A comprehensive analysis of the June 2026 seminar in Shenzhen, exploring the intersection of medical imaging, embedded AI, edge inference, and regulatory compliance, where x86 meets ARM, where the operating room meets the data center, and where evidence-based medicine collides with the hallucination problem of large language models.
The strategic landscape of medical AI at the edge, distilled from four hours of keynotes, product demonstrations, and roundtable discourse.
This seminar revealed three independent forces now converging on medical AI edge computing: (1) silicon evolution. Intel's Core Ultra roadmap pushing NPU TOPS from 35→99→180 in three generations; (2) regulatory inevitability, the EU Cyber Resilience Act (CRA) imposing €15M (1500万欧元) penalties, mandatory SBOM, Secure Boot, and 5-year security support obligations; and (3) clinical necessity. China's 20亿次 annual ultrasound exams served by only 200K doctors, creating an unsustainable labor gap that AI-assisted imaging must fill. Ankii positions itself at the intersection: a hardware integrator fluent in three ecosystems (Intel x86, Zhaoxin domestic x86, ARM/RK3588) with the medical certification expertise to bridge prototype and production.
The roundtable crystallized a non-negotiable constraint for medical AI: general-purpose LLMs (Qwen, Doubao, DeepSeek) are fundamentally unsafe for clinical use because they are optimized to tell users what they want to hear. Medical AI must be 循证医学 (evidence-based), offline, small-model, and incapable of self-learning from conversations. As Intel's 魏振明 stated: "I'd rather the model say 'I don't know' than make up a comforting answer."
| Speaker | Role | Session | Key Message | Highlight |
|---|---|---|---|---|
| 林总 | Ankii President | Opening Keynote | Hardware → Full-stack AIoT Platform | Ankii's evolution from board supplier to solution orchestrator |
| 魏振明 (Leo) | Intel China | Edge AI Keynote + Roundtable | CPU as AI orchestration brain; TOPS evolution | Panther Lake 180 TOPS; model routing saves tokens |
| 林启文 | Ankii IoT Division | 国产化 & Export Strategy | Dual-track: Rockchip+HarmonyOS vs Qualcomm+NVIDIA | CRA readiness; ROS integration for medical robots |
| 陈浩 | Ankii Embedded Systems | Medical Close Reading (46 min) | Embedded 4 criteria; ecosystem metaphor | McDonald's/Chinese burger/banquet analogy; 15-year Philips partnership |
| 王总 | Belinde 贝林德 | Roundtable Panel | Scenario-driven server integration | Pain-point analysis first: "What can we automate?" |
| 李总 | Ankii Compliance | CRA Briefing (9 min) | EU CRA by Sept 2026: SBOM, Secure Boot, OTA | €15M penalty; 24h critical vulnerability notification |
| Metric | Value | Context | Source |
|---|---|---|---|
| Panther Lake NPU TOPS | 180 TOPS | 3rd Gen Core Ultra, 2026 | 魏振明, Intel Keynote |
| Lunar Lake NPU TOPS | 40 TOPS | 2nd Gen Core Ultra, 2025 | 魏振明, Intel Keynote |
| Meteor Lake NPU TOPS | 99 TOPS | First Core Ultra with NPU | 魏振明, TOPS Evolution |
| Wildcat Lake NPU | 35 TOPS | Entry-level edge AI platform | 魏振明, Intel Keynote |
| Annual Ultrasound Exams (China) | 20亿次 | ≈2 billion exams/year | 陈浩, Medical Close Reading |
| Ultrasound Doctors (China) | 200,000 | ~10K exams/doctor/year | 陈浩, Medical Close Reading |
| Ultrasound Market Value | ¥200–300B | ¥100–800 per exam | 陈浩, Medical Close Reading |
| CRA Maximum Penalty | €15 million | EU Cyber Resilience Act | 李总, CRA Briefing |
| Critical Vuln Notification | 24小时 / 24 hours | CVSS ≥9 score; mandatory notification to authorities | 李总, CRA Briefing |
| Critical Vuln Fix Window | 72小时 / 72 hours | CVSS ≥9 score; patch must be deployed | 李总, CRA Briefing |
| Security Support Obligation | 5年 / 5 years | EU CRA mandated; minimum security update period | 李总, CRA Briefing |
| Ultrasound Training (General) | 3-5年 / 3–5 years | From novice to competent practitioner | 陈浩, Medical Close Reading |
| Ultrasound Training (Fetal) | 5-8年 / 5–8 years | Specialized fetal screening expertise | 陈浩, Medical Close Reading |
| Philips Medical Partnership | 15年 / 15 years | COM Express modular model partnership | 陈浩, Embedded Systems |
| Taiwan Nurse Shift Length | 12小时 / 12+ hour shifts | High patient ratio context driving automation | Roundtable, Ankii 林启文 |
| Ultrasound Sound Frequency | 20K Hz | Beyond human hearing; computational reconstruction required | 陈浩, Medical Close Reading |
| PCR Unit Surge (COVID) | 100→2,000/yr | 20x demand spike | 陈浩, IVD Case Study |
Speaker: 林总 (Ankii President). The strategic evolution of a company navigating three silicon ecosystems simultaneously.
