EV-4 (1992)

EV-4 (no code name) was the first public Alpha microprocessor. Dan Dobberpuhl was the project manager; Rich Witek was the chip architect and wrote the PALcode; Dick Sites was the overall Alpha architect; Jim Montano led the implementation team. EV-4, at 200Mhz, was the fastest microprocessor of its day, outstripping competitive designs by more than 2X.

EV-4 was implemented in DEC's 0.75u triple-metal CMOS4 process. EV-4 operated over a frequency range of 150Mhz to 200Mhz: the fastest microprocessor in the world.

EV-4 was shrunk into DEC's 0.5u CMOS5 process as EV-45. This version operated at frequencies up to 250Mhz and was also the fastest RISC processor of its day.

Name Number Size Transistors Comments
EV-4 660x546 1,681,999 EV-4 is the first Alpha microprocessor. It implements a complete 64b RISC processor on a single chip at unprecedented performance levels. Key features include:
  • Very high performance (200Mhz)
  • Fully-pipelined integer unit
  • Fully pipelined floating-point unit
  • Dual instruction issue (integer/floating point + load/store, integer + floating point)
  • 8KB on-chip instruction cache
  • 8KB on-chip data cache
  • 64 entry instruction TLB
  • 64 entry data TLB
  • Read-and-run, write-and-run, hit-under-miss memory operations

Power dissipation: 30W.

EV-4 shipped in late 1992 and supported the Alpha product line from workstations to mainframes. EV-45 shipped in 1994.

Personal Narrative

At end of the infamous Executive Committee meeting where Prism was cancelled, Ken Olsen turned to me and asked, very casually, if I would look into how to make VMS systems more competitive against RISC-based UNIX systems. This informal authorization became the basis of the Alpha Program.

Starting in July, 1988, a task team from DEC's chips, systems, and software groups met to discuss alternatives for faster VMS systems. The team was known as the Extended VAX, or EVAX, task force. Initial thinking focused on subsetting the VAX architecture further, to produce a "RISC-like" processor. Detailed analysis showed that this would not help performance on existing binaries, nor would it alleviate some of the principal performance bottlenecks of VAXen. Thereafter, the team focused on the question of whether VMS could be ported to a true RISC system in a reasonable amount of time, with some degree of binary compatibility.

Key breakthroughs came from Dick Sites and Nancy Kronenberg. Dick set out to "prove" that translating VAX binaries to RISC, with acceptable performance, was impossible. Instead, he demonstrated that it could be done, with very little hardware support. This allowed the nascent RISC architecture to dispense with a VAX compatibility mode. Nancy Kronenberg investigated the assumption that VMS, which was mostly written in VAX assembler, was too heavily tied to the VAX architecture and couldn't be ported. She concluded that VMS was actually reasonably well layered, that its dependencies on VAX architectural idiosynchosies were relatively few, and that VAX macrocode could be "compiled" to RISC code. This allowed VMS to be ported to a "hospitable" RISC architecture.

With these breakthroughs, the task team decided to recommend building a new family of 64b RISC systems that would run VMS and would achieve binary compatibility through translation. At the same time, the entire program was renamed Alpha, to "conceal" its focus on VMS. (Extended VAX lived on in the EV-xy naming of the Alpha chips, although the public explanation was that EV stood for "Electric Vlasic," honoring the Western Research Lab's famous Electric Pickle experiment.) Dick Sites and Rich Witek wrote the first draft (and all subsequent) drafts of the Alpha SRM in March, 1989. Dan Dobberpuhl's team started on chip development in June, 1989. Coordinated by Peter Conklin and other members of the Alpha Program Office, work also started in VMS, the languages group, software, and systems.

The Alpha chip team faced enormous challenges. Silicon was needed as quickly as possible to support the software development effort. At the same time, Alpha had to achieve breakthrough performance. Dan Dobberpuhl broke development into two phases. The first would create a "bare-bones" CPU chip and would be fabricated in CMOS-3. This internal chip, known as EV-3, threw everything deemed inessential overboard to achieve fast design time; for example, the first-level caches were implemented externally (in ECL!), and there was no floating point. The product chip, EV-4, would be fabricated in CMOS-4 and would include on-chip caches, the Rigel/NVAX floating point unit (with IEEE extensions), and a "friendly" pin interface. A team from DEC's Research Group, led by Chuck Thacker, signed up to design an EV-3-based system, known as the Alpha Development Unit (ADU). All the system groups focused on EV-4.

Despite a shortage of resources, the EV-3 chip team taped out in October, 1990 - just sixteen months after the project started. First wafers came back in 6 weeks (thanks to heroic efforts in Semiconductor Manufacturing) and were functional at probe. By the end of January, 1991, Ultrix (DEC's BSD-derived UNIX) was running on the ADU, and VMS had booted. EV-4 taped out in July, 1991, less than nine months after EV-3. It too was functional on first pass.

DEC showed both NVAX and EV-4 at the 1992 International Solid State Circuits Conference in February 1992. EV-4, with its unprecedented performance, new architecture, and novel circuit techniques, created a sensation. Following his presentation, Dan Dobberpuhl was mobbed like a rock star by excited, credulous, or awestruck chip designers from every company in the world. EV-4 single-handledly ignited the "megahertz wars" that would consume mainstream microprocessor development for the next decade. Alpha systems were shown publicly at DECWorld in the spring of 1992 and shipped for revenue in November, 1992.

Updated 24-Feb-2008 by Bob Supnik (simh AT trailing-edge DOT com - anti-spam encoded)