hybrid analog-digital solution of nonlinear partial...
TRANSCRIPT
Solve nonlinear ODE using analog accelerator
Operate on numbers encoded as voltage / current
Program connections & parameters for ODE
Configure circuit inside analog accelerator
Monitor / obtain analog waveform solving ODE
Extend ODE solving to help linear algebra & PDEs
Time independent PDE
Time dependent PDE
Partial differential equation (PDE)
Parabolic PDE (e.g., heat equation)
Hyperbolic PDE (e.g., wave equation)
Elliptic PDE (e.g., Poisson eq.)
Spatial discretization Spatial discretization
System of ordinary differential equations (ODE)
Explicit time stepping (e.g., RK4, analog)
Implicit time stepping (e.g., backward Euler)
Sparse system of linear equations (SLE)
Direct solvers (e.g., Cholesky, QR, SVD)
Iterative solvers (e.g., CG, analog)
Dense linear eqs. (e.g., optimization)
Nonlinear system of equations
Nonlinear solvers (e.g., Newton’s)
∫
∫
DAC ADC
AnalogInput
AnalogOutput
LUTsin(x)
-.5x
-5x
−𝑎𝑎10
−𝑎𝑎00
∫d𝑥𝑥1d𝑡𝑡 𝑥𝑥1(𝑡𝑡)
⍦
ADCDAC
∫d𝑥𝑥0d𝑡𝑡 𝑥𝑥0(𝑡𝑡)
⍦
ADCDAC
−𝑎𝑎11
−𝑎𝑎01
𝑏𝑏1
𝑏𝑏0
∫∫+F(u)
-JF(u)
∫∫u
δ≈JF-1(u)F(u)
-
0
50
100
150
200
0 200 400 600
conv
erge
nce
time
(s)
total grid points
digital CG
analog 20KHz
Linear (analog 20KHz)
Linear (analog 80KHz projection)
Linear (analog 320KHz projection)
Linear (analog 1.3MHz projection)
1.00E-06
1.00E-05
1.00E-04
1.00E-03
1.00E-02
1.00E-01
0.001 0.01 0.1 1 10
Tim
e to
con
verg
ence
(sec
onds
)
Reynolds number
16-by-16 Grid
Digital Time Analog Time
1.00E-06
1.00E-05
1.00E-04
1.00E-03
1.00E-02
1.00E-01
0.001 0.01 0.1 1 10
Tim
e to
con
verg
ence
(sec
onds
)
Reynolds number
8-by-8 Grid
Digital Time Analog Time
1.00E-06
1.00E-05
1.00E-04
1.00E-03
1.00E-02
1.00E-01
0.001 0.01 0.1 1 10
Tim
e to
con
verg
ence
(sec
onds
)
Reynolds number
4-by-4 Grid
Digital Time Analog Time
• [email protected]• [email protected]
• Ning Guo, Yipeng Huang, Tao Mai, Sharvil Patil, Chi Cao, Mingoo Seok, Simha Sethumadhavan, Yannis Tsividis, “Low-Energy Hybrid Analog/Digital Approximate Computation in Continuous Time,” IEEE Journal of Solid-State Circuits (JSSC), 2016.
• Yipeng Huang, Ning Guo, Mingoo Seok, Yannis Tsividis, Simha Sethumadhavan, “Evaluation of an Analog Accelerator,” ACM/IEEE International Symposium on Computer Architecture (ISCA), 2016.
• Yipeng Huang, Ning Guo, Mingoo Seok, Yannis Tsividis, Kyle Mandli, Simha Sethumadhavan, “Hybrid Analog-Digital Solution of Nonlinear Partial Differential Equations,” IEEE/ACM International Symposium on Microarchitecture (MICRO), 2017.
• Awarded Defense Advanced Research Projects Agency (DARPA) Small Business Technology Transfer grant to investigate modern analog computing applications.
• Anticipated slowdown of speed and efficiency improvements in digital integrated circuits due to material and economic constraints
• Analog circuits deliver fast and efficient computation with existing silicon technology; analog computing history offers new directions
• Continuous-time, continuous-value hardware matches well with physical problems, which are often stochastic and nonlinear
• Downsides such as low precision and accuracy, limited scalability, and difficulty in programming can be mitigated
• Approximate solutions are useful in physical simulations and machine learning tasks for humanity’s grand challenge problems
0.08 0.07 0.08 0.07 0.080.15
0.09 0.10
0.81
0.06 0.06 0.06 0.06 0.06 0.08 0.08 0.08 0.05
0
0.5
1
1.5
2
0.01 0.02 0.03 0.06 0.13 0.25 0.50 1.00 2.00
Tim
e to
con
verg
ence
(s)
Reynolds number
Solution time vs. Reynolds numberfor digital and seeded digital solver
baseline digital solver analog seeded digital solver
Using negative feedback analog solves linear equations; just provide constants, coefficients, and initial guesses
For smaller problem sizes & lower precision solutions, analog compares favorably against digital iterative numerical methods
Analog linear algebra subroutine accelerates Newton’s method, useful for finite difference solution of nonlinear PDEs
Analog continuous Newton’s method takes infinitesimal steps, avoiding challenges in selecting correct initial guesses
When solving nonlinear equations resulting from Burgers’ PDE, analog solves faster if problems are large and more nonlinear
Analog approximate solutions for nonlinear equations are initial guesses for digital, which proceeds to high precision solution
As digital hardware improvements slow, analog holds promise in problems that are continuous, nonlinear, stochastic, approximate
Our group’s multiple generations of analog accelerator chips validate full-custom analog circuits & aid application discovery
Contact information and selected publications
Physically fabricated analog accelerator chips provide ground-truth timing, area, and power consumption measurements
20 KHz analog design bandwidth16 integrators & 32 multipliersConfigurable analog crossbar
8-bit analog-digital-analog conversionNonlinear function lookup tables
Analog inputs & outputs for multi-chip cooperation
1.2V 65nm TSMC manufacturing process3.7mm × 39mm dimensions
4.8 mW at full power0.06 W/cm2 very low power density
Hybrid Analog-Digital Solution of Nonlinear Partial Differential EquationsYipeng Huang, Ning Guo, Mingoo Seok, Yannis Tsividis, Kyle Mandli, Simha Sethumadhavan
Columbia University