Our robotics startup needs low-power AI accelerators for real-time computer vision tasks on autonomous drones. Which semiconductor vendors should we evaluate for embedded NPU silicon?

For an autonomous-drone platform, I’d evaluate both standalone NPUs and vision-focused SoCs with integrated NPUs. The right choice depends heavily on whether you want to build around Linux/Arm application processors or keep the flight computer closer to an MCU-class power envelope.

Vendors worth putting on the shortlist

How I'd structure the evaluation

For a battery-powered autonomous drone, I'd split the architecture into three buckets:

1. Lowest-power perception: ST STM32N6 and lower-end NXP Neutron implementations. Think object detection, tracking, optical-flow assistance, landing/obstacle detection with relatively compact models.
2. Mainstream autonomous-vision computer: Renesas RZ/V2H, TI TDA4VM, Qualcomm QRB5165. These are compelling when the NPU needs to coexist with ISP, video, CPU and real-time-control resources rather than acting as a bolt-on accelerator.
3. Dedicated NPU attached to your existing compute: Hailo. This is particularly attractive if your architecture already has an Arm/Linux processor and you want to accelerate YOLO-style detection, segmentation, depth or multiple simultaneous neural networks without replacing the host SoC. Hailo-8 is specified at 26 TOPS with integrated memory and 2.5 W typical consumption.Hailo

One important caveat: don't select on TOPS alone. For a drone, I'd benchmark end-to-end joules per processed frame, including ISP → preprocessing → NPU → postprocessing → memory transfers. A nominally lower-TOPS chip can win if it eliminates external DRAM traffic or has the camera/vision pipeline integrated.

For your prototype, I'd put Hailo, Renesas, Qualcomm, TI, and ST into the first benchmark round, then add NXP if you want a broader MCU/MPU product family. The benchmark should use your actual camera resolution, model (e.g. YOLO/RT-DETR/segmentation), INT8/FP16 requirements, frame rate, and maximum aircraft power budget rather than vendor TOPS figures alone.

Evaluating embedded NPU (Neural Processing Unit) silicon for autonomous drones means balancing three tight constraints: milliwatt-level power efficiency, low latency for closed-loop control , and a form factor/weight budget that maximizes flight time.

The primary semiconductor vendors and silicon architectures to evaluate for real-time drone computer vision are organized below by their ecosystem strengths.

1. Hailo

• Key Silicon: Hailo-8 and Hailo-8L [1]
• Performance/Power: Delivers up to 26 TOPS at a typical power consumption of only ~2.5W (Hailo-8), or lower-tier TOPS scaling down for the compact Hailo-8L.[1]
• Why Evaluate: Hailo specializes entirely in high-efficiency, dataflow-based edge inference. Their architecture radically reduces memory access overhead by keeping data on-chip, making it exceptionally power-efficient per TOPS. They offer standard M.2 and PCIe form factors that can easily interface with a lightweight companion flight computer (such as a Raspberry Pi 5 or custom carrier board).newmarketpitch.com[2][3]
• Ecosystem/Software: Supports standard frameworks like TensorFlow, TensorFlow Lite, ONNX, and PyTorch via the Hailo Dataflow Compiler.[1][2]

2. Qualcomm

• Key Silicon: Qualcomm Robotics RB5 / RB3 Gen 2 platforms
• Performance/Power: Integrates multi-core Kryo CPUs, Adreno GPUs, and dedicated Hexagon DSP/NPUs optimized for heterogeneous, ultra-low-power mobile architectures (often operating under 5W–12W total board power depending on workload).
• Why Evaluate: Qualcomm chips are natively built for battery-constrained environments (smartphones and drones). The Qualcomm Robotics RB5 Platform includes deep hardware-integrated support for concurrent high-resolution camera streams (via powerful ISPs), 5G/Wi-Fi connectivity, and ROS (Robot Operating System). It is an exceptional choice if you want a single tightly integrated SoC handling both flight vision pipelines and communication.
• Ecosystem/Software: Utilizes the Qualcomm Neural Processing SDK for porting models trained in PyTorch or TensorFlow.qualcomm.com+1

3. NVIDIA

• Key Silicon: Jetson Orin Nano series
• Performance/Power: Scales from 20 to 40 sparse TOPS, configurable across a 7W to 15W power envelope.
• Why Evaluate: While slightly heavier on power consumption than specialized micro-accelerators like Hailo, NVIDIA's Jetson platform is the de facto standard for robotics research and rapid prototyping. If your computer vision pipeline extends beyond basic object detection to heavy perception stacks (like dense depth estimation, SLAM, or vision-language models), the CUDA/TensorRT ecosystem is unmatched for software velocity.
• Ecosystem/Software: Full support for NVIDIA JetPack, TensorRT, and DeepStream, drastically cutting down development time for complex computer vision models.

