Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf ·...
Transcript of Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf ·...
![Page 1: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/1.jpg)
Quantum Memory Hierarchies
Efficient Designs that Match Available Parallelism in Quantum Processors
Darshan D. ThakerTzvetan S. Metodi
UC Davis
Andrew W. CrossIssac L. Chuang
MIT
Frederic T. Chong
UC Santa Barbara
![Page 2: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/2.jpg)
Study tradeoffs between area - reliability - performance.
Goals of this research: Reduce overall area of the design.
Leverage conventional architectural techniques to improve performance.
Provide abstractions for further research.
Motivation
![Page 3: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/3.jpg)
Background and prior work.
Overview of quantum error correction codes.
Specialization into memory and compute regions.
Improving performance.
Results and discussion.
Outline
Background and prior work.
![Page 4: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/4.jpg)
Use ions trapped in electromagnetic fields.
Lasers acting on ions induce quantum gates.
Newer traps are micromachined.
Ion-Traps
Courtesy: C.Monroe at U.Michigan
Ion trapping region
T-junction
![Page 5: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/5.jpg)
Compute-anywhere design.
Teleportation based long-distance communication.
Exponential speedup when factoring large numbers.
Unresolved Issue: Size
Quantum Logic Array
Q Q Q
Q Q QQ
R
R
R
R
R
R
Sea of lower level
qubits
Single logical qubit
Repeaters
![Page 6: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/6.jpg)
Sea of qubits design.
Teleportation based long-distance communication.
Exponential speedup when factoring large numbers.
Unresolved Issue: Size
Quantum Logic Array
QLA: 90cm x 90cm
![Page 7: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/7.jpg)
Design Pyramid
Speed
ReliabilityArea
QLA
![Page 8: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/8.jpg)
Background and prior work.
Overview of quantum error correction codes.
Specialization into memory and compute regions.
Improving performance.
Results and discussion.
Outline
Overview of quantum error correction codes.
![Page 9: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/9.jpg)
Classical three bit code Equivalent quantum code
Comparison with Classical Codes
Single bit encoded as three bits.
Majority Voting.
Nine qubit Shor code
Protects against bit-flips and phase-flips.
![Page 10: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/10.jpg)
Classical three bit code Equivalent quantum code
Comparison with Classical Codes
Single bit encoded as three bits.
Majority Voting.
Nine qubit Shor code
Protects against bit-flips and phase-flips.
![Page 11: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/11.jpg)
Need greater reliability than provided by encoding a single time.
The No cloning theorem and restrictions on measurement require greater reliability.
Cannot use methods like checkpointing or make duplicates.
Solution: Use concatenated codes.
Greater Reliability
![Page 12: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/12.jpg)
Concatenated Codes
Reliability increases doubly exponentially.
Exponentially slower.
Exponentially greater resources.
1 logical qubit
Level 1: 7 physical qubits
Level 2: 49 physical qubits
Concatenated Steane Code
![Page 13: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/13.jpg)
Background and prior work.
Overview of quantum error correction codes.
Specialization into memory and compute regions.
Improving performance.
Results and discussion.
Outline
Specialization into memory and compute regions.
![Page 14: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/14.jpg)
Quantum Logic Array
Q Q Q
Q Q QQ
R
R
R
R
R
R
Sea of lower level
qubitsQ Q Q
Q Q QQ
R
R
R
R
R
R
Sea of lower level
qubits
R
R
Conventional wisdom: Max. parallelism necessary to minimize computation time and reduce prob. of failure.
![Page 15: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/15.jpg)
Shor’s quantum algorithm to find factors of very large numbers yields exponential speedup over classical algorithms.
Modular exponentiation is the most compute intensive part of Shor’s factoring algorithm.
Primary component: Draper carry-lookahead adder (quantum version of the classical adder).
Modular Exponentiation
![Page 16: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/16.jpg)
App. Constrained Parallelism
Create slower but denser memory region and faster but sparse compute region.
![Page 17: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/17.jpg)
Compute Region Memory Region
Specialization
Logical data qubits Logical ancilla qubits
An ion when idle has a lifetime of ~10 sec
Ancilla : Data 2 : 1
Ancilla : Data 1 : 8
![Page 18: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/18.jpg)
CQLA: Compressed QLA
Compute Block
Memory Block
![Page 19: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/19.jpg)
Area Reduction
-0.5
3.0
6.5
10.0
64-bit 256-bit 512-bit 1024-bit
9.1
7.4
6.66.4
Area Reduced Perf. Change
Shor’s Alg. Input Size
Fact
or o
f
-20%
![Page 20: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/20.jpg)
CQLA: Reduced Size
QLA: 90cm x 90cm
CQLA: 28cm x 28cm
![Page 21: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/21.jpg)
Design Pyramid: CQLA
Speed
ReliabilityArea
QLA
CQLA
![Page 22: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/22.jpg)
Background and prior work.
