Alpha Microprocessors — The 21164 and 21264 Families

Alpha Processor, Inc. built its business on two generations of the Alpha 64-bit RISC architecture: the in-order, very high clock 21164 and the out-of-order 21264. This page explains how the families differ, which API boards used which chip, and why the 21264 in particular became a reference point for high-performance design.

A clean 64-bit architecture from the start

Alpha was designed at Digital Equipment Corporation in the early 1990s as a pure 64-bit architecture with no 32-bit legacy mode to carry around. The instruction set was deliberately simple: fixed-length 32-bit instructions, 32 integer and 32 floating-point registers, a load/store model, and no condition codes or branch delay slots. The designers explicitly aimed for an architecture that could scale roughly a thousandfold in performance over its lifetime, through a mix of higher clock rates, wider issue and multiprocessing.

That minimalism had practical consequences. With little decoding complexity to worry about, Alpha design teams could push clock speeds harder than their contemporaries, and for most of the 1990s an Alpha chip held the title of fastest shipping microprocessor on integer and floating-point benchmarks. When API was formed in 1998, its job was to take that technical lead out of expensive workstations and into motherboards and systems that integrators, Linux users and technical customers could actually buy.

The 21164 family: EV5 and EV56

The 21164, known internally as EV5, was a four-way superscalar, in-order design. It was notable for being one of the first commodity microprocessors to put a second-level cache on the die: alongside 8 KB instruction and 8 KB data first-level caches it carried a 96 KB on-chip unified secondary cache, with support for an optional external third-level board cache. The original EV5 was built in a 0.5-micron process and ran at roughly 266 to 333 MHz.

The EV56 shrink moved the design to 0.35 micron and added the byte/word extension (BWX) instructions, which made it much easier to handle 8-bit and 16-bit data efficiently, an important fix for device drivers and for software ported from other platforms. EV56 parts reached clock speeds of up to about 600 MHz. Lower-cost variants, the 21164PC (PCA56 and PCA57), dropped the on-chip secondary cache and added motion-video instructions for multimedia work.

For API, the 21164 was primarily a legacy and support product. Many customers arriving at the company in 1998 and 1999 already ran 21164-based systems under Tru64 UNIX, Windows NT or Linux, and API continued to supply information and firmware guidance for those machines while steering new designs toward the 21264.

The 21264 family: EV6 and EV67

The 21264, or EV6, was a much more ambitious design. It introduced out-of-order execution to Alpha, with a large instruction window, register renaming, and the ability to issue up to four integer and two floating-point instructions per cycle. Its tournament branch predictor, which combined local and global history, was widely studied and imitated. First-level caches grew to 64 KB for instructions and 64 KB for data, both two-way set associative, and the chip drove an external backside cache of several megabytes through a dedicated interface.

Equally important was the system bus. EV6 used a point-to-point, source-synchronous, double-data-rate bus rather than a shared front-side bus. That design was later licensed by AMD for the Athlon processor, which is why API's first ATX board could use an AMD-derived northbridge. You can read more in the dedicated 21264 processor page.

The 21264A, or EV67, moved the core to a 0.25-micron process, lifting clock speeds into the 667 to 750 MHz range and beyond, and added count instructions (CIX) such as population count and leading/trailing zero count. Apart from higher performance and lower power per MHz, it was largely a drop-in successor, which let API refresh its boards without redesigning them from scratch. See the 21264A processor page for details.

Generation comparison

Figures are approximate and reflect the parts most commonly seen in API-era systems; speed grades varied by fab and by year.

Which API boards used which chip

API's own board line was built entirely around the 21264 family. The 21164 appeared in older Digital, Compaq and third-party systems that API supported but did not design.

• UP1000 — single-processor ATX board for the 21264 on a Slot B processor module, aimed at Linux workstations and cost-sensitive cluster nodes.
• UP2000 — dual-processor board for the 21264 and 21264A, with more memory capacity and I/O bandwidth for servers and technical computing.
• Later refreshes — updated single- and dual-processor boards and the CS20 dense server adopted faster 21264A and successor parts as they became available from the fabs.
• Legacy 21164 systems — supported through firmware guidance and Linux information rather than new hardware.

Choosing between them at the time

The rule of thumb API gave integrators was simple: if the workload was floating-point heavy, memory-bandwidth bound, or needed large address spaces, the 21264 was the processor to buy, and the dual-processor boards made most sense for anything shared by more than one user. The 21164 remained adequate for existing installations but offered no upgrade path onto API's newer platforms because of the completely different bus.

Manufacturing and supply

One of the reasons API existed at all was to diversify Alpha supply. Samsung, a founding investor, manufactured Alpha processors alongside Digital and later Compaq, and API could source chips and modules through those relationships. That second source mattered to board makers and integrators who were wary of committing to a platform with a single fab behind it.

Further reading

For the company context behind these products, see the company overview. Board installation, firmware and memory advice is collected on the technical support page.