Skip to content

Questions about the chips, the IP and evaluation.

Collected from the questions customers ask us most, so teams can judge the fit quickly.

What does Fanshi Semiconductor offer?

ASIC chips for edge LLM inference (Forge, Trident and Argus), custom chips built around a customer’s model, FPGA evaluation boards, and inference core IP (FS-Fusion, FS-Attention, FS-MoE and FS-Linear) that can be deployed into a customer’s own ASIC or FPGA.

What is an edge inference chip, and how is it different from GPU inference in the cloud?

An edge inference chip runs the model on the device itself, with no network round trip and no data centre. Compared with cloud GPUs there is no per-token bill, no network latency and no data leaving the device. The trade-off is that model size and update strategy are decided at design time.

What is the difference between compute-in-memory and near-memory compute?

Compute-in-memory performs multiply-accumulate inside the memory array, so the weights never leave the storage cells. Near-memory compute places compute units right next to high-bandwidth memory and uses on-chip streaming to keep data movement to a minimum. The first gives the lowest power and marginal cost; the second suits models that are still evolving.

When does MaskROM compute-in-memory make sense?

For edge devices with a stable model, a clear shipping volume and a hard requirement for low power and low marginal cost per inference, such as handheld translators and offline voice assistants. The weights are cast into the silicon at the mask stage of tape-out, so there is no external weight traffic at runtime.

Can the model be updated once it is cast into the chip?

Not on the MaskROM route: the weights are fixed at tape-out, so a new model means new masks. The SRAM compute-in-memory route loads weights at boot and can be updated. The near-memory route keeps weights in external memory and supports continuous iteration. That is why the first question is always: how often does your model change?

Which models do Fanshi chips support?

Forge targets Qwen 3.5 0.8B; Trident supports up to Qwen 3.5 35B; Argus runs SmolVLM, SegFormer and similar vision and multimodal models. The IP family covers Qwen3 / 3.5 (0.6B to 9B), SmolLM2, sparse MoE models and DeltaNet-style linear attention. See the supported models page for the full list.

Do you support Qwen3.6, Qwen3.8 and other new models?

Yes. Qwen3.6 and Qwen3.8 keep the Qwen3.5 hybrid layout: Gated DeltaNet linear-attention layers interleaved with gated full-attention layers. Fanshi’s FS-Linear and FS-Attention cores already cover both layer types, so a newer model in the same family is a new chip build on existing cores, not a new architecture. Parameter size, context length and performance targets are assessed per project.

What happens when a newer model is released after the chip is made?

On the MaskROM route, moving to a new model means re-making the mask layers that hold the weights. If the new model shares the same architecture, the core design, SDK and host integration carry over unchanged. On SRAM compute-in-memory you reload the weights at boot; on near-memory compute you update the weights in external memory.

How are figures such as 17,000 token/s defined?

They are decode-throughput design targets for the core model listed with each product. The exact model, precision, batch size and test platform are confirmed per project.

How does on-device inference keep data secure?

Inference runs entirely on the device, so neither the raw input nor the model output has to be uploaded, and the device can run fully offline. For air-gapped, classified or privacy-sensitive deployments, this is a guarantee of the architecture rather than a policy promise.

How much can a chip save compared with paying for API calls?

The chip is a one-time purchase, after which the marginal cost of each inference is close to zero, while an API bills every token. The cost calculator on the home page estimates the three-year total from your daily token volume and current model; actual figures depend on usage and quotation.

Can the inference IP run on an FPGA?

Yes. FS-Fusion, FS-Attention, FS-MoE and FS-Linear target both ASIC tape-out and FPGA prototyping, with an evaluation board for early validation, so an engineering sample can exist before silicon.

What does an IP licence include?

The inference core IP, an FPGA evaluation board, model mapping and interface adaptation, and design-in support for the project. Coverage, resource usage, performance targets and licensing terms are evaluated against the target device and model.

How do I request a sample or an evaluation kit?

Email info@fanshi-silicon.com with the target model, performance goal, power budget, interfaces and project timeline. We confirm the route and reply with a plan.

What is in the evaluation kit?

Packaged engineering samples on an evaluation board, host connectivity, the SDK and drivers, integration documentation, and direct support from the Fanshi engineering team.

Can we sign an NDA before discussing technical details?

Yes. Detailed specifications, evaluation data and development plans are shared once a mutual NDA is in place.

Do you work with customers outside China?

Yes. Fanshi works with teams around the world, in English or Chinese. We are based in Hefei, China (China Standard Time, UTC+8).

What process and supply chain does Fanshi use?

A mature SMIC 28 nm supply chain. The gains of compute-in-memory and near-memory compute come from architecture rather than leading-edge nodes, which keeps cost and supply predictable.

What is the relationship between Fanshi and Taalas?

There is none: no affiliation and no partnership. Fanshi works in the same broad direction of casting models into silicon, and develops its technology independently on a 28 nm supply chain.

What is the memory wall, and why does it decide whether edge LLMs work?

Every generated token requires reading all of the model’s weights from memory into the compute units, so bandwidth and power go mostly into moving data rather than computing. Edge devices have a small fraction of a data centre’s bandwidth and power budget, so as long as the weights have to move, large models cannot run at the edge. Compute-in-memory and near-memory compute exist to get around that wall.

Are the figures on this site final delivery specifications?

No. They describe the current product direction and design targets. Final specifications, test conditions, delivery versions and commercial terms are confirmed in project documents.

Which industries are Fanshi chips for?

Handheld translation and voice hardware, offline in-car voice, embodied robots, industrial edge servers, air-gapped and secure terminals, drone-based inspection, production-line defect detection and multi-sensor fusion.

Where is Fanshi based, and how do I get in touch?

Hefei Fanshi Semiconductor Technology Co., Ltd. is based in Hefei, Anhui, China. For business, IP licensing and careers, email info@fanshi-silicon.com.

Have another project question?

Send the target model, performance and power goals. We reply with a route and an evaluation plan.

Request an evaluation kit