• SEMIFIVE has begun mass production of HyperAccel's Bertha AI inference accelerator using Samsung Foundry's 4nm process
  • The first production contract is secured, while further orders are expected as HyperAccel expands its services

The fact

SEMIFIVE said on 8 September that it had begun mass production of HyperAccel's Bertha AI inference accelerator using Samsung Foundry's 4nm process. The chip is designed for large-language-model inference and has a die area of more than 500 square millimetres.

SEMIFIVE is providing work from design and verification through packaging, testing and volume manufacturing supply. The companies first announced their production agreement in October 2024. SEMIFIVE says an initial production contract has now been secured and expects additional purchase orders as HyperAccel expands its services. It has not disclosed the number of chips being produced, shipment volumes or a recurring production schedule.

The assessment

HyperAccel now has production hardware behind Bertha rather than only a development plan. That gives customers a chip they can evaluate for their own inference workloads and moves the project closer to deployment in data-centre systems.

The production announcement does not yet show how much hardware will be available. SEMIFIVE has secured the initial order, but further production depends on HyperAccel placing more purchase orders as its services expand. A customer planning capacity around Bertha would therefore need to understand both how the chip performs and how much supply HyperAccel can secure.

For BTW readers, Bertha is now closer to becoming usable AI infrastructure, but the size of that deployment is still unclear. Repeat production orders and named customer deployments would show whether the initial manufacturing run is turning into sustained capacity.

What to watch

Watch for follow-on purchase orders, disclosed shipment volumes and named deployments using Bertha. Customer workload results would provide evidence of how the chip performs after installation, while repeat manufacturing orders would show whether HyperAccel is expanding beyond the initial production run.