Beyond silicon

Silicon is not the only semiconductor in production.

"Beyond silicon" sounds like the future, but much of it already ships in volume. Compound semiconductors are made from two or more elements, usually from groups III and V of the periodic table, such as gallium and arsenic. They do not replace silicon for logic. They take over the jobs where silicon's physics falls short.

Gallium arsenide and indium phosphide move electrons faster than silicon does, and they have a direct bandgap, which means they emit light efficiently. Silicon's bandgap is indirect, so silicon makes a very poor light source. That is why the power amplifier in a phone's radio is often GaAs, and why the lasers that drive long-haul fibre links are built on InP.

Gallium nitride and silicon carbide are wide-bandgap materials. Their bandgap is about three times silicon's, so they withstand much stronger electric fields before breaking down, run hotter, and switch faster with lower losses. A power transistor made from them can be smaller and waste less energy as heat. SiC now sits in electric-vehicle traction inverters and solar inverters. GaN sits in compact laptop chargers, data-centre power supplies and radar and 5G radio amplifiers.

Why they stay niche

Compound wafers are smaller, more expensive and more prone to defects than silicon. SiC wafers are moving from 150 to 200 mm while silicon has been at 300 mm for over two decades. GaN is often grown as a thin layer on a silicon or SiC wafer to keep costs down. None of these materials has a native oxide as good as silicon's, so none of them is likely to replace it for dense logic.

Light moves data. Whether it computes is still open.

An electrical signal in a copper wire loses energy to resistance and capacitance, and the loss grows with distance and speed. Light in a waveguide or fibre does not have that problem. It can also carry many wavelengths, each with its own data stream, down the same path. Past a few metres at today's data rates, light already wins. The question now is how close to the chip it can get.

Silicon photonics builds optical parts with the same kind of process used for chips: waveguides that steer light, modulators that encode data onto it, and germanium photodetectors that turn it back into current. Silicon cannot make its own light, so a separate laser, usually indium phosphide, is attached or bonded on. Optical transceivers built this way already carry most of the traffic between racks in large data centres.

1. Electrical dataFrom the switch or chip
2. LaserIII-V light source
3. ModulatorEncodes bits onto light
4. Fibre or waveguideLow-loss path
5. PhotodetectorLight back to current
6. Electrical dataInto the next chip

Moving optics closer

In a pluggable transceiver, the signal still travels several centimetres of circuit board before it reaches the optics, and driving that board trace costs a lot of power. Co-packaged optics put the optical engines in the same package as the switch or processor, which shortens the electrical path to millimetres. Network switches with co-packaged optics moved from demonstrations to announced products in the mid-2020s. Putting optics directly between processors and memory is the next step, and it is still in development.

Optical computing

A mesh of optical interferometers can multiply a vector by a matrix as light passes through it, which is the core operation in neural networks. Research chips have demonstrated this. Three problems keep it early-stage. Converting between electrical and optical signals at every step costs energy. Precision is limited by noise and the accuracy of manufacturing. And light does not easily do the nonlinear operations a network also needs. Treat claims of large speed-ups with care until they include the cost of conversion.

When the channel gets too thin for silicon.

Course 01 showed the gate wrapping ever more tightly around ever thinner silicon. Below a few nanometres, silicon gets worse. Carriers scatter off its rough surfaces, and slight variations in thickness turn into large variations between transistors. The next step may be to change what the channel is made of.

Part of that change has already happened. Silicon-germanium has been used for well over a decade to strain p-type transistors so holes move faster, and SiGe layers are part of how nanosheet transistors are built. SiGe bipolar transistors are also standard in high-frequency radio chips. Those are adjustments to silicon, not replacements for it.

These materials are most likely to arrive first in combination with silicon rather than in its place: for example as a thin layer of transistors stacked above a silicon base, or as the channel in a stacked CFET. Roadmaps put this in the 2030s at the earliest.

Memories that keep their state.

SRAM and DRAM store bits as charge, and lose them when the power goes off. DRAM has to refresh itself thousands of times a second even while the power stays on. Flash keeps its data, but writing to it is slow and wears it out. Several newer memories store a bit in some physical state of a material instead. They keep data without power and can be built in the wiring layers above the transistors.

Spintronics

MRAM is the first mass-market product of spintronics, which uses the electron's spin as well as its charge. In STT-MRAM, a current of spin-aligned electrons flips the magnetic layer. In SOT-MRAM, which is still in development, a separate write path makes switching faster and wears the memory less, which may one day let MRAM stand in for some SRAM cache. The ferroelectric behaviour of hafnium oxide, first reported in 2011, matters for a similar reason: hafnium oxide is already the gate insulator in every leading-edge transistor, so the fabs already know how to handle it.

