What software ecosystem was highlighted during the event?

The system ran AMD ROCm 10.0, executing the Gemma4-E2B large language model to showcase open-source AI acceleration on a RISC-V host platform.

Back to feed

#AI & Compute

AMD and SiFive Validate ROCm Stack on RISC-V Datacenter Platform for AI Inference

Originally published as: AMD’s ROCm Runs on on SiFive’s BigSky Datacenter Development Platform

Embedded Computing DesignPublished 4 days ago

AI overview

Hardware teams evaluating alternative architectures can now look toward validated RISC-V and GPU pairings backed by an open-source software stack for AI inference.

At the AI Infra Summit, SiFive and AMD showcased a major software ecosystem milestone by running AMD ROCm 10.0 on the SiFive BigSky Datacenter Development Platform. The demonstration utilized SiFive Performance P870-D CPU cores as the host head node, paired with AMD Radeon AI PRO R9700 GPUs handling LLM inference offload for the Gemma4-E2B model.

This collaboration addresses a long-standing bottleneck in alternative architectures: the lack of mature, enterprise-grade open-source software stacks for non-x86 hosts. By porting ROCm to a high-performance RISC-V host processor, the companies validate a viable open-standard path for datacenter AI infrastructure, reducing software-dependency lock-in for hyperscalers and system integrators.

Hardware architects and platform developers should monitor how quickly this validation translates into production-ready software support, optimized compiler toolchains, and validated board support packages (BSPs). While currently positioned as a development and validation platform, expanding ROCm beyond traditional host ecosystems opens up new architectural possibilities for heterogeneous computing clusters.

Related components

Questions & answers

The demonstration featured SiFive's BigSky Datacenter Development Platform powered by SiFive Performance P870-D CPU cores acting as the host, paired with AMD Radeon AI PRO R9700 GPUs for inference offload.

Discover more

#Memory & Storage

Astera Labs Expands Leo Smart Memory Controller Family to Mitigate AI Server Constraints and Support Memory Reuse

Astera Labs has updated its Leo smart memory controller lineup to include the fabric-attached X-Series alongside new E-Series and P-Series devices, easing memory bottlenecks in agentic AI architectures. These components aim to stretch tight memory supplies by pooling resources and allowing legacy DDR4 reuse alongside DDR5.

EE Times

#MCU & Embedded

Power-Efficient Processing Considerations in Embedded Systems

Surging demand for real-time edge AI and signal processing in industrial systems forces hardware teams to balance high compute throughput with strict thermal and energy budgets. Selecting power-efficient silicon architectures is critical to preventing thermal throttling and meeting tight battery-life constraints in remote deployments.

Embedded Computing Design

#Optics & Connectivity

TeraSignal Introduces Analog Wave Processor for 200G/Lane AI Interconnects: Design Evaluation and System Impact

TeraSignal has unveiled an Analog Wave Processor architecture designed to perform math operations directly on broadband electrical waveforms for 200G/lane interconnects. This development impacts next-generation 1.6T and 3.2T AI infrastructure by delivering advanced signal correction without the high power consumption or latency of traditional digital signal processors.