[Edge Infrastructure ②] Key Solutions by Company

With the rise of agentic AI, the enterprise infrastructure landscape is shifting to a locally based distributed computing model. Behind that shift are unpredictable cloud API token costs and the risk of intellectual property leakage.

As a result, companies are responding by pushing routine, high-frequency inference tasks to the front lines, with the goal of optimizing TCO. In this trend, major vendors are accelerating the buildout of end-to-end hardware ecosystems spanning from the desktop to the data center.

Dell Technologies has fully unveiled its execution infrastructure portfolio, "Dell Deskside Agentic AI," to match this shift. The portfolio is designed to support the phased adoption and secure expansion of agentic AI in enterprises, and its defining feature is the systematic segmentation of hardware by customer system adoption maturity. In the early exploration stage, it recommends "Dell Pro Precision 9 T2" and "Dell Pro Max with GB10" so users can quickly test ideas and validate prototypes.

The expansion stage, which means moving from PoC to full application in actual work, centers on the tower workstation "Dell Pro Precision 9 T4," which supports data security and operational stability. The orchestration stage addresses complex multi-agent workflows, and "Dell Pro Max with GB300" and "Dell Pro Precision 9 T6" are deployed there. Based on this, Dell Technologies is pursuing end-to-end expansion through "Dell AI Factory."

Dell's key differentiator is its "End-to-End" architecture, which continuously spans from the desktop to the data center. When designing its deskside solutions, Dell made it possible to extend naturally to Dell PowerEdge servers and central infrastructure when needed. It has also linked them with the "Dell AI Factory with NVIDIA" ecosystem based on cooperation with NVIDIA, and has accumulated more than 3,000 enterprise deployment cases and benchmarks. Through this, customers can deploy agents validated on the desktop to the companywide data center scale while maintaining the same software stack and security framework.

Dell is also strengthening hardware-embedded security technologies to respond to terminal-level security threats. Dell's security direction is to reflect security from the hardware design stage, and it applies the "SafeBIOS" and "SafeID" features. It also preemptively applied quantum-resistance capabilities to enterprise PCs launched this year to counter quantum attacks. Based on this, Dell is focusing on regulated industries where outbound data transfer is strictly limited, R&D organizations that repeatedly run agentic workflows, and enterprise customers that need ultra-low-latency on-site processing.

Lenovo ISG is pursuing a "hybrid AI infrastructure" strategy that organically connects central data centers and the front lines. This strategy is built on a structure that connects central data centers and on-site edges.

For the field, it supplies edge-optimized servers including the "ThinkEdge" SE455 V3 line, along with SE100, SE350 V2, SE360 V2, and SE450. This edge product family is designed with an emphasis on durability, security, and remote management features, and targets factories, stores, and telecom base stations. It is also built to operate in harsh physical environments.

For the high-performance computing area linked to data centers and central infrastructure, it supplies the "ThinkSystem" SR675 V3. The "ThinkSystem" SR675 V3 is an AI inference-optimized server, and the accompanying "SR675i V3" is a product that maximizes inference capability. This high-performance computing product family supports the operation of low-latency AI applications.

Lenovo is expanding its hybrid infrastructure options by emphasizing its competitive edge in a global technology partner ecosystem and workload optimization capabilities, based on a strategy that is not tied to a specific chipset or a single architecture. To this end, it is building hybrid infrastructure in cooperation with NVIDIA, Intel, AMD, and Red Hat, and is offering a range of combinations tailored to customer environments, including "Hybrid AI Advantage with NVIDIA," Red Hat AI Enterprise, and CPU-only platforms based on Intel Xeon 6. In this process, the priority criteria for deployment review are system stability, management convenience, and service continuity rather than raw computing performance metrics.

Lenovo applied this approach to the integrated replacement of the aging IT infrastructure at Chang'an University, and configured it with the Nutanix HCI method based on the "ThinkAgile" HX650 V3. As a result, it reduced management burden by about 30%, and it is extending that stable infrastructure modernization experience into the edge area.

Lenovo said it will supply end-to-end portfolios centered on sensitive data industries such as manufacturing quality inspection, edge maintenance, retail customer behavior analysis, telecom edge traffic optimization, healthcare, and finance.

Movin has said it will target the physical AI and edge markets based on the combination of ultra-compact hardware and a high-performance NPU architecture relative to power consumption. As a key solution in its hardware portfolio, it presented the on-device AI USB stick "MLD-R1 AI USB."

The "MLD-R1 AI USB" is equipped with Movin's proprietary edge AI SoC, "REGULUS." The product delivers 10 TOPS of AI inference performance through a standard USB port connection alone, and is designed to minimize 24-hour operating burden with ultra-low power consumption at around 3W. It also supports Windows and Linux OS and x86 and ARM hosts, allowing existing laptops, industrial PCs, and kiosks to be turned into AI-dedicated devices without system replacement.

In addition to the ultra-compact USB stick, Movin is building a full edge hardware lineup spanning terminals to servers. It provides the "REGULUS AI single-board computer (SBC)" for standalone embedded equipment design, and this product operates in a low-power 6W to 8W environment and is equipped with a quad-core ARM CPU, ISP, video codec, and NPU, handling the entire process from camera video input to 4K AI inference to display output independently without a host. Movin is also supplying the ultra-compact system-on-module form factor "MLM-1 SoM." Alongside this, its on-premises and deskside infrastructure accelerator lineup under "ARIES" includes the 25W low-power PCIe card "MLA100," which delivers 80 TOPS, the high-performance "MLA400" with 320 TOPS, the on-site installation mini PC "MLX-A1," and the high-density inference server family "MLS," and Movin is broadly offering "MLA100 (80 TOPS)," "MLA400 (320 TOPS)," "MLX-A1," and "MLS" under the "ARIES" lineup.

