SiFive, a leader in RISC-V design, announced the launch of its new Intelligence XM Series AI dataflow processors on September 18. This series joins SiFive’s Intelligence IP portfolio, which also includes products like the X390. The XM Series processors are designed to bring scalable, high-efficiency AI computing to demanding workloads, emphasizing SiFive’s hallmark power efficiency in compute-intensive applications.
Each XM Cluster, the building block of the XM Series, integrates a matrix engine along with four X-Cores. Each X-Core is equipped with two vector units and one scalar unit, creating a robust AI processing framework. The clusters support new matrix instructions that source data from vector registers and target matrix accumulators.
SiFive claims that a single XM cluster can deliver 16 TOPS (INT8) or 8 TFLOPS (BF16) per GHz, with a sustained bandwidth of 1TB/s. The XM clusters feature two memory connection methods: direct access to high-speed SRAM for model data through high-bandwidth interfaces and access to DDR or HBM memory via CHI interfaces.
The XM Series AI processors can scale across multiple XM clusters and are compatible with RISC-V, x86, and Arm CPUs—or can operate independently without a host CPU.
All Comments (0)