POWER-DRIVEN MAPPING K-LUT-BASED FPGA CIRCUITS I. Bucur, N. Cupcea, C. Stefanescu, A. Surpateanu Computer Science and Engineering Department, University Politehnica of Bucharest, Romania Semiconductor Conference, 2009. CAS 2009. International
Dec 28, 2015
POWER-DRIVEN MAPPING K-LUT-BASED FPGA CIRCUITSI. Bucur, N. Cupcea, C. Stefanescu, A. Surpateanu
Computer Science and Engineering Department, University Politehnica of Bucharest, Romania
Semiconductor Conference, 2009. CAS 2009. International
Introduction• Power consumption is becoming one of the most
important considerations in VLSI design.• Especially in FPGA design.• In this paper is presented a new mapping approach for
decreasing the spurious power consumption in K-LUT based FPGA implemented circuits.
Background
FPGA POWER• Static power :
current leakage in transistors• Dynamic power :
signal transitions between logic-0 and logic-1•Functional transitions - necessary for the correct operation of the circuit•Glitch - unbalanced delays to the inputs of a logic gate (effect on power consumption)
• Dynamic powern : the number of nets in the circuitSi : the switching activity of net ICi : the capacitance of net If : the frequency of the circuitVdd : the supply voltage
Net switching activity estimation
- probability approaches- simulation-based approaches
This paper uses simulation-based simulator of SIS-1.2• This gate-level simulations provide both the functional and
total activity.• Spurious transition activity is computed as the difference
between them.
• Technology mappingtransforms the gate-level network into a network of cells in the target technology library.
Algorithm• K-feasible-cones enumeration from PIs to POs
• Make the selection among the K-feasible cones of each node guided using critical path in circuits and several appropriate cost functions from POs to PIs
• Depth Metric for node u is computed over one of the best depth K-feasible cone of u to quantify the depth criterion:
• In order to quantify the best suitable cone rooted in u is introduced the cost function bestCone. It is locally applied for the entire set of K-feasible cones of the node u:
• |LUT(v)|: the number of internal nodes of cone v previously (already) chosen as LUTs,
• |cone(v)|: the number of internal nodes of cone v.• activity(w): the spurious logic activity estimation of the
node w• |fanout(w)|: the number of nodes connected to the output
of node w• q1 (w) = 1 iff |fanout(w)| < 3• q1 (w) = 0 iff |fanout(w)| >= 3• q2 is the parameter controlling the influence of the
spurious activity estimation in multi-criteria cost
First part of multiCrit• The numerator of the quotient penalizes node duplications
by increasing the cost of cuts that encapsulate nodes that have already been labeled as root nodes.
• The denominator rewards cuts that encapsulate many nodes that have not been labeled as root nodes.
Second part of multiCrit• The q1 factor minimizes node duplication by favoring cuts
that reuse nodes. The activity factor minimizes the switching activity of the connections.
• The fanout size factor rewards cuts that have high-fanout input nodes.
• High-fanout nodes are difficult to encapsulate entirely-attempting to encapsulate them results in unnecessary node duplication.
• This is avoided by choosing high-fanout nets as root nodes.
• Using this cost function, nodes with large fanouts are likely to be chosen as root nodes.
• Both parameters, q1 and q2, were experimentally determined. Best results were obtained when q2 >> q1 reflecting the preference for optimizing power over depth.
Experimental result• To estimate power consumption using, it is required the
capacitance of each net. The fanout of each mapped node was considered as an estimate of the capacitance of it
• Optimum depth: keep optimum depth and search among power-aware equivalent solutions
• Optimal depth: optimal depth but performing with improved spurious power consumption.
• Optimal depth & area: optimum depth, with minimal area (number of used LUTs)
Conclusion• Trade-off between dynamic power and area.• Power-driven mapping both for depth and area optimal, it
appears to be more complex than mapping only for optimal depth.
• Actual working heuristics have to be upgraded because it was searched only a limited part of mapping solutions’ space.
• It is intended in the future approach to use dynamic programming together with refined heuristics to further develop PwDrvMap algorithm.