quantum architecture more unknowns than knowns
DESCRIPTION
Quantum Architecture more unknowns than knowns. Mark Oskin University of Washington. Outline. What / Why / How Design Rules and Technology Abstraction Quantum Architecture Simulation Infrastructure Programming languages. What is it?. Quantum Architecture:. - PowerPoint PPT PresentationTRANSCRIPT
Quantum Architecture
more unknowns than knowns
Mark OskinUniversity of Washington
Outline
• What / Why / How• Design Rules and Technology
Abstraction• Quantum Architecture• Simulation Infrastructure• Programming languages
What is it?
• (1) The organization and optimization of quantum and classical structures (i.e. the micro-architecture) and the interface (i.e. the ISA) for the efficient execution of quantum-enabled software.
• (2) A dark vast babble-space
Quantum Architecture:
Why?
Quantum architecture research can– Identify the weak spots in technology
• Point the way to solutions for some of them• Push the rest back to the physicists
– Discover what we don’t know• A surprisingly useful thing to know
– Bring a reality check to this process• Identify physical bounds that alter theoretical ones• Quantify the “known” aspects => quite large
– Maybe find the right abstraction?
- Now?
How
• Need expertise in both disciplines– Quantum theorist and physicist– Architecture Engineers
• Funding is the easiest part– NSF Nanoscale initiative– DARPA QuIST
• Students are available– Lots of interest– Need only simple background in
• Architecture• basic QC theory
– Can stay away from the dicey parts at first
How
• It’s not exactly SimpleQubit but…• Currently mathematical models• Working on an architecture simulator• Physicists working on component
simulator• “Applications” are well known:
– Its 99%++ error correction– They have all the things we like:
• Locality• Parallelism
Quantum Architecture
I. Abstracting technologiesII. Formulate design constraintsIII. Mold into building blocksIV. Form into architecturesV. Simulate application
performance
Technology abstraction
• First order assumptions:– Classical control of quantum gates– Silicon to interface and control– Individual control of quantum bits
Second order assumptions
• Choose a likely technology: Kane– Spin of 31P holds quantum state
20nm apart for quantum effect to occur 1.5Kelvin for reasonable coherence time
– Local magnetic field arbitrates gates• Controlled by “classical” pins
5nm classical pitch
• Driven by high frequency (10-100Mhz) clock• Gated by “lower” frequency (0.01 – 10) Mhz
• Similar to CMOS vs. TTL
1.5
Develop design rules
• 20nm spacing of qubits• 5nm spacing of control lines
– @ 1.5 Kelvin cannot drive AC current– 2 dimensions must be 100nm
• “pitch matching” issue– Implies sparseness of quantum state
Quantum architecture
• Abstractions– Interconnect– Memory– Processor
• Interfacing– Quantum ISA– Classical-Quantum interface
Specialization?
A Quantum Wire
• Short: swapping-channel– structural implications (sparseness)– Limited length
• Long: teleportation-channel– “Arbitrary” length– Architectural implications
• Overhead• Latency / bandwidth
A short quantum wire
• Constructed from swap gates
Unless the particle that holds the quantum state physically moves, the information “flows” in discrete steps from particle to particle.
Each step requires 3 quantum controlled-not operations, effectively performing a “swap” of the quantum states.
Straightforward approach
5nm access points contain only a handful of quantum statesfor their electrons at temperatures less than 1K, compromising correctoperation.
As two physical dimensions ofthe access point exceed 100nmthousands of electron states are held. Classically, these
states are restrictedto the access point,however, quantummechanically theytunnel downward,guided by the via,thus enabling control.
One solution…
100nm
5nm
20nm
100nm
100nm
Classical access points
Narrow tippedcontrol
20nm
100nm
Incompleteness of lines
Top-down view
QCAD Cell Implications
• Minimum wire length 200nm (10 qubits)– Excepting custom components
• Minimum junction point size 44 qubits square
• Realistic sizes will be larger– Assumes deep 5nm vias
Why short wires are short
• Limited by decoherence• Threshold theorem => distance
– 10-8 1.8mm
• Key difference from classical:– quantum information must be protected,
not just restored!!
• Can make longer with “repeater”– Essentially this is multiple short wires
separated by error correction blocks
Architecting long wires
• Key insight:– EPR pairs are known states
• No need to protect them– Purify the good ones– Discard the bad
Architecture of a long wire
EPRGenerator
Tele
pora
tion U
nit
Tele
pora
tion U
nit
Entropy Exchange
Purification
CodedTele-
Portation
Classical control channel
Quantum EPR channel
EPR channel
Long wires
• Can be of “arbitrary” length– A 10mm wire sustains nearly peak
bandwidth
• Low latency– Pre-communicate EPR pairs– Latency is constant: teleportation operation
• Code-conversation for “free”– Facilitates Processor <-> Memory
communication
Long wires
• Several architectural implications– EPR generation– Distributed entropy exchange (zero’s)– Purification– Teleportation
QCAD Cells
• Fundamental– Qubit– Zero– Measurement
• Basic– Line– Intersection
• Composite / Custom– Purify (custom error correct)– Error correct– Add? Multiply? Memory?