Overview of quantum error correction codes.
Specialization into memory and compute regions.
Improving performance.
Results and discussion.
Outline
Improving performance.
![Page 23: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/23.jpg)
Concatenated Codes
Reliability increases doubly exponentially.
Exponentially slower.
Exponentially greater resources.
1 logical qubit
Level 1: 7 physical qubits
Level 2: 49 physical qubits
Concatenated Steane Code
![Page 24: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/24.jpg)
Compute @ Level 1
Memory: Very reliable and slow.(Periodic error-correction)
Compute: Very reliable and fast.(49bit quantum operations and error-correction)
Level2
Encoding
Cache: Less reliable.(Infrequent Error-correction)
Compute: Less reliable, exponentially faster.(7bit quantum operations and error-correction)
Level1
Encoding
Transfer between encoding levels
![Page 25: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/25.jpg)
Faster CQLA
Memory Block
Compute Block
Cache @ Level 1 Compute @ Level 1
![Page 26: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/26.jpg)
Overall Results
0
4.75
9.50
14.25
19.00
256-bit 512-bit 1024-bit
4.96.26.2
18.217.417.4
9.1
6.05.0
Area Reduced L1 SpeedUp Total SpeedUp
Shor’s Alg. Input Size
Fact
or o
f
![Page 27: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/27.jpg)
Design Pyramid: QLA
Speed
ReliabilityArea
QLA
![Page 28: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/28.jpg)
Design Pyramid: CQLA
Speed
ReliabilityArea
QLA
CQLA
![Page 29: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/29.jpg)
Design Pyramid: CQLA v2
Speed
ReliabilityArea
QLA
CQLA v2
![Page 30: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/30.jpg)
Parallelism in quantum computing constrained by applications.
Different scheduling mechanisms of quantum operations.
Introduced a memory hierarchy for quantum computers.
Area reduced factor of 9 and speedup of factor of 4.
Discussion
![Page 31: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/31.jpg)
Even better results using the Bacon-Shor quantum error-correction code. ! Area reduced by a factor of 13. ! Speedup of factor of 8.
Details of transfer networks to enable change in encodings.
Discussion - 2
![Page 32: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/32.jpg)
Limited control signals: Incorporate studies of laser resources and laser power.
Incorporating fault tolerance into compiler optimization: Compiler techniques to reduce error-correction costs.
Future Work
![Page 33: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/33.jpg)
Project webpage: http://aar.cs.ucdavis.edu/qarc
Your questions...
¿ Questions ?
![Page 34: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/34.jpg)
Overall Results
![Page 35: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/35.jpg)
Concatenated Codes
Reliability increases doubly exponentially.
Exponentially slower.
Exponentially greater resources.
1 logical qubit
Level 1: 9 physical qubits
Level 2: 81 physical qubits
![Page 36: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/36.jpg)
Let memory remain at Level 2 encoding.
Compute at Level 1 encoding.
Drawbacks: Reliability degrades. Transfer between Level 1 and Level 2 is very expensive.
Use a cache to alleviate transfer costs.
Improve Performance
![Page 37: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/37.jpg)
Size Reduction
Input Size Compute Blocks
Area Reduced(Factor of) Speed Up
64-bit9 6.4 0.7
16 3.7 0.9
256-bit36 6.6 0.9
49 5.0 0.9
512-bit64 7.4 0.9
81 6.0 0.9
1024-bit100 9.1 0.8
121 7.8 0.9
![Page 38: Quantum Memory Hierarchies - Directorypeople.cs.uchicago.edu/~ftchong/33001/CQLA_isca_2006.pdf · Quantum Memory Hierarchies Efficient Designs that Match Available Parallelism in](https://reader030.fdocuments.us/reader030/viewer/2022021511/5ac483d47f8b9a220b8cda88/html5/thumbnails/38.jpg)
Overall Results
Parallel Transfers
Input Size
L1 Speed Up
L2Speed Up
Total Speed Up
Area Reduced
(Factor of)
5
256 10.4 0.9 4.0 5.0
512 10.4 0.9 4.0 6.0
1024 10.9 0.8 2.9 9.1
10
256 17.4 0.9 6.2 5.0
512 17.4 0.9 6.3 6.0
1024 18.2 0.8 4.9 9.1