Expedera Inc

08/20/2026

The edge AI landscape is evolving fast. In our newest blog post, Expedera Chief Scientist Sharad Chole unpacks a major shift: vision-only processors are giving way to hybrid systems where LLMs, VLMs, and generative AI run alongside traditional perception networks — all on-device, in smartphones, vehicles, and embedded gateways.

This isn't just an upgrade — it's a rethink of what edge processing needs to handle next. https://semiengineering.com/packet-based-npus-in-the-llm-era-from-compute-bound-cnns-to-memory-bound-edge-and-automotive-workloads/

08/19/2026

The future of AI compute won't run on just one kind of chip. In a new roundtable discussion, Expedera's Sharad Chole joins a panel of experts to explore how power, token cost, memory bandwidth, and workload complexity are pushing data centers toward heterogeneous clusters — blending CPUs, GPUs, NPUs, optics, and custom accelerators.

The conversation digs into inference disaggregation, parallelization strategies (tensor, data, context, pipeline), and why power and token efficiency are the two biggest challenges shaping what comes next. https://semiengineering.com/the-future-of-ai-compute-wont-run-on-just-one-kind-of-chip/

Power, token cost, interconnects, and software orchestration are pushing data centers toward heterogeneous clusters built from CPUs, GPUs, NPUs, optics, and custom accelerators.

06/11/2026

Expedera's Chief Scientist, Sharad Chole, explores how mismatches between layer dimensions and hardware blocks cause inefficiencies: smaller layers leave compute idle, larger ones require fragmentation with repeated memory accesses. https://semiengineering.com/how-to-start-building-edge-native-ai/

05/15/2026

Expedera's Athish Rahul Rao argues that the core hardware question is no longer how many TOPS can fit within a given power and area budget. It is whether an architecture is built around real multimodal workload behavior, especially memory movement, activation lifetimes, utilization under irregular graphs, and the software needed to schedule all of it effectively.

Peak TOPS is becoming a weaker proxy for delivered edge performance.

05/07/2026

Let’s define why agentic AI is different from generative AI. First and foremost, there is a notion of autonomy. Generative AI is a prompt, and then you come up with a response. Agentic AI has more autonomy in high-level tasks. You’ve given them high-level tasks, and they are responsible for orchestrating, planning it out, and coming up with how to follow through.

Long‑running agents, tool-calling LLMs, and multimodal chaos are rewriting edge compute rules, and making chip design more challenging.

05/01/2026

"If you think you understand how agentic AI will be used at the edge, then you don’t understand agentic AI yet." Agentic AI is likely the next evolutionary step in edge inference. How it all plays out always comes back to three things: how much power it consumes, how much data movement it requires, and how much compute it needs. Expedera Chief Scientist, Sharad Chole, shares his perspective in this Semiconductor Engineering Roundtable.

As models evolve faster than silicon cycles, experts weigh how much adaptability architects can afford without sacrificing power, area, or efficiency.

04/10/2026

Expedera Chief Scientist, Sharad Chole, explains that Edge AI processing is not just a hardware architecture challenge. It starts with the models, with quantization, and with the application. This is a whole-stack problem. https://semiengineering.com/fast-isnt-fast-enough-redefining-metrics-for-edge-ai/

Why latency guarantees, memory movement, power budgets, and rapid model deployment now matter more than raw TOPS.

04/09/2026

Expedera's Chief Scientist, Sharad Chole, makes an interesting case that Edge intelligence is hampered not by a lack of compute but by the underutilization of available resources.

Edge intelligence is hampered not by a lack of compute, but by the waste of it.

04/03/2026

We have to come up with hardware architectures that exploit the network architecture itself. Edge devices are essentially bandwidth-limited. Training is done using multiple HBMs. But on the edge, there is literally one LPDDR, or not even 64 channels, maybe even a smaller-channel LPDDR that gets deployed on low-cost edge devices. That means bandwidth management becomes a critical part of how we execute things on edge inference. This white paper examines technical challenges, architectural innovations, and benchmarks to help OEMs successfully transition to edge-native AI. https://www.expedera.com/next-generation-ai-transitioning-inference-from-the-cloud-to-the-edge/