Scale-up architecture increases a single system's capability by adding more resources, such as more GPUs tightly coupled by NVLink and NVSwitch within one node or NVLink domain, so they behave as one large accelerator with shared, coherent memory. It excels at workloads needing high intra-node bandwidth, like tensor parallelism, but is bounded by how many devices one domain can hold. It contrasts with scale-out across many networked nodes.
Category: IT & Compute Infrastructure · Subcategory: AI & HPC Infrastructure · Part of the AIDC Technical Terminology reference.