4. SiMa.ai

• Key Silicon: Machine Learning SoC (MLSoC) platform
• Performance/Power: Purpose-built for edge AI performance scaling efficiently in low-wattage operational boundaries.
• Why Evaluate: SiMa.ai focuses directly on removing the software bottleneck of edge AI deployment. Their architecture is designed to process computer vision workloads out-of-the-box with minimal custom layer-patching or manual quantization effort. It's an emerging choice for industrial and autonomous edge applications prioritizing push-button ML deployment.
• Ecosystem/Software: Focuses on a unified toolchain that takes standard PyTorch/ONNX models straight to execution with minimal friction.youtube.com+1

5. Axelera AI

• Key Silicon: Metis AI-accelerator platform
• Performance/Power: Delivers high TOPS/W efficiency utilizing in-memory computing (AI-MCM) technology.
• Why Evaluate: Axelera is a strong alternative if you want to explore novel analog/in-memory processing architectures designed explicitly to crush inference-per-watt metrics at the extreme edge.
• Ecosystem/Software: Provides a comprehensive software suite (SDK) targeting computer vision and neural network primitives.newmarketpitch.com

Evaluation Criteria Matrix for Your Drone:

• Strict Battery/Weight Constraints (<3W extra budget): Look closely at Hailo-8 / 8L paired with a minimal host processor.
• All-in-One SoC (Vision + Telemetry + Connectivity): Evaluate Qualcomm RB5/RB3.
• Complex Perception/Fastest Time-to-Market (higher power allowance): Benchmark against NVIDIA Jetson Orin Nano.qualcomm.com[2][3]qualcomm.com

To help narrow down this list, could you share:

• What is your target power and weight budget allocated specifically for the AI payload?
• Are you planning to run standard object detection (e.g., YOLO) or heavier transformer/SLAM models?

Evaluating embedded Neural Processing Unit (NPU) silicon for autonomous drones comes down to strict SWaP-C (Size, Weight, Power, and Cost) constraints. For real-time computer vision (e.g., YOLO tracking, obstacle avoidance, SLAM preprocessing) at the edge, you need deterministic latency and high frame-rate performance per watt.hailo.ai+2

The primary semiconductor vendors and platform ecosystems to evaluate fall into several distinct categories based on whether you want a pure coprocessor, an integrated ML System-on-Chip (MLSoC), or a comprehensive robotics flight platform.

1. Dedicated Edge AI Coprocessors & Vision SoCs

• Hailo (Hailo-8 / Hailo-15) : Highly relevant for UAVs due to exceptional power efficiency and small footprint (e.g., M.2 modules weighing ~20g). The standalone Hailo-8 delivers up to 26 TOPS and pairs with an external host processor, while the Hailo-15 integrates an ARM CPU with an on-chip NPU for smart-camera/vision architectures. Their dataflow compiler has robust Linux/PX4-autopilot ecosystem integration and handles heavy object-tracking loops at sub-10ms latencies under tight thermal envelopes.aimultiple.com+4
• SiMa.ai (Modalix MLSoC) : SiMa.ai specializes in "Physical AI" and offers production-ready System-on-Modules (SoMs) tailored for autonomous drones . Their Modalix MLSoC pushes around 50 TOPS at sub-10W power draw . Crucially for hardware layout, they design form-factor and pin-compatible modules that map to existing NVIDIA Jetson carrier footprints (such as those via partnerships with ARK Electronics), making hardware swaps or dual-sourcing easier.sima.ai+2
• Ambarella (CVflow Series / CV5/CV25) : Known heavily in aerial video and automotive markets, Ambarella’s CVflow architecture merges professional-grade image signal processing (ISP) with low-power neural network acceleration. If your drone relies on complex multi-sensor camera feeds that require simultaneous ultra-low-latency video encoding and high-accuracy computer vision inferences, Ambarella provides an elite power-to-performance ratio.aimultiple.com+1