Ankii's president 林总 opened the seminar by framing the company's transformation from a traditional embedded board supplier into a full-stack AIoT platform provider. This is not merely a branding exercise, it reflects a genuine architectural shift where Ankii now delivers integrated hardware, BSP software, AI SDK, medical certification guidance, and lifecycle management across Intel, Zhaoxin, and ARM ecosystems. The company's unique advantage is its ability to speak all three "languages" fluently, allowing medical device manufacturers to choose the silicon path that matches their regulatory and market requirements.
| Ecosystem | Metaphor (陈浩) | Key Silicon | Primary Market | Regulatory Path | AI Capability |
|---|---|---|---|---|---|
| Intel x86 | 🍔 McDonald's — standardized, predictable, global | Core Ultra, Panther Lake, Wildcat Lake, 12/13th Gen | Export / Global medical devices | EU CRA, FDA, CE | NPU 35–180 TOPS, OpenVINO |
| Zhaoxin 兆芯 | 🥡 Chinese burger chain — domestic ingredients, same recipe | Zhaoxin 6000G + 登临 NPU | Domestic / 信创 mandatory | 信创 compliance, CRA optional | Dual-domestic AI: 6000G CPU + NPU |
| ARM | 🏮 Chinese banquet — fragmented, anything goes | RK3588, Rockchip, Allwinner, HiSilicon, NXP, Qualcomm, TI, MTK, Renesas | Cost-sensitive / Specialized | Per-market; CRA preparation | RK3588 6 TOPS NPU + Hailo add-on |
| Tier | Form Factor | Examples | Target Use Case |
|---|---|---|---|
| Board Level | COM Express, COM-HPC, SMARC, Qseven, OSM, Raspberry Pi | Intel 12/13th Gen COMe, RK3588 SMARC | OEM carrier board integration |
| System Level | Fanless box, HPC server, AI Box (2U/3U/4U) | Multi-GPU AI Box, Video Matrix Box | Medical cart, operating room |
| Panel Level | Medical-grade display, industrial tablet, touch monitor | 21" Class III tablet, 32" nursing station | Nursing station, consultation room |
| Pre-Integrated | POS, mobile tablet, indoor positioning | Open-frame tablet (PCR), all-in-one | "Unbox and use" IoT deployments |
Three years ago we talked about metaverse. Now? Gone. AI — this is REAL. Companies are racing to make things cheaper. But the other path: embrace AI. Ankii can help. 陈浩, Ankii Embedded Systems (closing statement resonating with 林总's opening vision)
Speaker: 魏振明 (Leo), Intel China. CPU's renaissance as the orchestration brain for medical AI workloads.
魏振明 delivered a critical reframing of Intel's role in the AI era. While NVIDIA dominates the GPU narrative, Intel argues that the CPU is regaining importance as the central orchestrator of AI workloads. The Core Ultra architecture integrates NPU + GPU + CPU in a unified package, where the CPU's role is not parallel computation (that belongs to GPU/NPU) but AI orchestration: deciding what gets computed, by which accelerator, through which model, and at what cost. This is model routing, multiple paths to the same result, with Intel choosing the cheapest route.
| Generation | Codename | Year | NPU TOPS | Process Node | Key Architecture | Medical Edge Target |
|---|---|---|---|---|---|---|
| 1st Gen | Meteor Lake | 2024 | 99 TOPS | Intel 4 | Disaggregated tiles; first integrated NPU | Pilot medical AI deployments |
| 2nd Gen | Lunar Lake | 2025 | 40 TOPS | TSMC N3B | System-on-Chip; efficiency optimized | Mobile / portable medical devices |
| 3rd Gen | Panther Lake | 2026 | 180 TOPS | Intel 18A | Advanced NPU; RibbonFET + PowerVia | Real-time surgical navigation; ultrasound AI |
| Entry | Wildcat Lake | 2026 | 35 TOPS | Intel 18A | Cost-optimized edge AI | Nursing stations; basic diagnostics |
The CPU is not doing parallel compute — that's the GPU's job. The CPU is the brain: controlling what gets done, when, and how. It orchestrates which agent does what, which tool to use, compresses data in transit, and chooses the optimal workflow. 魏振明, Intel China — Roundtable, Q3
Speaker: 陈浩, Ankii Embedded Systems, 46-minute masterclass covering ultrasound physics, surgical robots, DR/CT/MRI, IVD, and infusion pumps.
