In honor of the LSF Community Summit in Santa Clara this coming Monday (June 1st, see you there!), I built a demo using BrainChip's AKD1000 chips with LSF. So, instead of Symphony, this time we're using IBM Spectrum LSF to enable batch runs with Akida as opposed to the service-oriented capability Symphony affords. LSF is actually widely used in the semiconductor design industry as well as one of the chief engines for training large language models at scale. When I say scale, I'm talking about thousands of nodes, multiple clusters worldwide, all using LSF for workload performance and efficiency.
Repurposing a chip is required for multi-tenancy and I've already built the capability in Symphony to use multi-modal, multi-model, multi-domain usage with the neuromorphic hive mind whether through these ten AKD1000s or through software simulation, on-prem, or in IBM Cloud. Implementing LSF with Symphony means I can attach a chip to run any batch job or, in this case, to switch models and have LSF monitor the Akidas through ELIMs.
Doing so exposes live chip state to LSF as custom resources we can schedule and maintain, monitor whether a chip is free, which model is currently mapped, which tenant last used it. This makes LSF an informed workload manager that coordinates multi-tenant access cleanly and keeps the fleet's duty cycle high, enabling queued access priority and availability at scale. We can do things like pre-warm models for available chips on the fly, making the inference immediately available as needed, cutting even more milliseconds out of the loop for performance sake.
What makes Akida a good citizen here in using both Symphony and LSF together is a compute ontology that doesn't lock you into doing things only one way or working with just one platform. You aren't trapped with a massive single vendor dependent GPU stack, either, and still get the benefits AI delivers. Constant repurposing is possible here, meaning, the same low-power fleet can absorb scheduled batch work overnight and serve online inferencing by day, LSF and Symphony together make this possible at scale.