None of these memories is better than DRAM and flash in every respect. Each trades off density, speed, endurance and cost differently, which is why each has found a niche rather than taking over.

Computing where the data already is.

Course 01 described the memory wall: moving a number costs far more energy than doing arithmetic on it. Analog in-memory compute avoids moving the data. Each weight of a neural network is stored as a conductance in a grid of memory cells. Inputs are applied as voltages along the rows. Each cell passes a current equal to voltage times conductance, which is Ohm's law doing the multiplication. The currents along each column add together, which is Kirchhoff's current law doing the sum. A whole matrix-vector product happens in one step, in place.

1. Digital inputsFrom the previous layer
2. DACNumbers to voltages
3. CrossbarMultiply by conductance
4. Column sumCurrents add
5. ADCCurrent back to numbers

The trade-off is precision. Cell conductances vary from one cell to the next and drift over time, noise adds error, and the converters at the edges can cost more energy than the array saves. Useful precision tends to be a few bits, so in-memory compute fits inference better than training. Research chips have shown strong efficiency, and a small number of products ship for low-power inference. It is not yet a general-purpose replacement for digital accelerators.

Neuromorphic chips

Neuromorphic hardware takes ideas from the brain. A spiking neural network sends brief events only when something changes, rather than computing every value on every clock tick, so a quiet input costs almost no energy. Research systems such as IBM's TrueNorth, Intel's Loihi and Manchester's SpiNNaker have shown this works for sensing and some control tasks. The hard part is software. Mature tools for training spiking networks to match the accuracy of conventional networks on mainstream workloads do not exist yet. It remains mostly research and pilot deployments.

Quantum computing is a different kind of machine.

A classical bit is 0 or 1. A qubit can be in a superposition of both, described by two weights that set how likely each outcome is when you measure it. Qubits can also be entangled, so their measured outcomes are correlated in ways no set of independent bits can copy. A quantum algorithm arranges these weights so that wrong answers cancel and right answers reinforce. Measurement then returns a right answer with high probability.

This is not a faster way to do everything. It helps on problems with a structure that quantum interference can exploit. The strongest cases are simulating molecules and materials, which are themselves quantum systems, and factoring large numbers, which would break the public-key encryption in wide use today. That threat is why post-quantum encryption standards were published in 2024, before any machine able to carry out the attack exists. Quantum computers are expected to work as accelerators attached to classical systems. They will not replace CPUs or GPUs, and there is no strong evidence yet that they will speed up mainstream AI.

Physical and logical qubits

Qubits are fragile. Heat, stray fields and imperfect control introduce errors far more often than in a classical transistor. Error correction spreads one logical qubit across many physical qubits and checks them continually. Depending on the physical error rate, one logical qubit can take hundreds to around a thousand physical ones. The largest machines today have hundreds to low thousands of physical qubits, and teams have shown that adding more physical qubits can reduce the logical error rate. Machines with enough logical qubits to factor useful numbers are still years away, and estimates of how many years vary widely.

Silicon spin qubits are where this course meets course 02. They are made with the same lithography, deposition and etch steps as transistors, often on isotopically purified silicon-28. If they can be scaled, the existing fab industry would be the way to manufacture quantum processors.

Further out, and how to judge a claim.

A few more ideas recur in research news. Each is real physics, and each is far from replacing a silicon accelerator.

• Superconducting digital logic. Josephson-junction circuits that switch in picoseconds at tiny energy per operation. The refrigeration overhead and low circuit density have kept it a research field.
• Reversible and adiabatic computing. Erasing a bit has a minimum energy cost, about 3 × 10⁻²¹ joules at room temperature (the Landauer limit). Logic that never erases information could in principle go below it. Today's chips use far more energy than this limit, so practical gains depend on other overheads first.
• Probabilistic and thermodynamic computing. Hardware that uses noise deliberately: random bits or physical systems that settle into likely states, for sampling and optimisation. Early prototypes exist.
• DNA storage and biocomputing. DNA stores data at extraordinary density and lasts for centuries in good conditions, but writing and reading it are slow and costly. Computing with living cells or tissue is at an early stage of research.

Five questions for any "new compute" claim

Maturity

Lab result, prototype, pilot or product?

Manufacturing

Can existing fabs make it, at what yield?

Energy per useful result

Including converters, cooling and I/O

Programming model

Does today's software run on it?

Error and precision

Compared on the same task as a digital chip