memory/cache optimization
DESCRIPTION
Memory/Cache Optimization. Lochana Narayanan Suchitra Chandran. A Distributed BIST Control Scheme for Complex VLSI Devices. Yervant Zorian AT&T Bell Laboratories Princeton, NJ, 08540. Introduction. The problem in today’s VLSI devices. Power, noise and area overhead. - PowerPoint PPT PresentationTRANSCRIPT
Memory/Cache Optimization
Lochana NarayananSuchitra Chandran
A Distributed BIST Control Scheme for
Complex VLSI Devices
Yervant Zorian
AT&T Bell Laboratories
Princeton, NJ, 08540
3/42
Introduction• The problem in today’s VLSI devices.
– Power, noise and area overhead.
• Solution – BIST! (Built in self test)
4/42
How BIST does it? It is used at all levels and reused at all consecutive
levels. Possible only when if BIST features are included in
devices. Devices are the building blocks – higher BIST levels
are built. So, device level BIST realization takes into account
the BIST requirements of all levels, device, MCM, board, system.
These requirements affect the BIST Execution Scheduling, hence the need for BIST Control Network.
5/42
Existing Scheduling approaches
• Focus on optimizing the test time for random logic blocks.
• Problems in executing BIST in parallel :» Power and noise dissipation
» Higher activity rate
» Overpass the device package limits.
• New BIST scheduling process :» Power dissipation less
» Global optimization – block type, device floor plan and test time.
6/42
New BIST Control Scheme
• Simplifies BIST Execution – Autonomous embedded block
– Simple and uniform interface protocol to communicate b/w controller and embedded blocks
– Distributed architecture
» Minimize routing cost – control elements placed close to BIST blocks
» Minimal control signals
7/42
BIST Scheduling Process
• Analysis process where no hardware modifications are allowed.
• Prior to scheduling, 2 analysis processes are performed:
– Device partitioning
– BIST scheme selection
8/42
Partitioning and BIST Control
• ASIC Z is a 132 pin standard device
» Partitioned into ten blocks
» Based on different criterion :
- Structural types – 4 RAMs 2 ROMs and Register File (RF)
-Random Logic blocks – RL1 and RL2
- High frequency blocks – RL3
9/42
Partitioning and BIST Control Contd.
Each block requires Test Pattern Generator (TPG) Output Data Evaluator (ODE) BIST Resource Controller (BRC)
RL1 and RL2 – Pseudo random based TPG and polynomial based ODE
RF and 4 RAMs – 100% fault coverage TPG
ODE(RF) - provides zero loss of fault coverage
ODE(RAM)- performs output data comparison and space compaction
2 ROMs – TPG and ODE -> perform exhaustive test sets
10/42
BIST Scheduling attributes
i. Power dissipation of each block
ii. Adjacency of blocks
iii. Type of each block
iv. Test time
11/42
i. Power Dissipation Analysis
• BIST is executed at system clock rate• Adv. : full benefit is obtained
• Disadv. : higher activity rate -> excess power dissipation
• P=F.N.PG• F : Frequency
• N : Block size = number of grids
• PG : number of watts per active grid and activity rate (0.5 for BIST modes)
• Different modes : Normal, BIST active and BIST idle
• BIST Active – during execution of a block
• BIST Idle – during periods when it waits for other blocks to execute BIST.
• Total power dissipation = power from each block + power from input and output buffers.
12/42
Power Dissipation Analysis Contd.
• Package limit of ASIC Z is 900mW.
• Based on the values in the table, the device power dissipation for parallel BIST execution is 2.332W.
• Hence BIST Scheduling for power optimization is necessary.
13/42
ii. Grouping and Ordering• Create groups and proper sequence of each stage – BIST Execution
• Grouping and Ordering is dependent on the physical distribution of the blocks on the floor plan of a device.
• Grouping is done based on 2 conditions :
• BIST power dissipation has to be less than device limit
• Blocks in a group have to be physically adjacent
on floor plan.
• A profile which is the result of grouping and sequencing at device level is in a matrix format :
• Bij - block, i – single BIST stage, j – number of blocks
• For our example, we have :
14/42
iii. Sharing and Test Time Optimization
• Sharing reduces hardware area by reducing redundancies.
• Sharing between blocks of same type: RAMs, random logic, since their BRCs are identical and TPGs and ODEs are different in parameters.
