Comp Arch the 4 th Epoch Looms Bob
Comp. Arch: the 4 th Epoch Looms Bob Colwell ISCA 2002 5/28/02 Colwell ISCA 2002
The 1 st Three Epochs 1. 1960’s: Mainframes Roamed The Earth ü 2. 1970’s: Experimentation, Minis, Supers ü 3. Birth of instruction set architecture “semantic gap”, caches, MP, langs, OSs ’ 80’s & ’ 90’s: Microprocessors Rule ü Perf is everything -> clock rate is everything, PC’s, www bob. colwell@attbi. com 5/28/02 2
The 4 th Epoch Dawns Technical Reasons n n Approaching Si limits Design complexity grows without bound Market Reasons n n No more “killer apps”, old PC good enough Volume growth flattening -> commoditization World Status n n Hackers, viruses, worms, IT warfare High tech economies wired to www bob. colwell@attbi. com 5/28/02 3
How to predict new epoch? History best predictor of future n Behavioral interviewing Glew’s ASC prophecy model w Accurate w Sequential w Conditional event + time + place trends if X costs $N, then trend Z will occur ISCA extrapolations My own guesses bob. colwell@attbi. com 5/28/02 4
“Accurate”…don’t count on it Available on many web sites The crystal ball with all its mystique and secrets is probably the best known method of divination for predicting the future. . . Crystal scrying is done in near to almost total darkness. In more ways than one bob. colwell@attbi. com 5/28/02 5
Computing, Then and Now 1 ISAs proliferate 30 years ISAs converging to few Implementation=MSI Implementation=VLSI Networking unknown www OS’s all do-it-yourself OS’s converging to few MP new & hard MP commonplace Computers calculate Computers communicate Big Business sets prices Volume pricing rules CPU speed = mem speed CPU speed >> mem speed bob. colwell@attbi. com 5/28/02 6
Computing, Then and Now 2 Languages abound Binary compatib=new idea System complexity = N No good CAD tools Language converge to few Binary compatib required Complexity = N*104 Lots of good CAD tools Next tech reasonably clear Unix is best OS Quantum. . Ga. N. . Molecular From “kids-in-new-toy-store” to …today. bob. colwell@attbi. com 5/28/02 7
29 Years of ISCA Extract session topics, all 29 yrs Tally across ISCAs Apply trending, extrapolate wildly Thanks Eugene Miya! bob. colwell@attbi. com 5/28/02 8
ISCA Topics Tallied Interesting… n n n MP & interconnect are among highest Cache, memory, CPU arch are high Dataflow? Logic prog? This chart does not reflect changes over time… bob. colwell@attbi. com 5/28/02 9
ISCA Topics vs Time Strong, sustained interest n Caches, memory, CPU arch, performance Early interest that waned n Dataflow, Logic programming, VLSI arch, pipelining (may be reviving!) Rising stars n ILP, Branching, MT, power-aware, emerging tech Never-were’s n Real-time, secure computing, embedded s… n o ti ic d re P l a ti n e u q e S e m o S r o F Now bob. colwell@attbi. com 5/28/02 10
ISCA Prediction 1: mop-up Mop-up operations will continue n Incremental improvements are lucrative w Enormous intellectual energy applied to extremely difficult targets (memory latency, low code parallelism) for small gains w Why? User path-of-least-resistance usually wins w Process tech improvements carry architecture n Some real gains will be posted w MT, CMP bob. colwell@attbi. com 5/28/02 11
ISCA Prediction 2: branch out Threats to cash cow will be rebuked n n Power-aware computing is a start Battery-life is even better w How to guarantee RT if CPU throttles on temp? w DVD movie players, audio players require this n Security w 1950’s -- TV ads told you not to leave car keys w Computer industry at that level of maturity bob. colwell@attbi. com 5/28/02 12
ISCA Prediction 3: leave home New opportunities arise n CMOS-based Moore’s Law will end w Whomever repeats this exercise in 30 yrs will have lived through it w Economics will stop MLaw before tech limits do n n Quantum, molecular, non-CMOS, nano, MEMS, Spintronics, optical ISCA shows some openness to these w At least, judging by the accepted papers bob. colwell@attbi. com 5/28/02 13
