Compute-in-memory: implementation evidence

← Return to the implementation matrix

Compute-in-memory performs an operation where data is stored, reducing movement between a processor and separate memory. The row credits integrated demonstrations, not every accelerator with a large cache.

Stage Mark Summary
Reference Executable models and mapped neural workloads
Physical Fabricated phase-change-memory compute arrays
Integrated 64 cores, digital functions, and on-chip network
Scaled Multicore chip, not a general large system
Access No externally programmable system located
Operational Complete experimental inference tasks

Reference — demonstrated

Claim
The matrix credits compute-in-memory at the Reference stage: executable models and mapped neural workloads.
Evidence
Toolchains map matrix-vector operations and neural-network layers onto analog memory arrays while modeling precision and device behavior. These executable mappings define what the hardware is expected to compute.
Criticism
Executable semantics do not establish purpose-built hardware, integration, scale, external access, recurring use, or comparative advantage.
Sources
IBM’s 64-core mixed-signal in-memory-compute chip

Physical — demonstrated

Claim
The matrix credits compute-in-memory at the Physical stage: fabricated phase-change-memory compute arrays.
Evidence
IBM reports a chip fabricated in 14 nm CMOS with backend-integrated phase-change memory. Computation occurs through the physical response of the memory arrays rather than by fetching every weight into a separate arithmetic unit.
Criticism
A physical realization does not by itself establish system integration, efficient scaling, external access, recurring use, or comparative advantage.
Sources
IBM’s 64-core mixed-signal in-memory-compute chip

Integrated — demonstrated

Claim
The matrix credits compute-in-memory at the Integrated stage: 64 cores, digital functions, and on-chip network.
Evidence
The chip combines 64 analog in-memory-compute cores with an on-chip communication network, digital activation functions, and processing for convolutional and recurrent layers. That integration is the strongest evidence in this row.
Criticism
A coherent system does not by itself establish efficient scaling, external access, recurring use, or comparative advantage.
Sources
IBM’s 64-core mixed-signal in-memory-compute chip

Scaled — limited

Claim
The matrix credits compute-in-memory only partially at the Scaled stage: multicore chip, not a general large system.
Evidence
Sixty-four cores establish multicore composition on one chip. The mark remains partial because conversion, calibration, weight updates, host coordination, and larger-system communication have not yet been shown as a broadly scalable computer.
Criticism
The mark is limited on the current public record: multicore chip, not a general large system. Composition at the reported scale does not establish useful scaling across workloads, favorable economics, or comparative advantage.
Sources
IBM’s 64-core mixed-signal in-memory-compute chip

Operational — limited

Claim
The matrix credits compute-in-memory only partially at the Operational stage: complete experimental inference tasks.
Evidence
The device repeatedly executes complete ResNet and long short-term-memory inference workloads with near-software-equivalent accuracy in experiments. This is more than an isolated memory-cell operation, but it is not a recurring external deployment.
Criticism
The mark is limited on the current public record: complete experimental inference tasks. Recurring work does not establish workload generality, independent reproduction, favorable economics, or comparative advantage.
Sources
IBM’s 64-core mixed-signal in-memory-compute chip

Stage not credited

No generally available system was located on which outsiders can load and run arbitrary supported workloads; the Access cell therefore remains a dash.

Source