| Metric | Value | Implication |
|---|---|---|
| Annual exams | 20亿次 (~2 billion) | Massive screening demand, overwhelming human capacity |
| Practicing doctors | 200,000 | ~10,000 exams/doctor/year — unsustainable workload |
| Per-exam price | ¥100–800 | Wide variance by region and exam complexity |
| Total market value | ¥200–300 billion/year | Larger than many national healthcare budgets |
| Training period (general) | 3–5 years | Long pipeline; AI can accelerate junior doctor capability |
| Training period (fetal screening) | 5–8 years | Specialized skill; highest AI impact potential |
| Ultrasound frequency | 20K+ Hz | Beyond human hearing; requires computational reconstruction |
陈浩 presented surgical navigation robots as a transformative technology. The workflow: CT/MRI scans → 3D point cloud reconstruction → overlay on patient anatomy. Positioning uses either magnetic field tracking or infrared camera systems. The robotic arm follows a pre-planned path and automatically avoids blood vessels. The clinical impact is dramatic: 2mm incisions versus 20–30cm without robotic assistance, meaning less blood loss, faster recovery, and fewer hospital-acquired infections. The hardware requirement is non-negotiable: a medical-grade computer that cannot fail during surgery. This is where Ankii's AI Box (2U/3U/4U form factors) with medical certification becomes critical infrastructure.
| Product | Business Model | Ankii Solution | Key Detail | Volume Spike |
|---|---|---|---|---|
| IVD (生化分析) | HP Printer Model: hardware cheap, reagents = cash machine | x86 over ARM for faster time-to-market | Ankii assisted with EMC/EMI/CFT medical certification | Steady clinical demand |
| PCR (核酸检测) | Instrument + consumables | Open-frame tablet: screen+touch+board in one supplier | Pre-COVID: 100–200 units/year | During COVID: 2,000 units/year (20×) |
| Immunoassay | Large equipment, high throughput | ARM OSM module | Competitors couldn't handle medical compliance (consistency, long-term supply, customization) | Stable demand |
| Infusion Pump | Precision drug delivery: dose, rate, duration control | 21" Class III medical tablet | Light-sensitive drugs need sealed delivery; short half-life drugs need continuous infusion | COVID-driven demand surge |
陈浩 provided a physics primer essential for AI engineers entering medical imaging. DR (X-ray): electron acceleration → tungsten target → radiation penetrates body → detector plate captures 2D overlap image. Radiation exposure is a concern. CT: X-ray slices computationally reconstructed into 3D volumes, higher radiation dose than DR. MRI: hydrogen atoms in magnetic field change alignment → image reconstruction. Safest modality, zero ionizing radiation. Contraindicated for patients with pacemakers or hearing aids due to the powerful magnetic field. Ultrasound: sound waves (20K+ Hz) → tissue echo → computational reconstruction. Fastest, safest, most accessible, but CPU-intensive, not merely GPU-bound. Understanding these physics differences is essential for designing appropriate AI acceleration strategies for each modality.
Speaker: 陈浩, Ankii Embedded Systems. The hardware ecosystem powering medical AI from the operating room to the traffic intersection.
陈浩 defined the four criteria that distinguish embedded systems from general-purpose computing: (1) Specialization, does ONE thing perfectly, not many things adequately; (2) Miniaturization, fits in tight physical spaces (medical carts, surgical suites, outdoor cabinets); (3) Low Power, battery or strict energy budgets, especially critical for portable medical devices; (4) Stability, 24/7 operation in harsh environments (factory floors, outdoor substations, operating rooms where failure is not an option). These criteria explain why embedded systems demand different silicon and integration approaches than cloud or desktop computing.
