XimplicXimplic

We compute inside the memory

Conventional chips spend most of their energy shuttling data between memory and logic. Ximplic does the computation inside the memory itself, so always-on AI runs on a fraction of the power.

The energy goes into moving data, not using it

In a conventional design, memory and compute sit apart, and data is moved between them for every operation. On small models that run constantly, that movement, not the calculation, dominates the energy.

Ximplic

Compute in memory
  • Data is used where it is stored
  • Almost no movement between blocks
  • Energy tracks the useful work
  • Standard memory, buildable today

Conventional

Memory apart from logic
  • Data moves for every operation
  • Movement dominates the energy
  • Battery drains on always-on models
  • Designed around constant movement

The memory does the computation

The model lives inside the memory array, and the array computes directly on it. Data goes in, the result comes out, and almost nothing moves in between.

  • Memory and compute are one block
  • Computation happens where the data is stored
  • Far less movement, far less power
  • Standard SRAM cells, designed to port across foundries (silicon in development)
Datagoes inMemory that computescomputed in placeResult
Data goes in, the array computes in place, and only the result comes out.

Most teams build compute-in-memory by hand

Hand-building one is a slow, specialised effort. Our compiler maps any trained model onto the array and validates it automatically, so a proven block takes weeks, not a multi-year custom project. The same flow carries any model from software to RTL without redoing the work by hand.

Where the energy goes

The computation itself costs about the same either way. The difference is the data movement around it. Remove the movement and the total energy drops sharply.

Energy one AI prediction burns. The math is the actual computation; everything else is energy wasted moving weights between memory and compute. Longer bar means more total energy.

the mathdata movement
today
with Ximplic

Same model, a fraction of the power. The energy that used to move data now does the math.

Illustrative split, not measured values. Per-workload figures come from Ximplic Vyzora.

Measured, not promised

Proven on real hardware

The design runs end to end on an FPGA today, with results correlated to the hardware design and calibrated against board measurement. Here's what's running now, and where we're upfront that a finished silicon chip is still to come.

Runs on an FPGA

The full design works end to end on an FPGA. Board matches simulation.

90%+ across three model families

Measured on hardware for speech, anomaly detection and spiking neural networks.

Measured, not promised

Reports separate measured results from targets, with detail under NDA.

Validated
FPGA-proven
The full design runs end to end on an FPGA today, not a simulation.
Architecture
SRAM in-memory
Built from standard digital SRAM cells, no exotic process.
Integration
AMBA AXI
Drops into your chip over a standard bus, like any other IP block.
Models
Any AI model
Bring any trained ONNX network; the compiler maps it onto the array.

Numbers for a specific workload come from the Vyzora. See also how it integrates into a chip.

Talk to us

Check the numbers on your own workload

The design you just read about runs on an FPGA today. Evaluate it under NDA on your own model, judge the measurements yourself, then license the same proven design into your SoC.

Email info@ximplic.com · Groningen, The Netherlands