a theory of coding for chip to chip communica6on
play

A Theory of Coding for Chip- to-Chip Communica6on Amin - PowerPoint PPT Presentation

A Theory of Coding for Chip- to-Chip Communica6on Amin Shokrollahi and the engineering team of Kandou Dagstuhl - August 2016 The Problem Dagstuhl - August 2016


  1. A ¡Theory ¡of ¡Coding ¡for ¡Chip-­‑ to-­‑Chip ¡Communica6on Amin ¡Shokrollahi ¡ and ¡the ¡engineering ¡team ¡of ¡Kandou Dagstuhl - August 2016

  2. The ¡Problem Dagstuhl - August 2016

  3. Chip-­‑to-­‑Chip ¡Communica6on Chip 1 Chip 2 Communica)on ¡ wires Dagstuhl - August 2016

  4. Abundant…. Dagstuhl - August 2016

  5. Noise Chip 1 Chip 2 Noise ¡scales ¡badly ¡with ¡frequency ¡of ¡transmission: ¡ Example: ¡-­‑40dB ¡at ¡frequency ¡ f, ¡ -­‑90dB ¡at ¡2 f Signal ¡strength ¡ Signal ¡strength ¡ drops ¡to ¡~1% ¡ drops ¡to ¡ ¡~0.003% Dagstuhl - August 2016

  6. Power Supercomputers Data centers Internal traffic in Giga- 1E+12 4E+11 bits per second Power in Mega-Watts 20000 8000 Mul6ply ¡by ¡4 ¡in ¡every ¡genera6on ¡(~2 ¡years) ¡ Very ¡par6ally ¡offset ¡by ¡Moore’s ¡law Dagstuhl - August 2016

  7. Point ¡in ¡Case “While ¡CPU’s ¡doubled ¡performance ¡every ¡two ¡years, ¡ evolu6on ¡from ¡1 ¡GigE ¡to ¡10GigE ¡took ¡12 ¡years, ¡and ¡WAN ¡ routers ¡increased ¡throughput ¡only ¡4-­‑fold ¡during ¡the ¡same ¡ 6me ¡period.” ¡[6] Andy ¡Bechtolsheim, ¡2012 Dagstuhl - August 2016

  8. Capacity ~20x 450 Throughput per wire (Gbps) 400 350 300 250 200 150 100 50 2 4 6 8 10 12 14 16 18 Today Std dev of noise (mV) Dagstuhl - August 2016

  9. Channel ¡is ¡NOT ¡Similar ¡to…. Or ¡any ¡channel ¡with ¡a ¡lot ¡of ¡“random” ¡noise Dagstuhl - August 2016

  10. Noise Deterministic Random SSO Ref Xtlk EMI CM ISI Thermal Almost ¡all ¡noise ¡is ¡determinis6c ¡ but ¡resources ¡are ¡6ght Dagstuhl - August 2016

  11. Rule ¡of ¡Thousands Throughput Energy/bit Recovery ¡)me/bit Wireless Mbps nJ nano-­‑second Chip-­‑to-­‑Chip Gbps pJ pico-­‑second Hardly ¡any ¡power ¡or ¡6me ¡to ¡recover ¡a ¡transmiaed ¡bit Dagstuhl - August 2016

  12. Differen6al ¡Signaling Driver ~ x Comparator b + Transmission ¡line sgn( x-­‑ y) ~ -b - y Transmits ¡one ¡bit ¡per ¡ a ¡pair ¡of ¡wires Dagstuhl - August 2016

  13. Chordal ¡Codes Brief ¡intro ¡into ¡theory Dagstuhl - August 2016

  14. Chordal ¡Codes Codewords ¡(signals) Comparators ¡(central ¡hyperplanes) Opera6onal ¡constraints Dagstuhl - August 2016

  15. Parameters ( C , Λ ) is ( n, N, I )-CC. • n is called the number of wires • log 2 ( N ) /n is the rate or the pin-e ffi ciency #bits ¡per ¡wire • | Λ | is called the detection complexity . Dagstuhl - August 2016

  16. Electronics: ¡Comparators ¡ Efficient, ¡High-­‑Speed ¡Electronic ¡Circuits x + Referenced ¡comparators sgn( x -­‑ Ref ) - Ref Dagstuhl - August 2016

