PCIe add-in card
IPROVF1 Inference Card
A full-height PCIe Gen5 ×16 accelerator for 3B–20B models at interactive latency, with up to ≈50B parameters resident in FP8.
IPROVF · PCIe Gen5 ×16 inference accelerator
Inference silicon that thinks faster.
Generating text means reading the whole model out of memory for every word it produces. That makes memory bandwidth, not raw compute, the thing that decides how quickly an answer comes back. MediaChipset accelerators are built around that: enough memory bandwidth that a single request runs at full speed, instead of waiting for others to pile up alongside it.
Pre-silicon specifications are design targets.
Overview
Silicon for running models, not training them.
MediaChipset builds inference accelerators: a PCIe card that drops into a server you already own, or a complete rack system. Both run the frameworks your team already uses. Alongside them, our high-speed switch IC connects servers and racks into one network.
Built for inference
Training a model and serving it are different problems. Our silicon is designed for the second one: answering requests from models that are already trained.
Fast on a single request
A chat turn, an agent step or a robot control loop arrives on its own. We size memory bandwidth so that case runs at full speed, rather than only paying off at very large batch sizes.
Runs your existing stack
Models arrive through PyTorch, Hugging Face and vLLM. Our compiler and runtime take them to silicon, and every configuration in the family runs the same stack.
Products
From a single card to the network that connects the racks.
Start with a card that drops into any PCIe Gen5 server, or deploy a MediaChipset server integrated and validated with the serving stack. Our high-speed switch IC links those servers and racks together.
Switch IC
High-Speed Switch IC
Ethernet switch silicon for data center and AI cluster networks, linking the servers in a rack and the racks across a data center at high bandwidth and low latency.
Rack server
AI Inference Server
Four or eight IPROVF1 cards in a validated system, shipped with the driver, runtime and vLLM serving stack installed. Serve many mid-size models side by side, or pair cards for 50B–100B models.
Customer data room
Datasheets and engineering documents, under NDA.
The full IPROVF1 datasheet and engineering documentation are shared with customers through a secure data room. Every download is logged and stamped to the recipient.
Request access
Tell us who you are and what you're evaluating. Use your business email.
NDA verified
We confirm a mutual NDA is in place, or send you one to sign, then approve your account.
Download
Set your password from the single-use invitation and download documents for the term of your NDA.