DSP Suppliers Take Different Paths: Motor Control, Audio AI and Server Interconnects
A DSP in a motor drive, a smart speaker and an AI-server optical module addresses three very different problems. Motor control depends on the time between sampling and an output update. Audio processing must fit sound-related workloads into power and latency budgets. Optical links need to recover high-speed data after transmission has distorted the signal.
Product announcements in 2026 make those distinctions clearer. Control vendors are developing multicore execution and protection; audio suppliers combine DSP and AI compute; interconnect companies advance 1.6T links and higher lane rates. For buyers evaluating Chinese suppliers alongside international companies, application, development support and delivery stage provide more useful comparisons than the shared DSP label.
DSP describes both a processor product and an integrated capability
Digital signal processors repeatedly execute operations such as filtering, multiply-accumulate and transforms. They can be independent chips or cores within a system-on-chip. High-speed communications products may use highly specialized processing architectures for a defined link.
Purchasing a controller, purchasing an audio SoC containing a DSP and licensing a DSP core are different transactions. Adding their revenue or shipments can double-count activity or attribute an entire SoC's revenue to DSP alone.
Control, audio/edge processing and high-speed interconnects provide three useful application groups, although DSP also serves baseband communications, radar and industrial measurement. The suppliers discussed here are not a complete market roster.

Figure 1. Suppliers grouped by the applications and supply-chain roles discussed in this article. Companies may span multiple areas. This is neither market share nor a complete portfolio map.
Control performance depends on latency and determinism from sampling to output. Audio evaluation involves channels, algorithm load, sound quality and power. Optical interconnects require lane rate, error performance and link-power analysis. MHz, TOPS and Tbps describe different things and cannot form a universal DSP ranking.
Control products compete through peripherals and tools as well as compute
A power, motor or inverter controller must complete feedback tasks within a specified interval. ADCs acquire current and voltage, the processor executes the algorithm, and PWM updates the output. Comparators and protection logic must respond to faults. Coordinating that sequence is more relevant than clock frequency alone.
Texas Instruments' C2000 remains an important reference. A peripheral guide updated in July 2026 and manuals for families including F29H85x updated in September show continuing development. Newer products include C29x cores and configurations such as lockstep and separate-core operation. A meaningful comparison selects a particular model and mode rather than treating one older device as the whole TI portfolio.
| Supplier | Product or disclosed progress | Evaluation focus |
|---|---|---|
| Jinxin Technology | AVP32F00157: 120 MHz floating-point control DSP with dual-core lockstep | Execution consistency, ADC/PWM coordination and fault protection |
| Hefei Qianxin | QXS320F28377LRevB: 300 MHz dual-core device with QX-IDE | Task allocation, debugging and control-code migration |
| Liuyue Micro | Industrial control DSPs and development software; recent new-product tape-outs | Returned-silicon validation, tools and supply progress |
| Texas Instruments | C2000 and newer real-time control products including F29H85x | Model/generation differences, peripheral behaviour and software-reuse costs |
These are different product approaches, not a ranking of equivalent devices.
The two Chinese dual-core examples should not be interpreted identically. Jinxin emphasizes lockstep execution, using corresponding computations and comparison to detect inconsistencies. Qianxin's dual-core implementation can divide control tasks. Execution checking and task allocation serve different objectives.

Figure 2. AVP32F00157 evaluation board from Jinxin's October 16, 2026 product article. The manufacturer stated that samples and development boards were available at that time.
Qianxin's manuals describe dual-core debugging, real-time variable refresh and project migration. Jinxin highlights the hardware protection path from comparator to PWM. Liuyue emphasizes the combination of cores, compilers and integrated development environments. Its October website article reports an August tape-out activity, which remains distinct from subsequent volume production.
On an actual board, engineers still need timely algorithm execution, aligned sampling and PWM, and defined output behaviour after a protection event. Suppliers create commercial value by helping customers complete that journey from code to circuit.
Audio DSP and AI compute are being integrated in different ways
Multichannel equalization, echo cancellation, mixing and spatial audio require continuous low-latency processing. Voice separation and intelligent noise reduction introduce neural-network workloads. Added capability must fit the power limits of devices that may operate for long periods.
Analog Devices and Cadence announced collaboration on a new generation of SHARC audio processors in September 2026. The ADSP-SC84x/2184x platform uses SHARC-FX architecture for applications including automotive audio, combining traditional signal processing with AI workloads. The announcement also covers compilers, development environments and audio-development frameworks.
This approach emphasizes programmable audio processing. Engineers continually adjust algorithms and signal paths for multiple speakers, noise cancellation and voice functions. Reusing existing algorithms and tools influences project cost alongside available processing capacity.
Actions Technology's ATS362X illustrates a different level of integration. September material describes a SoC in which CPU, DSP and NPU cooperate. The company's 2026 interim report says the series was already used in branded customers' volume-produced party speakers and was expanding toward professional audio.
The DSP performs signal processing, the NPU accelerates suitable neural-network operations, and the CPU handles system control and related work. Exact allocation depends on hardware and software. Adding an NPU does not eliminate the DSP's audio tasks.

