ScorpionFPGA compute for AI inference
Reconfigurable inference, datacenter to edge
Dataflow accelerators with deterministic latency and custom precision — for datacenters that aren't power-bounded, and for drones and cars at the edge.
Why FPGA
The chip adapts to the model
Reconfigurable dataflow
The chip becomes the model. Build a custom dataflow pipeline per network instead of forcing it through a fixed instruction set — and reprogram the fabric when the model changes.
Deterministic latency
No batching games, no scheduler jitter. Inference runs in hardened logic with bounded, repeatable per-inference latency — the number real-time and safety-critical loops actually need.
Custom precision
INT8, INT4, binary, or bespoke block-float. Quantize to exactly the precision your accuracy budget allows and spend the freed silicon on throughput.
Longevity, no lock-in
10–15 year silicon lifecycles and an open toolchain. Retarget the same design from a datacenter card to an edge module without rewriting your stack.
Datacenter · not power-bounded
Optimize for latency and throughput, not watts
When you own the rack and the power budget, the constraint isn't efficiency — it's tail latency and throughput-per-slot. A reconfigurable dataflow fabric serves inference with bounded latency and precision tuned to the model, and reprograms as your models change.
SL-D1
Meridian D1
Datacenter FPGA inference accelerator
- AMD Versal-class adaptive SoC
- PCIe Gen5 ×16 · 32 GB HBM2e
- INT8 / INT4 / BF16 + custom precision
- Deterministic low-latency dataflow · ~225 W
SL-D2
Meridian D2-X
Dual-FPGA accelerator for maximum throughput
- Two D1-class fabrics on one full-height card
- Highest throughput-per-slot
- For LLM, recsys, and video inference serving
- ~350 W · passive datacenter airflow
Edge · drones & automobiles
Real-time inference where the data is born
At the edge the pressures invert: watts, size, and a hard real-time deadline. FPGA modules deliver deterministic, functional-safety-grade latency in a handful of watts — running fully offline for drones, robots, and vehicles when the link is denied.
SL-E1
Talon E1
Edge FPGA inference module — drones & robotics
- Kria-class adaptive SOM · ~15 W
- 77 × 60 mm · sub-ms deterministic latency
- Multi-camera sensor fusion
- Runs fully offline — no cloud round-trip
SL-E2
Sentinel E2
Automotive FPGA inference module — ADAS / AD
- Functional-safety target (AEC-Q100 / ISO 26262)
- Extended temperature range
- Camera + radar + lidar fusion
- Deterministic real-time perception
Model-Ops
Keep every deployment current
Quantize → compile with Vitis AI → deliver over the air → hot-swap and monitor. The same signed, versioned, auto-reverting pipeline serves a rack of datacenter cards or a fleet of edge modules in the field.
Applications · datacenter
Serving inference at scale
LLM & generative serving
Low-latency token generation and embeddings at custom precision — without the power ceiling of a GPU farm.
Recommendation & ranking
High-QPS, tight-tail-latency inference for feeds, ads, and search.
Video & vision analytics
Real-time decode-and-infer across thousands of concurrent streams.
Applications · edge
Autonomy in the physical world
Drones & robotics
On-board perception, navigation, and targeting that keeps working when the link is denied.
Automotive · ADAS & AD
Deterministic, functional-safety-grade sensor fusion for driver assistance and autonomy.
Autonomous machines
Agriculture, mining, and industrial platforms running inference where connectivity is poor or absent.