, but that chip came mostly through Compaq rather than API. For context see the microprocessor overview and the company's news archive.