Figure 3. Conceptual allocation of common audio-SoC tasks. This is not an internal block diagram of a particular chip. Algorithms can be deployed differently across implementations.
ADI and Actions can be compared by technology direction and integration approach, but shared audio positioning does not establish direct substitution. Automotive and consumer products differ in interfaces, reliability, algorithms and lifecycles. An “AI-capable” label is insufficient to standardize the comparison.
For Chinese suppliers, a practical opportunity is to combine signal paths, algorithms and software around a defined speaker, wireless-audio or professional-device application. Standalone DSP revenue may not capture the value of that integrated offering.
Interconnect competition centres on bandwidth, link behaviour and power
Optical DSPs process high-speed link signals and work with equalization and error-correction mechanisms to preserve receive quality. A 1.6T bandwidth rating measures something different from a control processor's clock or an audio accelerator's compute capability.
Marvell expanded its 1.6T optical-DSP portfolio in March 2026, addressing power consumption, link reliability and rate conversion through products including Ara T, Ara X and Petra. Equal aggregate bandwidth can therefore coexist with different power budgets, interfaces and link requirements.
Broadcom's Taurus BCM83640, announced in March, emphasizes a 400G-class optical lane. Its product brief describes conversion between eight 200G-class electrical lanes and four 400G-class optical lanes, retaining 1.6T aggregate bandwidth. The change in lane count does not mean every interface becomes 400G.

Figure 4. Application diagram from Broadcom's March 2026 BCM83640-DIE product brief, showing the DSP with laser drivers, photodetectors and TIAs. This is a crop of the original document; 200G and 400G retain its rate-class terminology.
Higher lane rates require coordinated optical devices, electrical interfaces and packaging. A DSP's 1.6T rating alone cannot establish module power, error performance or reach. Module vendors must validate the complete link, while chip suppliers work with optical components, switches and system platforms.
AEC DSPs serve electrical interconnects in active copper cables rather than optical conversion. Marvell's October announcement of its second-generation 1.6T Alaska A AEC DSP identifies copper connectivity and sampling status. Optical DSPs, AEC DSPs and retimers may use related processing techniques, but their product identities and circuit positions are not interchangeable.
These products are closely associated with AI-infrastructure expansion and require high-speed interfaces, advanced processes and system validation. Success in control or audio DSP does not establish high-speed optical-DSP delivery capability.
Processor IP changes the relationships between suppliers
Cadence and Ceva supply processor IP and software tools for chip designers, rather than occupying exactly the same position as finished-chip suppliers. They can compete with proprietary cores while also partnering with semiconductor companies.
The ADI–Cadence audio collaboration is one example. Ceva's November disclosure about Mobilint's REGULUS project describes SensPro DSP and NeuPro-Nano NPU integration in one SoC for tasks including signal processing and neural-network inference.
Two distinctions are therefore necessary: whether the customer buys a finished device or design IP, and whether competition concerns an application solution or the core/toolchain. A finished chip can use licensed compute IP and differentiate through peripherals, algorithms and system implementation.
AI is changing DSP task allocation in several ways. Some DSPs gain AI capability, some cooperate with NPUs, and others remain focused on real-time control or transmission. The relevant comparison is the combination that completes the workload within its power and latency constraints.
Chinese suppliers show different entry paths. Jinxin, Qianxin and Liuyue focus on industrial control and development support; Actions participates through audio SoCs. High-speed interconnects require a separate assessment of interface technology, optical/electrical integration and module validation.
Switching costs also vary. Control products require code and fault-protection revalidation; audio products require algorithm, interface and sound tuning; interconnect products require renewed module and system interoperability tests. Price advantages become commercially persuasive when they cover that additional work. Reusable tools, algorithms and reference designs can support subsequent orders.
Progress should be traced through particular products: complete documentation, available samples, customer testing, volume production and reuse across later projects. Tape-out, sampling and end-product production answer different questions. Comparing suppliers against the same task, similar system conditions and documented delivery stages reveals where their advantages actually lie.



