Presentation on theme: "Josh Fisher Bob Rau Award Talk. Except for a very short talk at Mateo 2012, I haven’t given a talk in six years. I haven’t even kept up with the field."— Presentation transcript:
Josh Fisher Bob Rau Award Talk
Except for a very short talk at Mateo 2012, I haven’t given a talk in six years. I haven’t even kept up with the field. A talk on The Lefty’s Advantage in Tennis (I could do that well!) wouldn’t work. So what to do...
AWARD: For the development of trace scheduling compilation and pioneering work in VLIW (Very Long Instruction Word) architectures. I ’ m very proud of having done all that, and am truly honored to be receiving this award. And so it is with great trepidation — — that I say I did something else I ’ m equally proud of. I bring it up because I want to give advice: I think it hurts our science that it isn ’ t the kind of thing people get awards for. I wish that when (other) people do this sort of thing, they would get recognized.
From John Hennessey’s Keynote/Review (1998) of The First 25 Years of ISCA.
Dynamic ILP: 1. The general-purpose supercomputers of the day, and the scalar units of the vector machines that followed. Static ILP: 2. Horizontal microcode (here at Micro, “ The Microprogramming Workshop ” !). 3. Array processors/bit slice processors (e.g. for custom graphics). In the s, there were three different communities doing ILP None acted like the others existed.
“ The paper did present an ingenious method of rearranging the linear structure of a calculation and of departing from a linear structure and carrying out operations in parallel. ” Discussion involving Alan Perlis, John Backus, Robert Floyd, Maurice Wilkes, and others . “ If this approach is carried further in the coming years, processors may contain tens or hundreds of independent arithmetic units. … Ultimately, the problem of efficiency will fall on the compilers and the compiler writers. ” [Stone, 1967]. “... stores must be inhibited while fetches are in progress, and vice versa. This feature makes it infeasible to extend this design to include parallelism on a very large scale. ” [Schwartz, 1966]. (The later “ MAXPAR ” experiments made the pessimism worse.) It was real ILP: CDC 6600 / IBM Stretch
It was one of the most popular research topics here. We were trying to automatically turn vertical (sequential) microcode into horizontal (instruction-level parallel) microcode. At Micro, We Worried About “Horizontal Microcode Compaction” There was some awareness of horizontal microcode in the Array Processor/Bit Slice crowd (Bob Rau in particular bridged that gap in the 1980s). I was lucky enough to be in the right place (Courant) at the right time (the 1970s) to see and name this commonality among subjects. There is Real Value In This Commonality!
Something Similar: The “Invention” of RISC, and the RISC vs. CISC Wars. The real topic was “instruction complexity”. Of course there were “ RISC ” architectures before IBM/Berkeley/Stanford! The subject that was invented was the debate about instruction complexity. Although people often talked about instruction complexity before, it became a subject, a concept, with Cocke/Patterson/Hennessey promoting RISC. They framed a debate.
Something Similar: The “Invention” of RISC, and the RISC vs. CISC Wars. The real topic was “instruction complexity”. Work like Clark/Emer shed real light on the subject — it was much easier to see its value once the area was in sharp relief. Much more research like that should have been promoted. We weren ’ t stuck with dumb market-based arguments only. “ Such and such is better because it won in the marketplace. ” Though there is always a lot of that. (Doing better than that is our job as researchers.)
a few words about… Multiflow Computer
Multiflow is well-known for having demonstrated the practicality of VLIW architectures and their compilers. In the early 1980s, nearly everyone was very skeptical about this. e.g. Bob Colwell: Then Josh walked into my office, on a visit to CMU. After an hour of earnest discourse, he had moved me from default skeptic to “if there’s something insurmountable here, it’s not obvious what it is. And he seems very reasonable, given that he’s nuts.” Multiflow’s Architectural Claim to Fame
VLIW in relationship to marketplace forces has been, and continues to be, a tumultuous ride, but no one ever again called it impractical. (Or me nuts. Well, maybe some people.) e.g. Ray Simar, TI, in 1997: “I remember looking at the idea and saying these guys were nuts,... I thought there was no way it would work in the real world.” But … “At the end of the day what we thought was ridiculous was the best solution,” Multiflow’s Architectural Claim to Fame
But there is a second thing the Multiflow should also be known for: When you consider code development as well as use: I claim that the Trace was the most exotic/novel/different processor ever to be used as an ordinary computer. You compiled and ran, and got good performance. Even though the computer might have 1,024-bit instructions. Over time, many bizarre architectures have been proposed, but none have been programmed and used as ordinary computers. It was true in 1987, and 25 years later, it’s still true. The way the market has gone, it may always be true. Multiflow’s Architectural Claim to Fame
Dave Papworth Bob Colwell Geoff Lowney Josh Fisher Multiflow Also Produced Four Fellows at Major Computer Companies
Notable Minisupercomputer Companies (Wikipedia) Ametek Alliant Computer Systems American Supercomputer (Mike Flynn) Astronautics (Jim Smith) BBN Convex Computer Culler Harris Culler Scientific Cydrome (Bob Rau) DEC (VAX 9000) Elxsi Encore Computer Evans & Sutherland Flexible Computer Floating Point Systems Guiltech/SAXPY (Rob Schreiber) HAL Computer Systems IBM (ES/9121 w/ Vector Facility) ICL (DAP) Kendall Square Research Key Laboratories MasPar Meiko Scientific Multiflow Computer (Josh Fisher) Myrias Prisma Parsytec Pyramid Technology Scientific Computer Systems Sequent Solbourne SUPRENUM Supertek Computers Thinking Machines Corporation
The Preface and Table of Contents are available at: MultiflowTheBook.com I know I’m hardly an unbiased reviewer,… …but I really recommend this book to people interested in a startup that sprang from this community. It is an amazing story.
Preface The Eureka Moment Yale VLIW John and John Three Guys and a Car Multiflow Computer Don Eckdahl The Engineering Team Building the Trace Product Introduction No Vectors Le Flic Selling the Trace Doldrums Europe The End Epilogue Josh’s Technical Appendix Sources and Further Reading Table of Contents
Lx is very much inspired by the Multiflow Trace
Lx resulted in STMicroelectronics’ ST200 Family Here are two Lx cores on an STMicroelectronics SOC. ST uses Lx in digital video SOCs. Guess how many Lx cores STM has sold?? (And HP has put a good many more in their printers.)
A fanatic is someone who can’t change his mind, and won’t change the subject. Attributed to Winston Churchill
Multiflow was in business Multiflow’s computers cost between about $350k-$750k. How many systems do you think were sold??