• Combine Boundary Scan with pseudo random TPGs on one hand and with ODEs on the other.
• Test time is a limited amount of time that is dedicated per device.
• BIST Execution takes only a minor fraction of the total test time as the other tests are far more time consuming.
15/42
BIST Control Architecture
• Control operation has 4 functions:• External access to device level BIST.
• Control of BIST facilities of each block.
• Control of device BIST.
• Transfer of BIST response data.
• These functions are realized in 3 types of hierarchical controllers :
– External BIST Access Port
– BIST Resource Controllers (BRCs)
– BIST Control Network.
16/42
External BIST Access Port
• Boundary Scan TAP is the external access port for the device BIST.
• For simple tests – use 1 controller, TAP finite machine and boundary scan instructions.
• For complex tests – separate controllers are required for each one of BIST control functions.
• BIST execution and BIST response data transfer are done via BS TAP – no additional pins dedicated for external BIST access.
17/42
BIST Resource Controllers
• Individual BRCs have to be allocated because :– Number of blocks is in continuous growth.
– BIST schemes of each block are different.
• They are customized – TPG and ODE
• Sharing of non identical BRCs– low complexity – FSM
– Not cost effective
18/42
BIST Control Network
• Executes the device BIST based on predetermined schedule.
• Functions :– Receives BIST execution via TAP and provides activation
signals.
– Collects and transfers BIST response when BIST completes.
• Advantages :– Division of control network – each function controls subsequent
BIST operation.
19/42
Distributed BIST Control Network
• Centralized controllers for device level BIST control.
• Problems :– Routing area and number of control lines increase with the
increase in the number of BRCs
• For smaller number of blocks, centralized BIST control is possible.
• Optimizes cost of control lines. 20/42
SBRIC Finite State Machine
• Distributed network has control elements called Scheduled BIST Resource Interface Controllers
21/42
BIST control network
• Disadvantages :Accumulates signatures of all BISTed blocks to a common compactor – loss of diagnostic information.
•Advantages :
Provides not only group but also block level information.
22/42
Conclusions- Generic BIST scheduling process and effective BIST control architecture.
- Scheduling provides power optimization.
- Control architecture provides autonomous BIST activation and capability to identify failed blocks.
Future Works- BIST network with additional diagnostic capabilities can be used for reconfiguration and repair operation.
- Development of scheduling processes that take into account multi level test restrictions.
23/42
A Hybrid BIST Architecture and its Optimization for SoC Testing
Gert Jervan, Zebo Peng Raimund Ubar, Helena KruusLinköping University, Sweden Tallinn Technical University, Estonia
OVERVIEW
INTRODUCTIONINTRODUCTION
BIST
HYBRID BIST
APPROACH OF THE PAPER
TARGET HYBRID BIST ARCHIETECTURE
TARGET HYBRID BIST ARCHIETECTURE
COST of HYBRID BIST
COST of HYBRID BIST
CGEN : cost related to the time for generating L pseudorandom test patterns (number of clock cycles)CMEM : related to the memory cost for storing S pre-computed testpatterns to improve the pseudorandom test setΒ, α : constants to map the test length and memory space to thecosts of the two parts of the test solutions to be mixed.
COST CALCULATION FOR PSEUDORANDOM TEST
BIST Analysis Data
COST CALCULATION FOR STORED TEST
Notations used
ALGORITHM 1
Notations used
ALGORITHM 2
COST CALCULATION FOR STORED TEST
TABU SEARCH
RESULTS
L : length of pseudorandom sequence C: fault coverage CT : total cost of BISTS : number of test patterns generated by deterministic ATPG to be stored in BISTTG : the time (sec) needed for ATPG to generate the deterministic test setTA: the time(sec) needed for carrying out manipulations on fault tablesN : number of efficient patterns in the pseudorandom test sequenceT1,T1 : the time (sec) needed for calculating the cost curve by Algorithms 1 and 2T3 : the time (sec) to find the optimal cost by using Tabu searchTs: number of calculations in Tabu searchAcc : percentage accuracy of Tabu search solution compared to solution found from cost curve
RESULTS
Percentage of test patterns in the optimized test sets compared to the original test sets
RESULTS
Cost comparison of different methods.Cost of pseudorandom test is taken as 100%
SUMMARY
FUTURE WORK
QUESTIONS