Specialist Edge AI Silicon: Coral, Ambarella, SiMa.ai, Kneron and Beyond
Overview
Beyond the three challengers this knowledge base has already covered in depth, a second rank of specialist vendors builds edge AI silicon with sharply different bets: Google’s Coral Edge TPU (minimalist fixed-function INT8), Ambarella (vision SoCs with class-leading ISPs), SiMa.ai (whole-pipeline MLSoCs), Kneron and DEEPX (low-power camera NPUs), EdgeCortix (reconfigurable dataflow for transformers) and Mythic (analog compute-in-memory). None of these is a general-purpose platform; each wins inside a specific envelope of power, workload and integration style. This survey maps what each optimises for, with vendor-stated figures flagged as such, so you can shortlist realistically.
Key takeaways
- Google Coral delivers a reported 4 INT8 TOPS at around 2 W and remains fine for simple TFLite vision, but the ecosystem has visibly aged and should not anchor new designs.
- Ambarella pairs neural acceleration with best-in-class ISPs: CV7-class SoCs run vision and GenAI under a vendor-stated 5 W, while the N1 series targets on-prem GenAI under a vendor-stated 50 W.
- SiMa.ai’s MLSoC and Modalix put ISP, DSP, Arm cores and ML acceleration on one chip, with Modalix vendor-stated at 25 to 200 INT8 TOPS.
- DEEPX DX-M1 (vendor-stated 25 INT8 TOPS at 2-5 W) and EdgeCortix SAKURA-II (vendor-stated 60 INT8 TOPS at about 8 W, with BF16) are the strongest new M.2-class options.
- Mythic’s analog compute-in-memory is genuinely different physics: vendor-stated 25 TOPS at about 3 W, with maturity and toolchain risk to match its novelty.
Google Coral: the minimalist Edge TPU
The Coral Edge TPU executes INT8 TensorFlow Lite models at a reported 4 TOPS for roughly 2 W, packaged as USB sticks, M.2 and mini-PCIe modules and dev boards. It proved that tiny fixed-function silicon could handle real vision workloads, and thousands of deployments (including many hobbyist and NVR projects) still run on it. The honest position in 2026: the model zoo and tooling have stagnated, transformer support is absent, and its role now is maintaining existing fleets rather than anchoring new designs. It remains a useful baseline for what 2 W can do.
Ambarella: vision SoCs where the ISP is the point
Ambarella came from video processing, and its CVflow SoCs integrate a strong ISP, video codecs and neural acceleration on one die: the natural architecture for smart cameras, where image quality upstream of the network determines accuracy as much as TOPS. The 5 nm CV72S targets mainstream 4K security cameras on CVflow 3.0 with dual Cortex-A76 cores, and Ambarella states GenAI capability under 5 W on CV72-class parts. At the top, the N1 SoC series targets on-prem generative AI, with the vendor stating Llama2-13B at up to about 25 tokens per second under 50 W. Fit: camera-shaped products end to end; less suited to accelerating an existing x86 box.
SiMa.ai: the whole pipeline on one chip
SiMa.ai’s MLSoC argument is that real applications are pipelines, not models: its first-generation part combines Arm cores, ISP, DSP and an ML accelerator delivering a vendor-stated 50 INT8 TOPS in a single-digit-watt envelope. The second-generation Modalix family is vendor-stated at 25, 50, 100 and 200 INT8 TOPS variants aimed at multimodal and transformer workloads, with the 50-TOPS part stated to run Llama2-7B at more than 10 tokens per second at 8-10 W. Fit: teams that want one chip instead of host-plus-accelerator, and are willing to adopt SiMa’s Palette toolchain.
Kneron and DEEPX: low-power camera NPUs
Taiwan’s Kneron ships reconfigurable low-power NPU SoCs (the KL730 is vendor-stated in the single-digit-TOPS class) for smart-home, access-control and automotive-adjacent vision where cost and milliwatts dominate. Korea’s DEEPX DX-M1 is the more direct Hailo-8 competitor: vendor-stated 25 INT8 TOPS in a 2-5 W envelope on M.2 modules, aimed at drones, robots and multi-camera boxes. Both compete on price per stream against the challengers profiled in Hailo vs Axelera vs Blaize.