| Component | Specification | Purpose |
|---|---|---|
| CPU | Intel 12/13th Gen Core | Orchestration, workflow, model routing |
| GPU | Multiple NVIDIA GPUs | Parallel inference, image reconstruction |
| NPU | Multiple NPU accelerator cards | Specialized AI inference (low-power, high-throughput) |
| Front Panel | Customizable for doctor device connections | Plug-and-play into medical workflows |
| Form Factor | 2U / 3U / 4U rack-mountable | Medical cart integration, operating room deployment |
| Certification | 国医规 (National Medical Regulation) | Legal operation in Chinese clinical settings |
| Supply | Long-term guaranteed (5–10 year lifecycle) | Medical device regulatory requirement |
| Purpose | Data-center-level AI compute in the OR | Real-time inference without cloud dependency |
Zhaoxin (兆芯) represents China's domestic x86 alternative. Its strategic advantage is ecosystem compatibility: it borrows the entire Intel software stack, same operating systems, same drivers, same application binaries. This makes it the "easiest domestic transition path" for organizations under 信创 (Xinchuang) mandates. The Zhaoxin 6000G can be paired with a 登临 (Denglian) NPU to create a dual-domestic AI solution: Chinese CPU + Chinese NPU, fully 信创 compliant. 陈浩 presented two transition case studies: (1) Power grid fault recording at 220kV/110kV substations, records pre/post fault waveform data, locates break points; migrated from Intel to Zhaoxin with zero software changes. (2) Quadruped robot, same smooth transition, same codebase, demonstrating that the domestic x86 path preserves the entire development investment while achieving regulatory compliance. The quadruped robot case is particularly illustrative: a complex ROS2-based robotics platform with real-time SLAM and motor control, migrated from Intel to Zhaoxin without rewriting a single line of application code.
Rockchip's RK3588 is the ARM star: 8 cores, 6 TOPS NPU. Its strength is not raw TOPS (far below Intel's NPU) but multimedia processing: 4×HDMI output + 1×HDMI input for video matrix applications. Ankii has pushed it to its limits: running DeepSeek R1 and Qwen 2.5 locally, 8-channel real-time video decode for traffic analysis, and <1-second latency court transcription. When paired with a Hailo NPU add-on card, it competes with Intel solutions on specific workloads. The RK3588's versatility makes it the default choice for cost-sensitive edge AI applications that don't require the full Intel ecosystem.
RK3588 6 TOPS NPU DeepSeek R1 Qwen 2.5| Case | Silicon | Challenge | Ankii Solution | Result |
|---|---|---|---|---|
| AMR (Autonomous Mobile Robot) | RK3588 (x86→ARM) | Car factory, 100×100m zone, ROS2 SLAM | ARM transition with ROS2 integration | Point cloud SLAM, reduced BOM cost |
| ETC Toll Collection | ARM (custom→Ankii) | Customer self-designed board, unsustainable maintenance | Single-supplier lifecycle solution | Ongoing maintenance, single throat to choke |
| Traffic Lights | RK3588 + Hailo NPU | 8-channel real-time decode, 4-direction, 8-lane traffic analysis | "Squeezed every drop from RK3588" | Replaced Intel solution; equal performance, lower cost. This traffic lights application runs real-time vehicle detection, counting, and signal optimization entirely at the edge — no cloud round-trip, no latency penalty. The RK3588 processes 8 simultaneous video streams from 4 directions across 8 lanes, performing object detection, classification, and traffic flow analytics in real time. |
| Court Transcription | RK3588+1828 | Real-time speech-to-text + auto court record generation | Edge inference pipeline | <1 second latency; no cloud dependency |
One of Ankii's most practical innovations: a 32-inch display with 16GB RAM capable of splitting from 4-way to 16-way simultaneous video streams. All feeds, cameras, patient monitors, other computer signals, appear on ONE screen. This directly addresses the brutal workload of nurses (Taiwan context: 12+ hour shifts, high patient ratios). The AI layer adds VLM + AI agents that analyze real-time video to detect bed anomalies and auto-generate shift-change reports. Nurses spend minutes, not hours, on handover documentation. For doctors, the same multi-split technology overlays surgical video with vital signs, with foot-pedal zoom/switch and AI highlighting of suspicious regions.
Panel: Intel 魏振明 · Ankii 林启文 · Belinde 王总 · Moderator: Ankii 林总, 52 minutes of unfiltered discourse on AI deployment realities.
