By Dr. Jonathan Tapson, Chief Development Officer

The fastest execution path in trading and the most useful machine learning don’t have to live on the same hardware.

The tightest loop in high-frequency trading — tick-to-trade — belongs to FPGAs wired directly to the network interface. Some FPGA trading systems report sub-microsecond tick-to-trade latency — Algo-Logic describes a sub-500-nanosecond solution — while software, even kernel-bypassed, lands in the microseconds. What that tier doesn’t do is reason: turn an irregular firehose of quotes, trades, and order-book updates into better decisions, within a power budget that fits a colocation rack.

That is where Akida could fit — not as a replacement for the FPGA fast path, but as a second tier: a smart signal layer. To be clear up front: this is a proposed application architecture, not a benchmarked Akida trading deployment.

Market data arrives as an irregular stream of discrete events, not fixed frames. Akida’s event-based architecture computes only when activity occurs and skips zero-valued activations. One honest caveat: quotes and trades arriving asynchronously don’t automatically produce sparse neural activations — the feed handler and feature encoding determine how much sparsity the model actually sees. The architecture rewards good encoding; it doesn’t create sparsity for free.

Where the fit is real is in the model class. Akida 2.0 natively supports Temporal Event-Based Neural Networks (TENNs) — spatial-temporal convolutions with a recurrent inference mode — and BrainChip’s platform supports state-space models. Those are the architectures that map naturally to time-series work, with candidate applications in short-horizon price prediction, order-flow imbalance detection, microstructure anomaly spotting, and regime classification. The platform supports 8-, 4-, and 2-bit parameters, which suits small, fast networks — but low-bit accuracy is model-specific, and quantization to 4-bit or below must be validated on the target model, not assumed.

Distributing independent models across Akida devices could add inference capacity and specialization; throughput and total system power need to be measured together. A proposed multi-device implementation might run a different model per instrument, venue, or strategy — or use one device for feature extraction and another for classification, accepting the extra transfer that a pipeline implies — subject to validation of the target hardware and runtime. The existing trading stack retains risk and execution controls; the host fuses results.

Assigning independent models to dedicated devices can reduce accelerator-side contention. End-to-end latency still depends on host scheduling, transfers, feed bursts, and device coordination — and in trading, tail latency often matters more than the average, because one slow inference can mean a missed or adverse fill. Low power per device is what lets you fit many accelerators into a rack where power and space are priced at a premium.

Be clear-eyed about the boundary. A PCIe hop plus host software overhead will not fit a hundreds-of-nanoseconds tick-to-trade budget. The microsecond-to-low-millisecond timescale is the intended tier for this approach: a signal layer that would update parameters, regime flags, or probabilistic signals asynchronously, leaving the FPGA execution path independent of each new inference — or driving medium-frequency and market-making strategies where inference quality matters as much as raw speed.

We would not ask anyone to take this on faith. Before any deployment, benchmark: end-to-end latency and its tail under bursty feeds, model quality on time-ordered, held-out tick data, concurrent-device scaling, and host-plus-accelerator power. And remember that classification accuracy alone is not evidence of a profitable trading system — it is one input among many.

Akida’s opportunity is not to replace the FPGA execution path. It is to support a separate ML signal layer — with model quality, end-to-end timing, scaling, and system power validated on the target hardware.

If you’re evaluating ML for a trading stack and want to try Akida on your own market data, get in touch.