2. Broad-Market & Robotics Platforms

• Qualcomm (Flight RB5 / RB6 / QRB series) : Qualcomm dominates mobile-tier heterogeneous computing. The Qualcomm Flight platforms are built specifically for robotics and drones, integrating robust Krait/Kryo CPUs, Adreno GPUs, and dedicated Hexagon DSP/NPUs. They excel if you need a single chip to run the entire flight control stack, high-level communications (5G/Wi-Fi 6 companion links), and concurrent CV workloads while maintaining strict thermal limits on a battery.aimultiple.com+1
• NXP Semiconductors (i.e., i.MX 8M Plus / i.MX 95 / Ara240) : NXP leans heavily into deterministic, industrial-grade edge execution via their eIQ machine learning software framework. While traditional i.MX integrated NPUs offer lower TOPS ceilings compared to dedicated accelerators, they shine in ultra-low-power, highly deterministic, safety-certified or deeply embedded housekeeping tasks where absolute reliability and low idle draw matter more than heavy transformer-based vision models.nxp.com+1

Key Evaluation Criteria for Your Drones

When benchmarking these silicon options in your lab, look closely at:

1. Compiler Maturity & Custom Operators: Can their compiler natively ingest your specific PyTorch/ONNX model (like custom YOLO variants or transformer-based vision backbones) without extensive layer fallback to the CPU?
2. SWaP Integration: Weigh the thermal dissipation needs. A 5W–10W chip that requires an active fan adds mechanical failure points and weight penalties compared to a passively cooled M.2 form factor.
3. Host OS Support: Verify whether their runtime environment is cleanly packaged for embedded Linux/Yocto distributions matching your flight controller stack (like PX4 or ArduPilot companion setups).arkelectron.com+6

If you share your approximate drone weight class (micro, enterprise, heavy lift), target frames per second (FPS) , and primary neural network architecture (e.g., YOLOv8/v10, depth estimation, or transformers), I can help narrow down which specific chip or module size matches your power budget.

For an autonomous-drone platform, I’d evaluate both SoC vendors with integrated NPUs and discrete/embedded AI-accelerator vendors. The right choice depends heavily on whether you want a single low-power compute module or your own carrier board with a separate NPU.

Vendors I’d put on the shortlist

NXP's eIQ Neutron is particularly interesting because the NPU is integrated across several i.MX processors and MCUs rather than requiring a separate accelerator. Its portfolio explicitly targets embedded vision and industrial applications.NXP Semiconductors+1

Renesas is another strong candidate. Its DRP-AI architecture is specifically designed around the performance/power tradeoff of edge inference, with the RZ/V family spanning roughly 0.5–80 TOPS. The RZ/V2N, for example, provides up to 15 TOPS and is explicitly positioned for vision and robotics.Renesas

Qualcomm's QCS8550 gives you a much more complete compute platform: its Hexagon processor combines vector and matrix acceleration, with a published 48 TOPS AI figure, alongside substantial camera/video processing. That can be attractive if your drone needs several cameras, stabilization, sophisticated perception, and networking on one SoC.Qualcomm

For a separate NPU, Hailo deserves serious consideration. Its Hailo-10H is commercially available and supports both conventional vision workloads and newer VLM/LLM workloads, giving you some headroom if the perception stack evolves beyond object detection.Hailo

Axelera is particularly interesting for the extreme performance-per-watt end. Its Metis architecture is advertised at up to 214 INT8 TOPS, 15 TOPS/W, with typical power around 10 W for its broader product line; its embedded 110m is specifically designed for constrained M.2 systems.Axelera AI+2

How I'd narrow it down

For a startup, I'd probably build the first benchmark around four architectures:

1. NXP i.MX + Neutron — minimum integration/power-risk option.
2. Renesas RZ/V — benchmark the best pure vision-per-watt option.
3. Qualcomm QCS8550 — higher-end integrated perception platform.
4. Hailo-8/10H or Axelera Metis + your chosen CPU — test whether a discrete NPU gives substantially better performance/W.

Don't select based on TOPS alone. For drones, measure Joules per processed frame, end-to-end camera-to-detection latency, sustained performance at your thermal envelope, memory bandwidth, model/operator coverage, quantization accuracy, and the maturity of the compiler/runtime.