| Panelist | Character Choice | Interpretation | Strategic Implication |
|---|---|---|---|
| Intel (魏振明) | 兴起 (Rise/Emerge) | CPU regaining importance in the AI era. Not just GPU anymore. CPU controls orchestration, routing, workflow. | Intel's architectural bet: heterogeneous compute with CPU as ringmaster |
| Ankii (林启文) | 边端 (Edge/Endpoint) | Medical devices are inherently edge. AI acceleration must happen at the device side — real-time, immediate feedback. | Cloud is for big data. Edge is for life-and-death decisions. |
| Belinde (王总) | 未来 (Future) | AI penetrating all industries. Expecting more partnerships for digitalization + intelligence. | Server market shifting from spec-sheet selling to scenario consulting |
This was the most consequential discussion of the entire seminar. Intel's 魏振明 delivered a stark warning: General-purpose LLMs (Qwen, Doubao, DeepSeek) are your FRIENDS, they tell you what you WANT to hear. In medicine, this is FATAL. The scenario he painted: a doctor wants to hear "it's just indigestion, go home", but the patient actually has appendicitis or sepsis. The patient dies. Medical AI must be fundamentally different from general AI. The five non-negotiable requirements for medical LLMs:
A general LLM is like a friend who always tells you what you want to hear. 'Doctor, it's probably just indigestion — go home and rest.' But the patient has appendicitis. Sepsis sets in. The patient dies. That's not acceptable. In medicine, we'd rather the AI say 'I don't know.' 魏振明, Intel China — Roundtable, Q5 (the most quoted moment of the seminar)
| Panelist | Approach | Key Insight |
|---|---|---|
| Ankii (林启文) | Right SOC/NPU for model size | 1.5B vs 7B vs 13B models need different hardware. Memory architecture matters: shared memory vs dedicated NPU memory vs NPU with built-in RAM — same model, vastly different performance. Ankii's role: optimal hardware selection + software unification. |
| Intel (魏振明) | Model routing + orchestration | Don't generate more tokens than needed. Multiple paths to the same result — pick the cheapest. "Bus costs ¥10, highway costs ¥1000." |
| Belinde (王总) | Pain-point analysis first | "What can we automate? How many people can we free up?" — scenario consulting over spec-sheet selling. |
| Question | Answer | Key Detail |
|---|---|---|
| Modular hardware standards? | Ankii already supports COM Express, COM-HPC, SMARC, Qseven, OSM, Raspberry Pi form factors — all international modular standards for embedded computing. | Philips Medical: 15-year partnership using this modular model. The modular standards approach means: "Customer makes the carrier board, we provide core module + BSP + schematic review. Customer locked in for 5–10 years." This is the modularity paradox — the more standardized the interface, the deeper the long-term integration dependency. |
| Supply chain discussion? | "This is Solution Day. Supply Chain Day is separate. Add me on WeChat, I'll forward to procurement." | Clear boundary between technical solutions and commercial procurement — a professional supply chain separation that maintains focus on engineering while acknowledging the commercial reality. |
| Agentic AI — big vs medium vs small agents? | Intel: Most people don't need a 24/7 personal AI agent. Enterprise does (hospitals, data sovereignty). | The bottleneck is TOKENS, not agent size. "Unless you have 100 NVIDIA cards at home, you care about token cost." — 魏振明 |
Intel's 魏振明 on why token cost anxiety is the defining constraint of AI deployment, and how model routing changes the game.
魏振明 drew a powerful analogy: 10–15 years ago, people feared their mobile data bills, checking usage obsessively, afraid of overage charges. Today, the same psychology applies to AI tokens. He recounted a story from the seminar: "Someone left an AI server running overnight, burned a house worth of tokens." (The exact metaphor varied between "a house" and "a car" depending on the telling, but the point is devastating: AI tokens are expensive, unbounded, and terrifyingly easy to waste through simple negligence.) Intel's thesis: the company's role is NOT generating more tokens. It's making token usage MORE EFFICIENT. This story became the seminar's most visceral anecdote, a concrete, relatable horror story that crystallized the abstract concept of "token cost" into something every attendee could feel in their wallet.
| Scenario | Approach | Token Cost | Latency | Appropriate For |
|---|---|---|---|---|
| Cloud LLM (GPT-4 class) | Every query to cloud | $$$ (highest) | 500ms–2s | Complex reasoning, research |
| Cloud LLM (small) | Cheaper cloud model | $$ | 200–800ms | General Q&A, summarization |
| Edge LLM (7B–13B) | Local inference on AI Box | $ (one-time HW) | 50–200ms | Medical diagnosis, real-time analysis |
| Edge LLM (1.5B–3B) | Local on ARM/RK3588 | $ (lowest HW) | 10–50ms | Image triage, anomaly detection |
| Model Routing (Intel) | Route to cheapest model that satisfies quality | Optimal | Variable | Enterprise AI workloads |
Intel's model routing concept: there are always multiple paths to the same result. A simple question doesn't need GPT-4, it needs a deterministic lookup. Intel's orchestration layer picks the cheapest route that satisfies the quality requirement. 魏振明's analogy: "Bus costs ¥10. Highway costs ¥1000. You don't take the highway to go one block." The CPU's job is to make this routing decision in real-time, compressing data in transit, selecting the optimal model, and choosing which accelerator (NPU, GPU, or CPU) handles each sub-task. This is AI orchestration, like Manus or Lobster, many agents behind one interface, with Intel orchestrating which agent does what.
