1. †UCSD, ‡UCSB, EHTZ*, UNIBO*
Energy-Efficient GPGPU Architectures via Collaborative Compilation and Memristive Memory-Based Computing
Abbas Rahimi†, A. Ghofrani‡, M. A. Montano‡, K-T Cheng‡, L. Benini*, R. K. Gupta†
†UCSD, ‡UCSB, EHTZ*, UNIBO*
Micrel.deis.unibo.it
/MultiTherman
Variability.org

2. Energy-Efficient GPGPU
Thousands of deep and wide pipelines make GPGPU high power consuming parts
NT and VOS achieve energy efficiency at costs to
Performance loss
Increasing timing sensitivity in the presence of variations
✓SIMD 
× conservative guardbands  loss of operational efficiency 
Total delay: corner + 3σ stochastic delay
Kakoee et al, TCAS-II’12
guardband

3. Variability is about Cost and Scale
Eliminating guardband 
Timing error 
Bowman et al, JSSC’09
error rate × wider width
Wide lanes
Costly error recovery for SIMD 
Recovery cycles increases linearly with pipeline length
quadratically expensive
Deep pipes

4. Taxonomy of SIMD Variability-Tolerance
Guardband
Adaptive
Eliminating
No timing error
Timing error
Hierarchically focused guardbanding and uniform instruction assignment
Error recovery
Rahimi et al, DATE’13
Rahimi et al, DAC’13
Exact / approximate
computing
Exact
computing
Predict & prevent
Memoization
Independent recovery
Recalling recent context of error-free execution
(approximately / exactly)
Lane decoupling through private queues
Pawlowski et al, ISSCC’12
Krimer et al, ISCA’12
Rahimi et al, TCAS’13
Rahimi et al, DATE’14
Detect-then-correct

5. Contributions
Efficient spatiotemporal reuse of computation in GPGPUs by collaborative
Micro-architectural design
An associative memristive memory (AMM) module is integrated with FPUs − representing partial functionality
Compiler profiling
Fine-grained partitioning of values (searching space of possible inputs)
Pre- storing high-frequent sets of values in AMM modules
Ensure their resiliency under voltage overscaling for Evergreen GPGPUs

6. Collaborative compilation framework and memristive-based computing
Training datasets
OpenCL
Kernel
Profiler
Profiling
Highly frequent computations
one-off activity
Customized clCreateBuffer to insert AMM contents
2) Code
generation
AMM contents
Kernel
lunching kernel
programming
FPU
AMM
3) Runtime =?

7. Return pre-stored result
AMM with FPU
Error  No
Recovery 
AMM:
Software programmable
Mimics partial functionality of FPU
Two pipelined stages
Return pre-stored result
Search
Operands
TCAM: a self-referenced sensing scheme†, 2-bit encoding, 15% positive slack at 45nm
Memory block: avoids read disturbance
Ternary content addressable memory (TCAM)
Crossbar-based
1T-1R memristive memory block
†Li et al, JSSC’14

8. Programming before lunching kernel
OpenCL Sobel
AMM Hit Rates
Profiler
+: {a, b} → {q}
*: {a, b} → {q}
√ : {a} → {q} …
train
test1
offline
Programming before lunching kernel
FPU+
AMM+
test2
FPU*
AMM*
FPU√
AMM√ …
test3
runtime
test4

9. Efficiency under Voltage Overscaling
33%
30%
36%
19%
17%
33%
28%
32%
39%
29%
37%
28%
Reduce timing errors from 38% to 24%
At 1.0V, without any timing error, 36% average energy saving (7 kernels)
At 0.88V, on average 39% energy saving

10. Conclusion
Static compiler analysis and coordinated microarchitectural design that enable efficient reuse of computations in GPGPUs
Emerging associative memristive modules are coupled with FPU for fast spatial and temporal reuse
GPGPU Kernels exhibit a low entropy yielding an average energy saving of 36% on the 32-entry AMMs