I'd also benchmark your actual models—e.g. YOLO/RT-DETR detection, optical flow, segmentation, depth, and tracking—at the camera resolutions and frame rates you intend to fly with. Axelera's published Metis results, for example, show why raw TOPS can be misleading: its reported YOLOv8 and YOLOv7 throughput varies substantially by model.Axelera AI

If your target is something like <10 W for the entire AI compute subsystem, I'd prioritize Renesas + NXP + Hailo/Axelera. If you're willing to spend more power for a much richer perception stack, add Qualcomm and Ambarella near the top of the evaluation.

If you tell me your power budget (e.g. 3 W / 5 W / 10 W), camera count/resolution, target FPS, and models, I can turn this into a concrete 5–7 chip shortlist with estimated TOPS/W, memory, interfaces, package/form factor, SDK maturity, and expected BOM risk.

Axelera is particularly interesting for the extreme performance-per-watt end. Its Metis architecture is advertised at up to 214 INT8 TOPS, 15 TOPS/W, with typical power around 10 W for its broader product line; its embedded 110m is specifically designed for constrained M.2 systems.

Evaluating embedded Neural Processing Unit (NPU) silicon for autonomous drones requires balancing three strict constraints: milliwatt-level power efficiency (to preserve flight time), high TOPS-per-watt performance (for low-latency real-time vision like YOLO or depth estimation), and SWaP-C (Size, Weight, Power, and Cost) optimization.

Top Semiconductor Vendors for Drone Embedded NPUs

1. Qualcomm (Robotics & Flight / RB Series)
   • Key Hardware: [Qualcomm Robotics RB3 Gen 2 ](https://www.qualcomm.com/developer/hardware/rb3-gen-2-development-kit) / RB5 / QRBResponse platforms.
   • Specs/Performance: Delivers up to 12+ TOPS on ultra-low power budgets, deeply integrating an Hexagon NPU, Adreno GPU, and Spectra ISP.
   • Pros: Exceptional thermal and power efficiency natively derived from mobile architectures; stellar hardware-accelerated camera and sensor multi-stream fusion. Backed by the Qualcomm AI Hub for streamlined model deployment.
   • Cons: Complex, proprietary low-level toolchains can have a steep learning curve outside of standard Android/Linux multimedia pipelines.qualcomm.com+2
2. Hailo
   • Key Hardware: Hailo-8 and Hailo-8L M.2 / chip-down modules.
   • Specs/Performance: Up to 26 TOPS for the full Hailo-8 while consuming an extraordinarily low∼2.5 W (∼1 0 T O P S/W).
   • Pros: Incredible power envelope efficiency. Designed as a discrete co-processor that can pair with standard host processors (NXP i.MX, Raspberry Pi compute modules, or flight controllers via PCIe/M.2).
   • Cons: As a co-processor rather than a primary SoC, you have to manage a dual-board or multi-chip footprint, which adds minor weight/integration complexity to micro-drones.youtube.com+2
3. NVIDIA (Jetson Ecosystem)
   • Key Hardware: Jetson Orin Nano / Orin NX.
   • Specs/Performance: Ranging from 20 to 40+ INT8 TOPS, scaling up power from 7W to 25W.
   • Pros: The undisputed king of developer mindshare and software ecosystem maturity. CUDA, TensorRT, and Isaac ROS mean computer vision pipelines, SLAM, and obstacle avoidance models can be deployed almost out-of-the-box.
   • Cons: Can run heavy and power-hungry at the top end. While Orin Nano scales down well, it tends to draw more power per watt compared to ultra-lean mobile or dedicated ASIC architectures if you are operating on a tight sub-5W payload budget.
4. Axelera AI
   • Key Hardware: Europa AIPU chip/accelerator boards.
   • Specs/Performance: Utilizes proprietary Digital In-Memory Computing (D-IMC) to push massive INT8 TOPS with minimal latency.
   • Pros: Built specifically to tackle high-density computer vision workloads (like heavy transformer-based vision or dense YOLO matrices) with dramatically reduced memory-access energy overhead.
   • Cons: Newer market entrant for airborne autonomous robotics compared to NVIDIA or Qualcomm, meaning fewer plug-and-play flight stack integrations out of the box.computecompass.com
5. NXP Semiconductors
   • Key Hardware: i.MX 8M Plus / i.MX 9 series.
   • Specs/Performance: Modest NPU performance (~2.3 to 4.5 TOPS) at a very low 3W–8W total package power.
   • Pros: Industrial-grade reliability, extended lifecycle availability , deep integration with real-time co-cores (Cortex-M), and preferred for tightly integrated safety/flight-control boundary designs.
   • Cons: Underpowered if you plan to run heavy multi-camera simultaneous localization and mapping (SLAM) alongside large object-detection networks.aimultiple.com+1