  17. Electronics: ¡Comparators ¡ Efficient, ¡High-­‑Speed ¡Electronic ¡Circuits x + Differen6al ¡comparators sgn( x − y ) - y • Circuit ¡should ¡not ¡have ¡any ¡gain. ¡ • Therefore, ¡only ¡convex ¡combina6ons ¡ allowed. x 1 + Convex ¡combina6on x k ⇣P k ⌘ i =1 a i x i − P m Mul6-­‑Input ¡comparators ¡ ¡ j =1 b j y j sgn y 1 (MIC) - Convex ¡combina6on y m Dagstuhl - August 2016

  18. Geometry: ¡ ¡ x 1 + x k ⇣P k ⌘ i =1 a i x i − P m j =1 b j y j sgn y 1 Central ¡Hyperplanes - y m + + + • A MIC corresponds to a central hyperplane + -­‑ • Each hyperplane subdivides space into two halves + -­‑ • Each codeword should ideally lie on one side or another -­‑ -­‑ • Not all codewords should lie on the same side -­‑ Dagstuhl - August 2016

  19. Transmission ¡Chain Analog ¡drivers ¡ MIC’s ¡create ¡a ¡ Digital ¡encoder ¡ transmit ¡codeword ¡ sequence ¡of ¡bits ¡ selects ¡unique ¡ on ¡collec6on ¡of ¡ from ¡received ¡wire ¡ codeword wires values Digital ¡decoder ¡ Input ¡bits ¡enter ¡the ¡ recreates ¡bits ¡from ¡ transmission ¡system MIC ¡values c 0 + b 0 - b 0 c 1 + - b 1 b 1 + - Decoder Encoder MIC’s + - + - + - + - • MIC-signature of a codeword is sequence of outputs of MIC’s. • Necessary: Every codeword has unique MIC signature. Dagstuhl - August 2016

  20. Chordal ¡Codes No ¡gain Unique ¡MIC ¡signature ¡ Dis6nguishability Dagstuhl - August 2016

  21. First ¡Bound Given n and | Λ | , determine the largest N . What ¡is ¡the ¡largest ¡rate ¡for ¡a ¡given ¡detec6on ¡complexity? n − 1 ✓ | Λ | ◆ X (1 + ( − 1) n − 1 − i ) N ≤ i i =0 Zaslavsky’s ¡Formula ¡for ¡the ¡max ¡number ¡of ¡chambers ¡ of ¡an ¡arrangement ¡of ¡central ¡hyperplanes Dagstuhl - August 2016

  22. Unbounded ¡Rate n − 1 ✓ | Λ | ◆ X (1 + ( − 1) n − 1 − i ) N ≤ i i =0 | Λ | = cn = ⇒ Rate ∼ 1 + log 2 ( c ) But: ¡ • Asympto6c ¡results ¡are ¡not ¡really ¡relevant ¡ • Didn’t ¡take ¡into ¡account ¡noise Dagstuhl - August 2016

  23. Small ¡Chambers ¡ Suscep6bility ¡to ¡Noise ERROR Dagstuhl - August 2016

  24. Noise: ¡Common ¡Mode a a + x b b + x c c + x Common mode noise d d + x e e + x • Bad ¡for ¡signal ¡integrity ¡ • Common ¡mode ¡should ¡be ¡rejected ¡at ¡receiver ¡ Means ¡that ¡comparators ¡should ¡evaluate ¡to ¡0 ¡on ¡vector ¡ ¡(1,1,1,...,1) ¡ • Codewords ¡should ¡have ¡no ¡common ¡mode ¡component ¡ Common ¡mode ¡component ¡is ¡along ¡vector ¡(1,1,1,...,1) ¡ Means ¡that ¡the ¡sum ¡of ¡the ¡values ¡on ¡the ¡wires ¡should ¡be ¡constant. Dagstuhl - August 2016

