IN-MEMORY COMPUTING

Do the math where the data lives

In an analog crossbar, each weight is stored as a conductance G at the crossing of two wires. Apply input voltages V to the rows: by Ohm's law each cell passes a current I = G × V, and by Kirchhoff's law the currents add up on each column. A whole matrix-vector multiply happens in one step, without moving any weights.

Physics does the multiply-accumulate, so the memory wall disappears for the weights.

THE CATCH

Noise and drift

Analog values are noisy and drift over time and temperature, and converting results back to digital (ADCs) costs area and energy. Analog chips usually reach only a few bits of effective precision. Companies such as IBM and Mythic, and many university labs, work with resistive memory, phase-change memory and flash cells.

Analog is very efficient, but getting it accurate enough is the hard part.

NEUROMORPHIC

Chips that spike

Neuromorphic chips imitate the brain more closely: neurons send short spikes only when something changes, and sit idle otherwise. IBM's TrueNorth (2014) had a million neurons running on about 70 mW; Intel's Loihi chips are research platforms for spiking networks.

The brain runs on about 20 watts, which keeps inspiring chip designers.

PHOTONICS

Computing and talking with light

💡

🔀

Photonic math

Interferometer meshes can do matrix multiplies with light

📡

Optical links

Co-packaged optics put lasers and fibers right beside the chip to move data with less energy

🏁

Status

Optical links are arriving in AI networks first; optical computing is still mostly in research

Light is already moving AI's data; computing with it is still ahead.

MATCH THEM

Which future chip?

Match each description with the kind of chip it describes.

Each approach attacks the memory wall or energy cost in a different way.

UNIT 20 STUDY COMPLETE

🔮

Ready for the Fab Challenge?

You've seen how a crossbar does a whole matrix multiply with Ohm's and Kirchhoff's laws, and met spiking and photonic chips.