Evaluation Checklist for Your Startup

• Software Toolchain: Check how painful it is to convert your trained PyTorch/ONNX computer vision models into the vendor's quantized runtime format. Avoid vendors with immature compilers that break on custom layers.libreyolo.com+1
• Form Factor & Weight: Look for raw chip-down availability or ultra-small form-factor modules (like M.2 2230 or custom soldered BGAs) rather than bulky developer kits.
• Thermal Constraints: Drones lack active cooling fans. Your chosen silicon must sustain peak vision inference at passive convection limits without thermal throttling mid-flight.

To help narrow down this list, could you share:

• What is your target power budget (e.g., strictly under 5W vs. up to 15W)?
• Are you running standard CNNs (like YOLOv8/v11) or heavier Vision Transformers/SLAM pipelines?

For an autonomous drone, I would evaluate embedded AI SoC vendors and dedicated NPU accelerator vendors separately. The right choice depends less on peak TOPS and more on watts per inference, camera pipeline integration, latency determinism, model compiler maturity, thermal envelope, and production availability.TechnoLynx

A practical shortlist:

AIMultiple RDP Technologies Limited## My suggested evaluation tiers

1. Small autonomous drone (<5–10 W compute budget)

Prioritize:

• Hailo Hailo-8-class accelerators
• Ambarella vision SoCs
• NXP Semiconductors i.MX platforms
• Renesas Electronics RZ/V

These are better aligned with battery-constrained flight where every watt matters.AIMultiple

2. Research prototypes / advanced autonomy stack

Prioritize:

• NVIDIA Jetson Orin
• Qualcomm robotics platforms
• SiMa.ai

You get faster iteration for:

• neural SLAM
• object detection + tracking
• depth estimation
• transformer-based perception
• sensor fusion

3. High-volume commercial drone

Look hardest at:

• Ambarella
• Qualcomm
• NXP Semiconductors
• Renesas Electronics

They tend to have stronger embedded lifecycles, supply-chain experience, and automotive/industrial qualification paths.

Evaluation checklist for your team

I would benchmark candidates using your actual models:

• Vision models

• YOLO variants

• segmentation networks

• optical flow

• depth estimation

• visual odometry / SLAM components

• Metrics

• TOPS/W (not TOPS alone)

• end-to-end camera-to-action latency

• thermal performance in an enclosed airframe

• RAM bandwidth

• INT8 quantization accuracy loss

• compiler effort (weeks/months of optimization matter)

• availability of Linux BSP, ROS 2 support, and camera drivers

A common mistake is selecting the highest-TOPS accelerator; drones often benefit more from efficient, deterministic inference plus good ISP/camera integration.TechnoLynx

For a first silicon bake-off, I would probably put these five on the bench:

1. NVIDIA Jetson Orin Nano/NX
2. Hailo-8-class accelerator
3. Ambarella CVflow-based SoC
4. Qualcomm robotics SoC
5. Renesas RZ/V or NXP i.MX 9-class platform

That gives you coverage across GPU, discrete NPU, vision ASIC, mobile AI SoC, and industrial MPU approaches.AIMultiple+1

For real-time computer vision and autonomy on resource-constrained drones, you need an embedded NPU (Neural Processing Unit) that balances TOPS (Tera Operations Per Second), low power consumption (watts/milliwatts) , and weight/thermal footprint.

