EV7, EV8 and the End of the Alpha Roadmap
Alpha did not fade because it ran out of ideas. Its last shipping design, the 21364, and its cancelled successor, the 21464, were arguably the most forward-looking processors of their day. What ended the roadmap was a business decision in 2001, and the ideas went on to shape much of what came after.
Where the roadmap stood in 2000
By 2000 the 21264 and its die-shrunk variants were shipping in everything from API's UP-series boards to Compaq's largest AlphaServers. The roadmap beyond it had two main steps. EV7, the 21364, would reuse a refined EV6-class core but surround it with a radically different system architecture. EV8, the 21464, would be an entirely new, much wider core with simultaneous multithreading. Alpha Processor, Inc. and Samsung were part of the broader ecosystem around this roadmap, and the December 1999 strategy announcement between Compaq, API and Samsung framed Alpha as a platform with a long future.
EV7: the system on the chip
The 21364 was first described publicly in the late 1990s but did not ship in volume until early 2003, in the AlphaServer GS1280 and later the ES47 and ES80. Its core was essentially an EV68, so single-thread performance at a given clock was familiar. The novelty was everything around the core.
- Large on-die L2 cache. Roughly 1.75 MB, on the chip rather than on the board, removing the external B-cache that earlier Alphas depended on.
- Integrated memory controllers. Two Direct Rambus controllers on die, giving each processor its own local memory with bandwidth on the order of 12 GB/s.
- Integrated router. Four inter-processor links, conventionally labelled north, south, east and west, plus an I/O port, so processors could be wired directly into a two-dimensional torus with no external switch or northbridge.
The consequence was that memory bandwidth and interconnect capacity grew with every processor added. A 64-processor GS1280 was a mesh of processors, each with local memory, routing coherence traffic for its neighbours. Latency to remote memory rose with hop count, but there was no single shared bus or crossbar to saturate.
| Feature | 21264 (EV6/EV68) | 21364 (EV7) | 21464 (EV8, planned) |
|---|---|---|---|
| Issue width | 4 integer + 2 FP | Same core family | 8-wide |
| Threads per core | 1 | 1 | 4 (SMT) |
| L2 cache | External B-cache | ~1.75 MB on die | ~3 MB on die (planned) |
| Memory controller | In chipset | On die, Rambus | On die |
| Multiprocessor links | Via chipset | On-die router, 4 links | On-die router |
| Transistors | ~15 million | ~150 million | ~250 million (planned) |
| Clock | Up to ~1.25 GHz (late EV68) | ~0.8–1.3 GHz | Target ~1.2–2 GHz |
EV8: an ambitious multithreaded core
If EV7 was about the system, EV8 was about the core. The 21464 was designed to fetch and issue eight instructions per cycle, with a very large register file and deep out-of-order window. On its own, such a wide machine would struggle to find enough independent instructions in a single thread. The design team's answer was simultaneous multithreading: four hardware threads sharing one core, with instructions from different threads issued in the same cycle to fill otherwise idle execution slots.
SMT had been explored in academic research during the 1990s, with Digital engineers collaborating on some of that work. EV8 would have been one of the first high-end commercial processors to implement it aggressively. Combined with EV7's integrated memory controller and router, it would have given Compaq a server processor with very high throughput per socket and near-linear memory scaling.
June 2001: the Itanium decision
In June 2001 Compaq announced that it would standardise its high-end servers on Intel's Itanium architecture over the following few years, and that Alpha development would wind down after EV7. As part of the arrangement, Intel took on Alpha intellectual property and a substantial number of Alpha engineers. EV8 was cancelled. EV7 would still ship, along with a modest speed bump, and existing customers would be supported for years.
The reasons were largely economic. Compaq was carrying the full cost of processor design, fabrication relationships and compilers for a platform with a relatively small installed base, while Intel was promising an industry-standard 64-bit architecture with enormous volume behind it. Keeping two high-end architectures, Alpha and Itanium, alive in the same company made little sense to Compaq's management, and HP, which Compaq merged with in 2002, was already committed to Itanium as the successor to PA-RISC.
For API, the decision confirmed a shift already under way. The company had been moving toward networking and HyperTransport silicon through API NetWorks, and the Alpha board business had limited prospects without a roadmap behind it.
What happened to the engineers
Alpha alumni scattered widely, and their influence is visible across the industry:
- Intel. Many EV8 engineers joined Intel's server processor efforts. Later Itanium generations, and Intel's move to an integrated memory controller and a point-to-point interconnect with QuickPath in 2008, bear a clear family resemblance to EV7.
- AMD. Several Digital engineers had already moved to AMD in the late 1990s, contributing to the Athlon and later the K8 architecture. Opteron's integrated memory controller and HyperTransport links echo the same design philosophy.
- Startups and elsewhere. Alpha veterans turned up at low-power and embedded design houses, networking silicon vendors and, eventually, at mobile processor teams, where the Digital tradition of aggressive custom circuit design found new applications.
It is speculative to draw a straight line from EV8 to any particular later product. Intel's Hyper-Threading, for instance, shipped on the Pentium 4 in 2002 and had its own development history. But the concentration of SMT, integrated memory and router expertise that Compaq released in 2001 clearly raised the baseline for everyone who hired from it.
The long wind-down
EV7 systems sold steadily to customers running OpenVMS and Tru64, and the final EV7z parts appeared around 2004. HP continued to sell AlphaServers until roughly 2007, with support contracts running long after. OpenVMS was eventually ported to Itanium and, later still, to x86-64, which means code first compiled for Alpha is still running in production in 2026 on processors that never heard of the instruction set.
What EV7 and EV8 teach us
Looking back, the 21364 is the more influential of the two. Within a decade of its launch, every mass-market server processor had an integrated memory controller and point-to-point links between sockets. EV8's SMT is now standard in server cores too, although rarely at four threads on such a wide machine. The Alpha team got the architecture right early; what it could not do was fund the next decade of development on its own. That tension between technical leadership and economic scale is the recurring theme of the whole Alpha story, from DEC to API.
Further reading on this archive
- Why the Alpha 21264 still matters covers the core that EV7 built on.
- Lessons from the Alpha ISA examines the instruction set design choices.
- From LDT to HyperTransport follows the point-to-point interconnect story on the AMD side.
- Microprocessors lists API's archived processor product pages.