Duration/Format

Half-day workshop

• Invited talks

• Technical paper presentations

• Industry technical sessions

• Benchmarking competition presentations

Technical Session (2.5 hours)

Curated presentations covering recent and emerging research in AI systems and infrastructure, including work aligned with ASPLOS, ISCA, HPCA, and MICRO themes.

Example topics:

• GPU-NPU heterogeneous systems

• KV-cache optimization, CXL memory systems

• AI accelerator evaluation

• Compiler and kernel optimization

• Runtime optimization for heterogeneous infrastructures

The goal is to promote discussion between academia and industry researchers.

Benchmarking Competition (1.5 hours)

A distinguishing feature of MOA is a practical benchmarking and optimization competition powered by the AI-BMT platform (https://www.ai-bmt.com/) and organized in collaboration with FuriosaAI (https://furiosa.ai/).

Participants will benchmark and optimize modern AI workloads using the FuriosaAI virtual ISA (vISA) stack and execute workloads on remotely accessible RNGD AI accelerator systems. The competition combines benchmarking methodology with hands-on optimization, enabling participants to evaluate runtime behavior and quantify optimization effectiveness on production AI hardware.

Target Workloads

• Transformer models

• Diffusion models

• State Space Models (SSMs)

• Speech-to-Text models

Optimization Targets

• Kernel scheduling

• Memory access optimization

• Operator fusion

• Tensor layout transformation

• Compiler-assisted lowering and mapping

The session will feature finalist presentations together with invited technical talks on benchmarking methodologies and optimization techniques for modern AI accelerator systems.