| Agent Size | Use Case | Token Demand | Who Needs It |
|---|---|---|---|
| Big Agent | 24/7 personal assistant, multi-domain reasoning | Very high (hundreds of thousands of tokens/day) | Enterprise only — hospitals, data centers. "Most people don't need this." |
| Medium Agent | Domain-specific (e.g., radiology AI, surgical navigation) | Moderate (targeted inference bursts) | Specialized medical devices; real-time edge inference |
| Small Agent | Single-task (e.g., anomaly detection, bed exit alert) | Low (deterministic, lightweight) | Nursing stations, IoT sensors, embedded devices |
People think they want a 24/7 AI agent. But unless you have 100 NVIDIA cards at home, you care about token cost. The bottleneck is TOKENS — not whether the agent is big, medium, or small. 魏振明, Intel — Roundtable Audience Q&A: agentic AI
Ankii's 林启文 clarified the SDK strategy during the roundtable: the SDK is on GitHub, free to use, but only on Ankii hardware. It's not open-source in the traditional sense. The app is the demo; the SDK is the actual product. This is a razor-and-blades model adapted for AI hardware: give away the software framework to lock in the hardware platform for 5–10 year medical device lifecycles.
Speaker: 李总, Ankii Compliance, 9-minute briefing on the regulatory earthquake hitting medical device manufacturers in September 2026.
The EU Cyber Resilience Act (CRA) represents a regulatory paradigm shift for medical device manufacturers exporting to Europe. The deadline is September 2026, three months after this seminar. The maximum penalty is €15 million. But the real cost is not the fine: it's market exclusion. Devices without CRA compliance cannot be sold in the EU. For Chinese medical device manufacturers using Ankii hardware, this is existential. 李总's briefing covered the four pillars of CRA compliance and Ankii's preparation strategy.
| Pillar | Requirement | Deadline / Timeline | Ankii Solution |
|---|---|---|---|
| 1. Secure Boot | From the moment of power-on, the system must be in a verified state. Boot chain integrity from firmware → bootloader → OS → application. | Mandatory by Sept 2026 | Hardware-rooted Secure Boot across Intel, Zhaoxin, and ARM platforms |
| 2. SBOM (Software Bill of Materials) | Mandatory component inventory: every library, dependency, and version must be listed and traceable. Vendors MUST provide SBOM to customers. | Mandatory by Sept 2026 | Export-ready SBOM generation tooling; automated dependency tracking |
| 3. Vulnerability Management | CVSS scoring required. Critical vulnerabilities (CVSS ≥9): 24-hour notification to authorities, 72-hour fix window. Regular patches for all severities. | Ongoing; 24h/72h SLA for critical | Partnership with Swiss company for cloud/on-premise OTA update management |
| 4. 5-Year Security Support | Manufacturers must provide security updates for a minimum of 5 years from the date of product placement on the market. | 5 years from launch | Long-term support infrastructure; regular + critical patch cadence |
| Device Class | CRA Burden | Ankii Readiness | Key Risk |
|---|---|---|---|
| Class III Medical (infusion pumps, surgical robots) | Maximum — life-critical devices | Medical-certified hardware + SBOM + Secure Boot | Any security gap = immediate market withdrawal |
| Class II Medical (ultrasound, patient monitors) | High — diagnostic devices | Full CRA preparation; OTA infrastructure | SBOM completeness; third-party library CVEs |
| Class I Medical (examination lights, hospital beds) | Moderate — non-invasive | Standard CRA baseline | Legacy device retrofitting |
| Non-medical IoT (ETC, traffic, court) | Moderate — connected devices | CRA-ready but not mandatory for domestic | Export-only requirement for now |
Ankii has partnered with a Swiss company to provide OTA (Over-The-Air) update management, available both as a cloud service and an on-premise deployment. This is critical because medical devices cannot rely on public cloud infrastructure for security patches (air-gapped hospital networks, data sovereignty requirements). The OTA platform handles regular patches, critical vulnerability fixes (the 72-hour window), and SBOM version tracking. For medical devices specifically, the requirements are particularly strict, because a compromised infusion pump or surgical robot is not just a data breach; it's a potential fatality.
Synthesizing the patterns that recurred across all four sessions, the unspoken assumptions, the strategic tensions, and the emergent consensus.
The seminar revealed a computing landscape far more complex than "x86 vs ARM." Ankii navigates three distinct ecosystems (Intel x86, Zhaoxin domestic x86, ARM with 8+ CPU vendors and 5+ operating systems) while maintaining a unified product and support fabric. The food metaphors from 陈浩: McDonald's, Chinese burger chain, Chinese banquet, are not just colorful analogies; they encode the actual fragmentation that medical device manufacturers must navigate. The strategic implication: no single silicon vendor can serve all medical device markets. Export devices need Intel for CRA compliance and global recognition. Domestic 信创 devices need Zhaoxin. Cost-sensitive or specialized devices need ARM. Ankii's value proposition is being the translator across all three.