  25. Noise: ¡Inter-­‑Symbol ¡Interference Error Leads ¡to ¡errors Dagstuhl - August 2016

  26. Geometric ¡Interpreta6on Dagstuhl - August 2016

  27. Chordal ¡Codes Common ¡mode ¡resilience ISI ¡resilience Dagstuhl - August 2016

  28. Parameters ( C , Λ ) is ( n, N, I )-CC. • n is called the number of wires • log 2 ( N ) /n is the rate or the pin-e ffi ciency #bits ¡per ¡wires The ¡fewer ¡comparators ¡the ¡ • | Λ | is called the detection complexity . beaer ¡(for ¡power/area) • I is called the ISI-ratio (if equality holds for some λ , c, c 0 ). Small ¡ I ¡means ¡beaer ¡resilience ¡to ¡ISI Ali ¡Horma6 ¡[14,15] Dagstuhl - August 2016

  29. Fundamental ¡Problem Given n and N , determine smallest I such that there is a ( n, N, I )-CC. Alternatively Given n and I , determine largest N such that there is a ( n, N, I )-CC. Dagstuhl - August 2016

  30. Examples ¡ Differen6al ¡Signaling Same ¡distance Same ¡magnitude ISIR ¡= ¡1 ¡ (2,2,1)-­‑CC Dagstuhl - August 2016

  31. Examples ¡ 3 ¡Wires 2x ¡magnitude ¡ ra6o ISIR ¡= ¡2 ¡ (3,6,2)-­‑CC Dagstuhl - August 2016

  32. Bounds Obvious Every ¡comparator ¡gives ¡at ¡most ¡one ¡bit ¡of ¡informa6on Dagstuhl - August 2016

  33. Construc6ons Some, ¡not ¡all…. Dagstuhl - August 2016

  34. Tampering ¡Process What if sum of coordinates is not zero? Start with any set of codewords and comparators. • Construct ( n − 1) × n -matrix with – All rows orthogonal – Row-sum = 0 for all rows • c ∈ C : c · A . • λ ∈ Λ : λ · A . Tampering ¡process Dagstuhl - August 2016

  35. Example ✓ 1 ◆ 0 − 1 1 1 − 1 2 2 Dagstuhl - August 2016

  36. Linear ¡Chordal ¡Codes Scaling, ¡so ¡ coordinates ¡are ¡ between ¡±1 Apply tampering process to C = 1 m ( ± 1 , ± 1 , . . . , ± 1) · A • Vertices of the hypercube and Λ = scaled versions of rows of A • The coordinate axes.   1 − 1 1 − 1 C = 1 3( ± 1 , ± 1 , ± 1) · 1 1 − 1 − 1 C = ( ± 1) · (1 , − 1)   1 − 1 − 1 1 Λ = { (1 , − 1) } Λ = { (1 , − 1 , 1 , − 1) / 2 , (1 , 1 , − 1 , − 1) / 2 , (1 , − 1 , − 1 , 1) / 2 } Differen6al ENRZ Dagstuhl - August 2016

  37. Op6mal ¡Chordal ¡Codes • Op6mal ¡number ¡of ¡comparators ¡ • Op6mal ¡number ¡of ¡codewords Dagstuhl - August 2016

  38. Examples 0 0 0 0 Phantom CNRZ-­‑5 Dagstuhl - August 2016

  39. Other ¡ISI ¡Ra6os ⇒ N ≤ (1 + I ) n − 1 . • Conjecture: ( n, N, I )-CC = • Max rate . log 2 (1 + I ) • Can show rate ∼ log 2 (1 + I ) for integer I . Dagstuhl - August 2016

  40. Construc6on ¡Methods ¡ Relaxa6on • Define stripe around every hyperplane – Codewords inside a stripe are “inactive” for that hyperplane (and vice versa) – Codewords outside stripe are “active” for that hyperplane (and vice versa) • Any two codewords are separated by at least one active hyperplane • For ISI-ratio only active hyperplanes are considered Dagstuhl - August 2016

  41. Relaxa6on • ISI-ratio without relaxation = ∞ • Inactive pairs: lines and incident points • Any two points separated by mutually active line √ • New ISI-ratio = (1 + 5) / 2. Dagstuhl - August 2016

Download Presentation
Download Policy: The content available on the website is offered to you 'AS IS' for your personal information and use only. It cannot be commercialized, licensed, or distributed on other websites without prior consent from the author. To download a presentation, simply click this link. If you encounter any difficulties during the download process, it's possible that the publisher has removed the file from their server.

Recommend


More recommend