An Interview with Josh Fisher | Inventing VLIW, Multiflow, Itanium, VLIW's Massive Success

TL;DR

VLIW pioneer Josh Fisher corrects the myth that VLIW failed, explaining its origins, Multiflow's collapse, Itanium's flaws, and its dominance in embedded computing.

Key Points

• 1.VLIW is not a failed technology — it ships in roughly 12–15 billion processors annually. Fisher estimates ~5 billion VLIW units sold per year, each often containing multiple processors, making it 45x more prevalent by unit count than x86, though mostly in inexpensive embedded chips.
• 2.Fisher invented VLIW while automating horizontal microcode scheduling at NYU's Courant Institute. He developed 'trace scheduling' to compile high-level code into wide instruction words, recognizing that hardware had to be co-designed with the compiler — not retrofitted afterward.
• 3.The Bulldog compiler, written as a Yale PhD thesis by John Ellis, won best computer science thesis worldwide. It was a plastic compiler that could generate code for any machine described to it, forming the technical foundation for Multiflow.
• 4.Multiflow built machines issuing 7, 14, or 28 operations simultaneously and consistently won on price-performance. The company raised $70 million total, used slower TTL circuits deliberately to ship faster than competitors like Cydrome, which used ECL and never sold a single unit.
• 5.Multiflow failed due to inability to raise a next financing round, not technical failure. Killer micros — single-chip processors from Sun and others — convinced investors a transformation was coming, making Multiflow's multi-board VLIWs seem unscalable regardless of their performance edge.
• 6.Key marketing mistakes included limited geographic presence and pricing too high early on. Fisher and co-founder John O'Donnell pushed to lower prices and ship on-site tech support with machines, but the CEO — from NCR's old-school minicomputer world — resisted, ceding Germany and Japan largely to competitor Convex.
• 7.Fisher joined HP Labs after Multiflow closed in early 1990 and immediately disliked the Itanium architecture. He found it over-complicated, violating his co-design principles, and felt it would be extremely hard to compile for, though publicly he promoted it enthusiastically given what it meant for VLIW's credibility.
• 8.Itanium faced a structural problem: by the late 1990s, recompiling for a new ISA was no longer practical. The massive x86 application base, immature binary translation technology, and Intel's apparent lack of internal commitment all compounded the architecture's inherent complexity issues.
• 9.Fisher pivoted to embedded computing at HP Labs Cambridge, partnering with STMicroelectronics on the ST241 chip family. Sold into cable set-top boxes, cell base stations, and HP printers, the ST241 became a major commercial success and exemplified VLIW's embedded strengths: low power, small die size, and predictable timing.
• 10.VLIW dominates embedded because its static scheduling delivers low power, low heat, and predictable timing. Companies like Cadence Tensilica and CEVA ship billions of ~8-issue VLIW processors annually into IoT devices, smartphones (Bluetooth chips), storage, cars, and Google's TPUs, which switched to VLIW for training Gemini.
• 11.Fisher nearly launched what would have been an early smartphone at HP around 2000, leveraging VLIW's media-processing advantage. Marketing executives under Carly Fiorina rejected the convergence concept, insisting consumers would buy separate single-purpose devices — a decision Fisher views as a major missed opportunity for HP.