Belinde's 王总 crystallized a shift that echoed across every session: in the last 5–6 years, customers have stopped asking "how many cores, what frequency?" and started asking "what does my SCENARIO need?" This shift, from spec-sheet selling to scenario consulting, is the defining transformation of edge computing. Cloud is fragmenting by vertical. Medical, industrial, traffic, and court transcription all need different compute profiles. The integrator's role (Ankii's role) is no longer matching specs to requirements; it's understanding the clinical or operational workflow and designing the compute architecture around it.
The most important cross-cutting theme is the hallucination boundary: the point beyond which general-purpose AI's tendency to be agreeable becomes medically lethal. This theme appeared in the roundtable (Q5), was reinforced by Intel's model routing philosophy, and is implicitly embedded in every medical case study. The implication: medical AI requires a fundamentally different training and deployment philosophy. Evidence-based (循证医学), offline, small-model, non-self-learning, these are not preferences; they are safety requirements. The "friendly" LLM that tells a doctor what they want to hear is not a feature; it's a fatality waiting to happen.
The audience question about modular hardware standards (COM Express, COM-HPC, SMARC, Qseven, OSM, Raspberry Pi) revealed a sophisticated business model: modularity creates customer lock-in through value, not coercion. The Philips Medical partnership : 15 years, works because Ankii provides the core module + BSP + schematic review, and the customer designs their own carrier board. The customer gets customization; Ankii gets a 5–10 year lock-in. This is the modularity paradox: the more standardized the interface, the deeper the integration dependency.
Token anxiety emerged as the unifying economic constraint across cloud, edge, and hybrid AI deployments. Intel's story about someone "burning a house's worth of tokens" by leaving an AI server on overnight is not just an anecdote, it reveals the fundamental economic structure of AI: tokens are unbounded costs in a world of bounded budgets. Every technical decision, model routing, AI orchestration, edge vs cloud, big vs small model, reduces to a token economics optimization. This is why Intel frames its role as making token usage MORE EFFICIENT, not generating more tokens. The bottleneck is not model capability; it's the cost of exercising that capability.
Two seemingly opposite regulatory forces. China's 信创 (domestic technology independence) and the EU's CRA (cybersecurity regulation), are converging to create a dual compliance requirement for Chinese medical device exporters. A device must be both 信创 compliant (for domestic approval) and CRA compliant (for EU market access). Ankii's three-ecosystem strategy is not a luxury; it's a necessary response to this dual sovereignty demand. The Intel track handles CRA export; the Zhaoxin track handles 信创 domestic; the ARM track bridges both for cost-sensitive applications.
| Theme | §1 Ankii Opening | §2 Intel AI | §3 Med Imaging | §4 Embedded | §5 Roundtable | §6 Tokens | §7 CRA |
|---|---|---|---|---|---|---|---|
| Ecosystem Fragmentation | Three-ecosystem strategy | x86 NPU evolution | Modality-specific compute | ARM case diversity | Food metaphors | Model routing across models | Per-platform Secure Boot |
| Scenario Consulting | AIoT platform narrative | Medical edge focus | Ultrasound economics | Pain-point analysis | Belinde server shift | Right model for scenario | Risk-based device classes |
| Hallucination Boundary | — | — | 循证医学 requirement | AI as "2nd eyes" not replacement | Q5 critical discussion | Small model > big model | Auditability via SBOM |
| Modular Lock-In | Board → system → solution | NPU as modular accelerator | Ultrasound form factors | COM Express standards | 15-year Philips case | SDK lock-in on HW | OTA platform dependency |
| Token Economics | — | Efficient token usage | Edge inference = free tokens | Local models (DeepSeek R1) | "Burned a house worth" | Central theme | — |
| AI Sovereignty | 国产化 dual track | Intel global ecosystem | China market dominance | 信创 Zhaoxin transition | Supply chain separation | Data sovereignty at edge | Dual compliance burden |
| Quote | Speaker | Context | Significance |
|---|---|---|---|
| "Metaverse is gone, AI is real." | 陈浩 | Closing statement, Embedded session | Strategic pivot from hype to practical AI deployment |
| "I'd rather the model say 'I don't know' than make up a comforting answer." | 魏振明 | Roundtable Q5 on hallucination | Defines the safety boundary for medical AI |
| "Someone left AI server on overnight — burned through a house worth of tokens." | 魏振明 | Roundtable on token anxiety | Concrete illustration of unbounded AI cost |
| "Bus costs ¥10, highway costs ¥1000." | 魏振明 | Roundtable on model routing | Core Intel philosophy: cheapest path to quality result |
| "Intel = McDonald's. Zhaoxin = Chinese burger. ARM = Chinese banquet." | 陈浩 | Embedded ecosystem overview | Memorable encoding of ecosystem fragmentation |
| "We squeezed every drop of compute from the RK3588." | 陈浩 | Traffic light case study | Ankii's optimization capability claim |
| "Customer locked in for 5–10 years." | 林启文 / 陈浩 | Modular standards Q&A | Business model: modularity as long-term partnership |
| "This is Solution Day. Supply Chain Day is separate." | Ankii Team | Audience Q&A | Professional boundary-setting |
The Ankii & Intel Medical AI Edge Computing Seminar was not a product launch. It was a strategic alignment demonstration. Ankii proved it can speak Intel's language (TOPS, NPU, model routing), Zhaoxin's language (信创, domestic transition), ARM's language (fragmented ecosystem optimization), and the medical industry's language (循证医学, CRA compliance, 15-year partnerships). Intel proved that its CPU renaissance narrative. CPU as AI orchestration brain, not just a GPU sidekick, has a concrete deployment story in medical edge computing. And the roundtable proved that the industry's most important conversation is not about bigger models or more TOPS, it's about safety boundaries, token economics, and the hallucination gap between "helpful AI" and "medical-grade AI." The metaverse may be gone, but the real work, making AI safe, efficient, and deployable at the edge of life-and-death decisions, has only just begun.