EdgeCortix and Mythic: transformers and analog physics
Japan’s EdgeCortix SAKURA-II is a reconfigurable dataflow accelerator vendor-stated at up to 60 INT8 TOPS in roughly 8 W, notable for BF16 support and explicit targeting of transformer and GenAI workloads on M.2 and PCIe cards (dual-chip cards vendor-stated up to 120 INT8 TOPS). Mythic takes the most radical path: its M1076 analog matrix processor computes multiply-accumulates inside flash memory arrays, vendor-stated at up to 25 TOPS at about 3 W, with a 16-chip PCIe card stated at up to 400 TOPS at 75 W. Analog compute-in-memory offers striking efficiency, but note that analog TOPS are not precision-labelled like digital INT8 figures, and the toolchain and ecosystem are the least proven in this survey.
Specialist platforms at a glance
| Platform | Architecture bet | Vendor-stated compute | Power band | Realistic fit |
|---|---|---|---|---|
| Google Coral Edge TPU | Fixed-function INT8 ASIC | About 4 TOPS (INT8, reported) | About 2 W | Legacy TFLite vision fleets, hobbyist NVR |
| Ambarella CV72-class | Vision SoC with ISP + codecs | Not directly TOPS-rated by vendor | Under 5 W | 4K smart cameras, video-first products |
| Ambarella N1 | GenAI vision SoC | Llama2-13B about 25 tok/s (vendor-stated) | Under 50 W | On-prem video plus GenAI appliances |
| SiMa.ai MLSoC / Modalix | Whole-pipeline MLSoC | 50 / 25-200 TOPS (INT8) | 5-10 W | Single-chip vision and multimodal systems |
| Kneron KL-series | Low-power reconfigurable NPU SoC | Single-digit TOPS class | Sub-2 W class | Smart home, access control, cost-first vision |
| DEEPX DX-M1 | Camera NPU on M.2 | 25 TOPS (INT8) | 2-5 W | Drones, robots, multi-camera edge boxes |
| EdgeCortix SAKURA-II | Reconfigurable dataflow, BF16 | 60 TOPS (INT8) per chip | About 8 W | Edge transformers and GenAI on M.2/PCIe |
| Mythic M1076 | Analog compute-in-memory | 25 TOPS (analog, vendor-stated) | About 3 W | Power-critical vision, experimental designs |
How to shortlist among specialists
Match the bet to your product shape: camera products favour ISP-integrated SoCs (Ambarella, SiMa.ai); retrofit accelerator cards favour M.2 parts (DEEPX, EdgeCortix, and the challengers covered in the Hailo deep dive); transformer roadmaps favour BF16-capable or large-memory parts. Then test your actual model through the vendor’s compiler before committing, because operator coverage varies far more among specialists than among incumbents; the process is laid out in the selection framework, and the wider market context sits in the landscape hub.
Frequently asked questions
Is Google Coral discontinued?
Coral hardware remained purchasable through distributors at the time of writing, but toolchain and model-zoo updates have largely stalled for years. Treat it as maintenance-mode silicon: fine for existing fleets, risky as the anchor of a new multi-year product.
Why does an integrated ISP matter for edge AI?
Detection accuracy depends on the image the network sees. An SoC-grade ISP handles HDR, low light and noise before inference, often improving real-world accuracy more than extra TOPS would. It also removes a discrete ISP from the bill of materials.
Which specialists can run LLMs or multimodal models?
Among this group, Ambarella’s N1 series, SiMa.ai’s Modalix and EdgeCortix’s SAKURA-II publish explicit GenAI and transformer claims. Small fixed-function INT8 parts such as Coral, Kneron’s low-end SoCs and DX-M1 remain CNN-first.
Are analog TOPS comparable to INT8 TOPS?
Not directly. Analog compute-in-memory performs multiply-accumulates in the analog domain, so its effective precision differs from digital INT8 arithmetic. Compare candidates by accuracy and throughput on your own model, never by the headline number.
How risky are smaller specialist vendors for a 7-year product?
Assess funding runway, published longevity commitments and second-source options honestly. Several specialists offer 10-year availability statements, but a small vendor’s commitment is only as strong as the company. Dual-qualifying an alternative module is prudent for long-lived fleets.
Ready to deploy?
Talk to an RDP architect about power, cooling and lead time.