Building Block (I)Building Block (I)
• Measurement unit – computational & Bell basis
Measure
0
Qubit to measure
Zero qubit
Classical control
Classical {0,1} outputwith probabilitydetermined by
Building BlockBuilding Block
• Entropy exchange unit
0 0 …
EX
P
PolarizedLight
Polarized ElectronsElectric Field
Ground
Macro BlockMacro Block
• EPR generation unit
EPR
EPR Generator
0.....0Zero qubits
Classical controlQuantum outputof an EPR state
2
1100
Macro BlockMacro Block
• Purification unit – error correction
Pur
Purification UnitEPR states to purify
Classical control
Purified EPR statesZero bits 0.....0
Garbage state (to Entropy Exch) NE
10
M 10
Quantum Memory
Quantum memory?
• Is dedicated memory viable?• Yes
– DRAM like (needs refreshing)– Hierarchical error codes?
• Quantum caches
– DFS (Decoherence Free Subspace)?• Really phase coherent subspace• Need less error correction/qubit
• No– Qubit Refresh almost as complex as computation!– Big “Almost” => No T gate / all transversal
Quantum ALU / ISA
Quantum Functional Unit
• Complex, have to tightly integrate:– Measurement– Zeros– Quantum I/O– Irregular classical logic
• Maybe custom data-paths for:– H/X/Z– CNot– T– Complicated by hierarchical error coding
Processing
• Likely to use just-in-time compilation– Huge O(n*c^k) savings with error
correction:• Optimize overhead to data size• Clustering
– Smaller O(n*c) savings:• Packing / unpacking• Application specific error processing
– Phase error independence– Bit-flip error independence
Flexible execution units
Classic analogy: MMX (except more complicated to combine)
Interfacing and Control
• Quantum operations occur at different speeds– ~ 10-100Mhz for single qubit rotations– ~ 10-100Khz for two-qubit operations– ~ 1Mhz on average (50/50 split)
• Even at 1Mhz operation– Ample opportunity for interesting classical work…– Error correction creates even more time for top-
level control (5^k)– Low-level must simultaneously decide on the
control of millions of qubits/Mhz
Controlling the classical control
• Highly parallel– O(n) operations per-cycle!– Required for fault-tolerant operation
• Specialized classical processors?– Certainly ASIC logic for drive/control– Quantum co-processor ISA interface?
Quantum ISA
• Single qubit rotations– rotate(qubit, axis, angle)
• Controlled operations– rotate(qubit, axis, angle, {on list})
• Just-Enough-Compilation– Control error correction overhead– Invoke(program, input, input
complexity)
Simulation
• Architecture Simulation– Abstraction layer
• QCAD Cells• Macro blocks (memory, etc)
– Classical interfacing• Bolt onto SimpleScalar??
– Design path• QVHDL -> Cell Layout
How?
• Quantum simulation is O(2^n) hard– Obtaining the right algorithmic answer is
not going to happen
• “Symbolic” simulation is only O(n*t)– Classic n-body simulation– Eminently Parallelizable– Look for this in the Fall
Programming Abstractions
• Quantum computing lacks a clear abstraction for computer scientists– Matrix algebra just isn’t intuitive
enough
• Difficult to abstract– 2^n states for n bits– entanglement
A Classical Representation of Quantum Circuits
Example: Quantum Teleportation
0
0
H
H
X Z
Not obvious that this measurementaffects the probability distributionfor this quantum bit
Not explicit that these qubitsare now entangled…
Critic
+ Concise+ Familiar+ Classical decisions are explicit- Super-position is hidden- Entanglement is hidden
Alternative Representation
000
001 001
000
101
100
001
000
111
110
011
000
101
110
011
000
101
110
010
001
100
111
000
110
010
100
110
010
010
110
H H XC C
Critic
- Not very concise (exponential!)- Not very familiar (where are the
qubits?)- Classical decisions are implicit+ Super-position is exposed+ Entanglement is exposed
Ideal Programming Abstraction
• Concise• Familiar within reason• Integrates Classical/Quantum• Exposes super-position and
entanglement
Conclude
• Choose your area of interest and there is work to do:– Design rules / cell development– Architecture abstractions– Classical-Quantum interfacing– Programming languages
Notes / Graduate course
• http://www.cs.washington.eduhomes/oskin/quantum-tutorial
• Notes based on book by Michael Nielsen and Isaac Chuang (with some info from John Preskill)
• Graduate course w/UG’s on request• Geared for computer scientists
– Begins with linear algebra review– Ends with error correction
• Sequence of programming assignments in QCL
QARC Project
• Quantum Architecture project– Isaac Chuang, MIT– Fred Chong, UC Davis– John Kubiatowicz, UC Berkeley– Mark Oskin, UW