| Product Category | Hardware | Demonstration | Target Audience |
|---|---|---|---|
| AI Box | Intel 12/13th Gen + Multi-GPU + Multi-NPU | Real-time medical image AI inference; DeepSeek R1 + Qwen 2.5 running locally | Medical device manufacturers; surgical navigation |
| Multi-Screen Display | 32" Medical Display · 16GB · 4→16 split | 16 simultaneous video streams on one screen; nurse shift-change AI report generation | Hospital nursing stations; ICUs |
| Video Matrix Box | RK3588 · 4×HDMI Out + 1×HDMI In | Medical video walls; real-time multi-source display; ad distribution | Hospital command centers; public health displays |
| ARM Development Boards | RK3588 SMARC · OSM · Raspberry Pi form factors | ROS2 SLAM; ETC toll processing; court transcription pipeline | System integrators; software developers |
| Zhaoxin Systems | Zhaoxin 6000G + 登临 NPU | Fault recording playback; quadruped robot control; 信创 compliance demo | Power grid operators; government procurement |
| Medical Tablets | 21" Class III · Open-frame · Industrial | Infusion pump integration; PCR instrument control; immunoassay interface | IVD manufacturers; hospital equipment purchasers |
| HPC Servers | Multi-GPU rack servers | Data-center-class AI training and inference; medical image reconstruction | Large hospital data centers; research institutions |
| Stakeholder | Key Takeaway | Action Required | Time Horizon |
|---|---|---|---|
| Medical Device Manufacturers (Export) | CRA compliance is non-negotiable by September 2026. Ankii+Intel provides a ready path: Secure Boot, SBOM, OTA, 5-year support. | Begin CRA gap analysis immediately. Engage Ankii for SBOM generation and Secure Boot enablement on existing designs. | 3 months (Sept 2026) |
| Medical Device Manufacturers (Domestic/信创) | Zhaoxin 6000G + 登临 NPU = viable domestic AI path. Smooth transition from Intel with no code changes. | Evaluate Zhaoxin for 信创 compliance. Plan dual-sourcing: Intel for export, Zhaoxin for domestic. | 6–12 months |
| Hospital IT Departments | Edge AI (AI Box in the OR) eliminates cloud dependency for real-time surgical navigation and ultrasound analysis. | Pilot AI Box deployment in one surgical suite. Measure latency, accuracy, and clinician satisfaction. | 12–18 months |
| AI Model Developers | Medical AI must be evidence-based, offline, small-model, non-self-learning. General LLMs are unsafe for clinical use. | Retrain models on curated medical literature only. Disable internet access and conversation-based learning. | Immediate |
| System Integrators | Ankii's three-ecosystem strategy means one partner for Intel, Zhaoxin, and ARM needs. Modular standards enable carrier board customization. | Evaluate which ecosystem fits each customer project. Leverage Ankii's BSP + schematic review for faster time-to-market. | Ongoing |
| Regulatory Bodies | The CRA is creating a bifurcated market: compliant devices for EU, 信创 devices for China. Dual compliance is becoming a competitive moat. | Monitor how CRA enforcement interacts with medical device certification (CE marking, FDA). Expect harmonization pressure. | 2026–2028 |
Three years ago we talked about metaverse. Now? Gone. AI — this is REAL. Companies are racing to make things cheaper. But the other path: embrace AI. Ankii can help. The question is not whether AI will transform medicine — it's whether we build it safely, deploy it efficiently, and regulate it intelligently before someone gets hurt. 陈浩, Ankii Embedded Systems — and the unspoken thesis of the entire seminar