MemryX AI Unveils MX3+ Accelerator With More Than 1 POPS of Compute
MemryX AI has unveiled MX3+, a next-generation AI accelerator designed to deliver up to 10 times more peak AI compute and nearly 10 times more on-chip memory than MX3 within the same low-power operating envelope. The company is targeting the accelerator at edge devices, industrial systems and high-density inference infrastructure.
- MX3+ is designed for up to 10 times more peak AI compute than MX3.
- The accelerator offers nearly 10 times more on-chip memory within MX3's low-power envelope.
- A planned 16-chip system is designed to exceed 1 POPS and provide 2.16 GB of distributed memory.
- Engineering samples are targeted for late 2027, with volume production targeted for mid-2028.
- MemryX AI is accepting early-access requests from select partners and enterprise customers.
A scalable accelerator architecture
MX3+ builds on MemryX AI's production-proven MX3 dataflow architecture. Planned configurations will range from a single accelerator to 16-chip systems, with multiple accelerators operating as one logical device. According to MemryX AI, a 16-chip MX3+ configuration is designed to deliver more than 1 peta-operation per second, or POPS, of peak AI compute.
That configuration is also designed to provide 2.16 GB of distributed on-chip memory. MemryX AI says its distributed dataflow architecture allows the accelerators to work together without requiring the host processor to manage individual chips. The company's Neural Compiler automatically maps neural networks across the available accelerators, allowing developers to scale performance without manually partitioning their models.
Support for more complex AI workloads
MemryX AI says the additional compute and memory will expand the workloads MX3+ can address. These include higher-resolution vision, transformer-based models, small language models, vision-language-action models, hybrid architectures and more sophisticated multi-model AI pipelines.
The accelerator is also being positioned for industrial automation, robotics, intelligent transportation, smart infrastructure, video management systems, healthcare devices, retail and consumer electronics. Its near-memory dataflow architecture keeps model data close to computation, which the company says minimizes data movement and supports larger models and higher-resolution inputs.
MX3+ is designed for efficient, predictable, low-latency inference at a batch size of one. MemryX AI also says it will carry forward the development experience established with MX3, allowing developers to use a familiar process for compiling and deploying applications as model complexity and system scale increase.
Engineering samples targeted for late 2027
MemryX AI is targeting engineering samples of MX3+ for late 2027 and volume production for mid-2028. The company says additional product specifications, system configurations and availability information will be announced as development progresses.
MemryX AI is accepting early-access requests from select strategic partners and enterprise customers interested in co-validating pre-production hardware and software platforms. The company describes MX3+ as part of its longer-term architecture roadmap, preserving the modularity, scalability and distributed dataflow principles of MX3 while adding capabilities intended to carry forward into MX4, its future platform for advanced multimodal and large language model inference.
For customers, the announcement points to a planned route from embedded deployments to larger inference systems using the same architectural foundation. MemryX AI President and CEO Ross Jatou said MX3+ is designed to bring more capable AI into real-world systems while lowering total cost of ownership for high-density edge deployments.
PR Newswire: MemryX AI Unveils MX3+, Scaling AI Inference Beyond 1 POPS
