← TRIGLER Lab
Silicon Strategy — Reference

Accelerator Comparison

Whether an updatable soft CPU core / eFPGA fits Kestrel's power envelope — and why the same fabric is a better fit for an AI or robotics platform than for a gaming handheld.

Document KST-ACC-001
Revision A
Issued 2026-09-15
Status Reference — engineering estimate
1.15–2.05W
Kestrel SoC power budget
20–400MHz
Reconfigurable soft-core ceiling
<300µW
QuickLogic EOS S3, always-on class
2roles
Fits as coprocessor / AI accelerator, not main CPU
01

Why This Doc Exists

Kestrel Silicon KX2 (the concept ARM part in the architecture doc) doesn't exist yet — same problem the SteamOS POC works around for launcher/UX validation. An "updatable soft CPU core on FPGA" looked like a possible way to sidestep committing to fixed silicon for compute too: build the main processor as reconfigurable fabric, patch it post-ship instead of taping out again. This doc checks whether that actually works for Kestrel, and where it does work.

Short answer: no for Kestrel's main CPU, yes for a different product shape. Both halves are below.
02

The Soft-Core Tradeoff

A soft core runs on generic LUTs and reconfigurable routing instead of custom-laid-out transistors — that generality is what makes it reconfigurable, and it costs on two axes from the same root cause: more transistors switch per logic operation than a hardened gate doing the same job (power), and the same routing adds propagation delay to the critical path (max clock frequency). The two aren't independent problems — pushing clock speed to compensate for weak throughput just makes the power problem worse.

PartReconfigurable ("soft") modeHardened mode on the same die
Efinix Sapphire RISC-V SoC20–400 MHz, patchableUp to 1 GHz quad-core, fixed at fab time
The catch: "updatable" and "fast enough to drive a modern OS/UI" sit on opposite ends of the same tradeoff on this chip family. Pick reconfigurable, you get MCU-class speed — nowhere near a mobile Cortex-A73/A55-class core. Pick the fast core, you've given up updatability entirely; it's fixed silicon next to the fabric, not the fabric itself.
03

Vendor Landscape

CompanyProductCategoryNotes
AMD (Xilinx)MicroBlaze VSoft-core IPRISC-V soft core for adaptive SoCs/FPGAs
Intel (Altera)Nios VSoft-core IP~5× perf of Nios II/e, still soft-core class
MicrochipMi-V ecosystemSoft-core IP + hard RISC-VPolarFire SoC pairs hard RISC-V cores with fabric
EfinixSapphire RISC-V SoCSoft + hard coreSee §02 — same family, two very different tradeoffs
LatticeSiFive E2 (partner IP)Soft-core IPTargets Lattice FPGAs via SiFive collaboration
AchronixSpeedcore eFPGAeFPGA-in-ASICLicensable fabric embedded directly in a custom ASIC
Flex LogixEFLX1K / EFLX4KeFPGA-in-ASICTileable eFPGA cores for ASIC/SoC integration
QuickLogicEOS S3MCU + small eFPGACortex-M4F + eFPGA glue logic, always-on voice/sensor class
AMD (Xilinx)Kria K26 SOMHard CPU + fabricZynq UltraScale+: hard Cortex-A53/R5 + FPGA, ROS 2 support
04

Why It Fails Kestrel's Envelope

Checked against the SoC power budget in the architecture doc's §03 (1,150 mW mixed gaming / 2,050 mW peak / 480 mW video) and the ~45 mW whole-system suspend baseline.

CandidateFits the power budget?Why it still doesn't work as the main CPU
Efinix Titanium/Sapphire (soft mode)Unverified — no published Icc tableEven if power fits, 20–400 MHz is too slow to drive a modern gaming UI/OS smoothly
QuickLogic EOS S3Trivially, <300µWWrong product class entirely — an MCU+sensor-fusion part, not an application processor
Where either one might still earn a place in Kestrel: as a tiny always-on coprocessor for the ~45 mW suspend/quick-resume state, not as KX2's replacement — wake-on-button, wake-on-network, or sensor polling while the main SoC is fully suspended, field-updatable without touching the main silicon.
05

Why It Fits an AI/Robotics Platform Instead

The fix isn't a better chip generation — it's that AI/robotics workloads don't route through a soft CPU's scalar instruction pipeline at all. The natural implementation is dataflow: dedicated parallel pipelines built directly in the fabric, next to a hardened CPU doing orchestration.

This half of the analysis has since grown into its own concept project — Raptor Core (separate repo, sibling to this one) — rather than staying a Kestrel side-section. The summary below is kept here for context; the current spec lives there.
Gaming handheld (Kestrel)
Branchy scalar code
OS, launcher UI, Proton translation
Soft CPU core
Worst case for LUT/routing fabric — general instructions, one at a time
AI/robotics platform
Orchestration (ROS nodes, planning)
Runs on a hardened CPU core, not the fabric
Dataflow pipelines in fabric
NN inference (DSP-block MAC/BFLOAT16), motor/sensor control loops, sensor bridging — parallel hardware, not interpreted instructions
WorkloadWhy the fabric helps
NN inference / sensor fusion mathDense MAC is what DSP blocks are built for — e.g. Efinix Ti60 has 160 DSP blocks, each capable of INT8 quad-multiply or a fused BFLOAT16 multiply-add in hardware
Motor/sensor control loopsCycle-accurate, deterministic timing — no OS/RTOS scheduler jitter, a real capability advantage independent of power
Sensor bridgingMany simultaneous parallel I/O interfaces without contending for a shared bus/DMA like an MCU would
PartRole in an AI/robotics platform
AMD Kria K26 SOMHard Cortex-A53/R5 + fabric, ROS 2-ready — the reference heterogeneous pattern
Microchip PolarFire SoCHard RISC-V + fabric, real-time control + reconfigurable I/O as the core pitch
Efinix SapphireSame chip as §02, but here the BFLOAT16 DSP path is the point, not soft-core clock speed
Lattice CrossLink-NX / CertusPro-NXLow-power edge AI/vision sensor bridging
QuickLogic EOS S3Legitimate fit here — always-on sensor/voice wake is its actual target market, not a stretch
06

Recommendation

For Kestrel K1
Don't use soft-core FPGA as the main CPU — the power/clock-speed tradeoff in §02 is structural, not a maturity gap. A small eFPGA as an always-on suspend/wake coprocessor is the one role worth revisiting.
For an AI/robotics platform
Heterogeneous hardened-CPU + reconfigurable-fabric (Zynq/Kria/PolarFire-style) is a legitimate architecture, not a workaround — the fabric does dataflow work it's actually good at, and the CPU stays hardened for orchestration.
Scope note
This is a separate product shape from Kestrel, not a pivot of it — the workload shape, not the silicon budget, is what makes the fabric a good idea here. It now lives as its own project, Raptor Core.
Kestrel K1 — Accelerator Comparison Rev A Figures are public datasheet specs and engineering estimates, not a sourced teardown