femtoAI
SPU — Competitive Landscape Confidential
femtoAI · Investor Diligence Materials
SPU Platform — Competitive Analysis
The edge AI chip market splits into chips that are too weak (Syntiant, AON — wakeword only) and chips that are too large and power-hungry (Hailo, Mythic — vision NPUs designed for cameras, not ears). femtoAI SPU occupies the white space: purpose-built audio AI inference silicon covering SPU001 through SPU2xx, with no external DRAM and hearable-class die area. All data sourced from public competitor datasheets.
100K+
Effective TOPS/W (SPU221)
6.7mm²
Smallest die in class (SPU210)
0
External DRAM required
700M
Max params, DRAM-free
Filter — Select Devices to Display Show All  ·  Hide All
Quick filter by DRAM requirement
Interactive Chart
Efficiency Frontier
Full Comparison
All Chips — Data Table
Company / Device Type DRAM Model Class eGOPS/W eGOPS/mm² Die Area Power Source
Why It Matters
femtoAI SPU Key Differentiators
100K
Effective TOPS/W10–100× efficiency lead over audio-adjacent competitors on target workloads.
6.7mm²
Smallest die in classvs. 289mm² (Hailo-8L). Fits inside TWS earbuds and hearing aids.
0
External DRAM required700M-param models via on-chip SRAM + PSRAM. Competitors all require DRAM at this scale.
5×
Simultaneous workloadsWakeword + SLU + noise reduction + AEC + speaker ID — no competitor handles all five at this power.
✓
Silicon in productionSPU001 is in mass production, in-hand with customers. SPU150 is sampling. Not a roadmap — silicon exists and is shipping.
18mo
Software stack leadCompiler + audio model library + reference designs. Moat is the full stack, not just the chip.
Why eGOPS/W matters more than raw TOPS: Raw TOPS numbers measure peak compute under ideal conditions. eGOPS (effective GOPS) accounts for memory bottlenecks — whether a chip can sustain throughput when models don't fit on-chip. femtoAI's PSRAM architecture removes the memory wall constraining every competitor in the sub-10mW class. Syntiant NDP250: 30 raw GOPS. femtoAI SPU240: 2,400 raw GOPS → 240,000 eGOPS. That 8,000× eGOPS gap vs. a 80× raw GOPS gap explains why customers evaluating Syntiant hit a throughput ceiling and switched.