Non-ISCA Prediction 4: new opportunities Incremental progression leads to step-function changes E. g. , non-volatile storage capacity doubling every year n n This will lead to entirely new products Prediction: when people can carry everything about their own lives with them, they will w Every homework assignment, letters, old email, favorite articles, music, addresses, phone numbers w Carry downloadable user profile to all other devices w Music, movies…Hollywood, get over it bob. colwell@attbi. com 5/28/02 14
Prognostications Linear extrapolation: 2012, Pentium 8 n 12 GHz, 8 -way CMP, 4 -way MT, 100 MB L 4 cache on die, 20 GB/s optical I/O, 250 W, 2 TB hard drive What will users do with this? Sequential prophecy alternative A n What they do now. Somewhat faster. Commoditized. Sequential prophecy alternative B n They won’t want them. They’ll want affordable, mobile, ubiquitous, secure computing based on huge local nonvolatile storage with auto-backup Comp Arch must anticipate both alternatives! bob. colwell@attbi. com 5/28/02 15
So What Should You Work On? Can’t tell you what to design n n Combine my predictions with yours Try to intersect future at profitable spot Guidelines? Those I’ve got. bob. colwell@attbi. com 5/28/02 16
10 Hints for Chip Designers 1. Design systems, not chips Forces good questions about I/O, caching, SW, usage models, feature weighting per market 2. Design-for-100 M units is 100 times harder than design-for-1 M unit Know which one you’re targeting 3. Observe Golden Rule of Silicon: Who Ships Volume Wins Even the biggest Si companies forget this 4. If 100 apps, architect’s creation should bottleneck 100 different ways Balanced systems are the goal bob. colwell@attbi. com 5/28/02 17
An Unbalanced System bob. colwell@attbi. com 5/28/02 18
10 Hints for Chip Designers 5. Comp. Arch needs the FT community The more capable computers get, the more taken-forgranted they become, the more irritated users get if they fail, and the less they care about why 6. If you didn’t test it, it doesn’t work You are responsible for overall product Architects can prevent otherwise-unfindable bugs 7. Don’t Overemphasize the Quantifiable Performance not always more important than quality, reliability, simplicity This pendulum has swung too far bob. colwell@attbi. com 5/28/02 19
10 Hints for Chip Designers 8. Undefined flags don’t mean don’t care Size of affected user decides issue 9. Adding new instructions is hard Removing old obsolete ones impossible 10. Economics rule, not technology VHS beat Beta…x 86 outlived Alpha Don’t design until you understand why bob. colwell@attbi. com 5/28/02 20
And then there’s Blue Crystals Laundry detergent “New! Improved! Now with Blue Crystals!” Microprocessors now in that same consumer space Pentium MMX was marketed accordingly Remember dancing bunny suits? Slot 1 hologram vs Celeron bob. colwell@attbi. com 5/28/02 Photos courtesy Intel Corp. 21
We hate Blue Crystals, but… Need to balance A. B. Urge to do whatever sells Urge to innovate ect t i h c , ar Intel, 2 n o t 0 Hin aire at nner ‘ n i n Gle aordin ard W r ext es Aw k Wil Both legitimate concerns Sea change in our industry ~1995 bob. colwell@attbi. com 5/28/02 y unn b g cin Dan on s per Photos courtesy Intel Corp. 22
A Ten-Year Plan Two Tasks 1. We must fix what we’ve made Easier to use, more dependable Safe, secure 2. Catch the 4 th epoch wave Ruled by economics Driven by new implementation technologies bob. colwell@attbi. com 5/28/02 23
Example Fix: Mission Availability Is computing system fulfilling its mission? n n Web server’s job is to satisfy legitimate requests, not just pings/sec (Do. S attacks) Real world: human errors, malicious attacks, HW & SW failures Promising avenues n n Recovery-oriented computing Autonomic computing (IBM) bob. colwell@attbi. com 5/28/02 (UCB, Stanford, Mills College) 24
But There Be Dragons Here Observation: Oo. O machines easily restarted Opportunity! Justifiable n n P 6 retirement watchdog timer Watches for fwd progress beyond longest latency Arguable n Pentium 4 deadlock/livelock detectors w Scattered around core of u. Arch w Reason: perf tweaks cause rare lockups Real problem: complexity in pursuit of perf, compensated by more complexity in recovery bob. colwell@attbi. com 5/28/02 25
Half of what you know about computers will help you in the 4 th epoch. 4 th Epoch : 1 st 3 : : www : PC bob. colwell@attbi. com 5/28/02 26
- Slides: 26