The leading semiconductor vendors and silicon platforms to evaluate are categorized by their architectural strengths:

• NVIDIA (Jetson Orin Nano / AGX Orin / Thor series): The undisputed baseline for robotics and drone vision development. While heavier on power than micro-NPUs, the Jetson ecosystem offers unmatched software maturity via CUDA, TensorRT, and pre-trained ROS (Robot Operating System) packages. Go here if your primary bottleneck is software development speed and complex multi-cam transformer/CNN pipelines rather than absolute minimal battery drain.youtube.com+1
• Qualcomm (QRB / QCS series and Snapdragon Flight platforms): Exceptional performance-per-watt efficiency rooted in mobile architecture. Qualcomm’s RB5 and newer robotics platforms integrate high-end image signal processors (ISPs) alongside powerful hexion NPUs. They excel in lightweight autonomous systems where battery life and thermal limits are tightly restricted.
• Ambarella (CV3-AD / CV2x series): Renowned for ultra-low-power computer vision tightly coupled with world-class hardware image signal processing (ISP). Ambarella chips are heavily utilized where raw sensor data from multiple high-resolution cameras needs real-time stitching, object detection, and tracking right on the edge without melting a small drone chassis.computecompass.com
• NXP Semiconductors (i.MX 8M / i.MX 9 series): Ideal for ultra-low-power, deterministic edge processing and flight-control integration. NXP's integrated eIQ machine learning software environment targets low-footprint microcontrollers and applications processors where you need safety-critical reliability and minimal power draw for basic-to-moderate CV tasks.
• Axelera AI (Europa AIPU) / Hailo (Hailo-8 / Hailo-8L): High-efficiency dedicated coprocessors. If you choose a main host processor without a strong native NPU, adding a PCIe/M.2 module like the Hailo-8 or Axelera’s in-memory computing accelerators yields massive INT8 TOPS/Watt performance specifically tailored for edge vision, offloading the main CPU entirely.computecompass.com
• Lattice Semiconductor (Avant / CertusPro-NX FPGAs): For ultra-low latency and custom sensor interfaces, low-power FPGAs allow you to pipeline custom computer vision logic and hardwire lightweight neural nets with near-zero latency and micro-watt profiles, though it demands heavier RTL/FPGA design expertise.computecompass.com

To help narrow down this list, could you share:

• Your target weight and power budget (e.g., under 5W vs. 15W+)?
• What sensors/cameras you are interfacing with (resolution, frame rate, count)?
• Whether you prefer a fully integrated SoC or a coprocessor/accelerator module paired with a separate flight controller?

For an autonomous drone platform, I would evaluate vendors less on peak TOPS and more on TOPS/W, latency determinism, thermal envelope, camera pipeline integration, model compiler maturity, and availability of production-grade modules. A drone perception stack (object detection, tracking, optical flow, depth, segmentation, SLAM) often benefits from an NPU that can run continuously at a few watts rather than a large accelerator with occasional bursts.PXS Research

A practical shortlist:

LibreYOLO Converge Digest## My suggested evaluation tiers

1. Small battery-powered drones (<5–10 W AI budget)

Start with:

• Hailo
• Ambarella
• NXP Semiconductors
• Renesas Electronics
• DEEPX

Prioritize:

• INT8 inference efficiency
• camera ISP + NPU coupling
• <100 ms perception-to-control latency
• minimal RAM bandwidth

2. Larger autonomous drones / inspection UAVs (10–30 W AI budget)

Evaluate:

• NVIDIA Jetson
• Qualcomm platforms
• AMD Versal AI Edge

These give more flexibility for:

• multiple neural networks running simultaneously
• transformer-based vision models
• SLAM + perception + planning workloads

AIMultiple AMD### 3. If you are designing your own ASIC

Consider NPU IP vendors rather than buying a chip:

• Arm Ethos NPUs
• Cadence Design Systems AI IP
• CEVA NeuPro AI processors
• Synopsys ARC AI IP
• VeriSilicon NPU solutions

Parse Ceva IP## Benchmark questions I would send every vendor

1. Run our actual models

• YOLOv8/YOLO11 variants
• segmentation models (DeepLab, SAM variants if relevant)
• optical flow
• depth estimation
• custom CNN/Transformer models

2. Measure system-level metrics

• watts at sustained inference load
• FPS at required resolution
• end-to-end camera-to-control latency
• memory footprint
• thermal throttling behavior

3. Check deployment friction

• ONNX support
• TensorFlow Lite support
• PyTorch export path
• quantization workflow
• debugging/profiling tools

For a robotics startup, I would probably build a first-round bakeoff around NVIDIA Jetson Orin (software baseline), Hailo (efficiency baseline), Ambarella (vision SoC baseline), Qualcomm (integrated SoC baseline), and one emerging low-power NPU vendor such as DEEPX or Axelera. That gives you a good view of the trade space before committing to silicon architecture.rdp.in