Movin presents deployment flexibility that covers new build environments and existing control infrastructures as its core differentiation strategy. For new sites, it deploys REGULUS-based SBCs and intelligent cameras, while for existing control rooms that operate many conventional cameras, it responds by adding ARIES accelerator cards to process video. It also lowers the deployment barrier by supporting the same software environment in both new and existing environments.

Movin has completed validation of more than 490 models, from vision AI to generative AI. It presents low TCO and high power efficiency as its strengths. Based on this, it is expanding references into public video surveillance, in-house sLLM-based document search, and robotics.

Meanwhile, NVIDIA is rolling out a portfolio across deskside and edge computing. NVIDIA is positioning "RTX Spark," a next-generation Windows PC platform for slim laptops and compact desktops, as its lead platform, and has equipped it with the "RTX Spark Superchip" designed for AI and graphics processing. The chip consists of a Blackwell architecture GPU and a Grace CPU, and RTX Spark provides up to 1 PFlops of AI computing performance and up to 128GB of unified memory. Alongside this, NVIDIA is building its edge lineup with the Linux-based desktop development environment "DGX Spark," "RTX PRO" for professional workstations, and "Jetson IGX" for robots and manufacturing equipment.

RTX Spark's strength lies in its structure that natively supports the CUDA and RTX ecosystems on-device. As a result, developers can use familiar development tools, and local systems can be used for prototyping large language models (LLMs) and long-running AI agents. The expansion path for workloads then leads to data center DGX systems and the cloud. In addition to AI inference, RTX Spark also accelerates high-load creative tasks at the same time, handling 90GB-plus large 3D scene rendering, 12K video editing, and real-time graphics generation by accelerating them simultaneously on a single device. NVIDIA's core focus is its close integration with the software ecosystem.

NVIDIA is working with Microsoft to build "NVIDIA OpenShell" and the Windows security framework. OpenShell includes features for safely running agent isolation environments and for allowing users to directly control access permissions. NVIDIA is also pursuing creative pipeline optimization with Adobe and Blender. In Korea, Krafton and NCSoft are working on running their own titles on RTX Spark systems. The finished product is expected to be supplied through major manufacturers such as Dell, Lenovo, HP, ASUS, and MSI, and it will be broadly introduced to the market soon.

Intel proposed a distributed infrastructure based on its x86 portfolio. The scope covers client AI PCs, workstations, and edge servers, and the goal of the distributed infrastructure is flexibility and cost efficiency. For personal devices, Intel supplies the "Core Ultra" Series 3 processor that integrates CPU, GPU, and NPU on a single chip, and for workstations it proposed a combination of Intel processors and the "Intel Arc Pro" B Series GPU. The configuration is aimed at developers and engineering organizations, and supports stable local operation of relatively large AI models and agentic workloads. For high-throughput on-site edge systems, it supplies the "Xeon 6" processor.

Intel's differentiation strategy lies in a structure that does not depend on a specific accelerator, the use of existing x86 IT assets, and a progressively expandable "open and hybrid architecture." Accordingly, it said companies do not need to introduce expensive external GPUs uniformly for every task. Routine work such as document summarization and in-house knowledge search is distributed to the CPU, integrated graphics, and low-power NPU, while external GPUs are additionally deployed for high-load tasks in workstations. Intel explained that this supports TCO optimization.

It also said it would implement local-first execution by applying intelligent task routing software called "SuperClaw." Intel said it has demonstrated results that cut cloud token consumption for enterprise workloads by up to 70% through this approach.

Intel presented OpenVINO and oneAPI as key elements of software optimization. Through this, developers can use the same code without being tied to a specific piece of hardware, and they can optimize and deploy that same code across PCs, workstations, and edge servers. On this basis, Intel is accelerating penetration into the enterprise and physical AI markets by combining industry-specific references.

Intel cited examples of cooperation including LG Innotek in manufacturing smart factories, and JLK and Samsung Medison in medical imaging analysis. These references have undergone field validation.

AMD has entered the deskside infrastructure market. Its flagship products are the "Ryzen AI Max PRO 400" series and the desktop AI accelerator "Radeon AI PRO R9700." The "Ryzen AI Max PRO 400" series is the first x86 client processor to run a 300 billion parameter model directly locally, and the Ryzen AI Max PRO 400 platform supports high-performance local AI and simultaneous agentic workflows with a single-chip configuration that combines strong local AI computing performance, large unified memory, and enterprise-grade management features. The Radeon AI PRO R9700 processes about 18 million tokens per day and has proven its cost breakeven versus the cloud in 3 months.

AMD presents two core technology philosophies: renewed focus on the CPU in the age of agentic AI and heterogeneous computing. AMD explains that as AI agents become more advanced, they require not only model acceleration but also file navigation, system command execution, subtask coordination, and broad orchestration, and that the core engine responsible for this underlying execution is the CPU. Accordingly, AMD is proposing a system-level full-stack architecture that organically combines CPU, GPU, and NPU, with the CPU handling general-purpose tasks and orchestration, the GPU handling parallel computing inference, and the NPU performing low-power always-on processing.

AMD is presenting platform flexibility and openness as its key weapons, with the goal of moving away from dependence on any specific proprietary ecosystem. To this end, it is investing in the open source-based ROCm software stack and developer ecosystem, and is leading open industry standard alliances such as the Ultra Ethernet Consortium and UALink. AMD's policy is to use an open full stack as the base environment and support the porting of workloads developed and optimized on local AI PCs to the cloud and data centers without code rework, thereby supporting long-term enterprise AI investment.

Source: IT DAILY · Kwon Young-seok
Original: https://www.itdaily.kr/news/articleView.html?idxno=241938

References

• https://www.itdaily.kr/news/articleView.html?idxno=241938