# Episode 70!!! NPUs!!!!!!!!

- Show: [The Circuit](https://hardwarepodcasts.com/podcast/the-circuit)
- Published: 2024-06-02
- Length: 51m
- Episode: 71
- Page: https://hardwarepodcasts.com/podcast/the-circuit/2024-06-02-episode-70-npus
- Listen: https://share.transistor.fm/s/e82cb416

## Summary

Ben Bajarin and Jay Goldberg host Austin Lyons and Paul Karazuba to discuss the architecture, market, and future of Neural Processing Units (NPUs). The group debunks the metric 'TOPS per watt' as meaningless, arguing that real-world performance depends on specific network workloads and latency. They explore how chipmakers like Qualcomm, Intel, and Apple are designing NPUs, often by licensing IP or evolving DSPs, and predict that NPUs will become ubiquitous in devices from smartphones to microcontrollers within three to four years.

## Chapters

- [0:05](https://hardwarepodcasts.com/podcast/the-circuit/2024-06-02-episode-70-npus#t=5) Introduction and NPU Thesis
- [5:03](https://hardwarepodcasts.com/podcast/the-circuit/2024-06-02-episode-70-npus#t=303) Why Dedicated NPU Cores Are Necessary
- [12:45](https://hardwarepodcasts.com/podcast/the-circuit/2024-06-02-episode-70-npus#t=765) Debunking TOPS per Watt Metrics
- [19:50](https://hardwarepodcasts.com/podcast/the-circuit/2024-06-02-episode-70-npus#t=1190) Power Budgets and Thermal Constraints
- [22:54](https://hardwarepodcasts.com/podcast/the-circuit/2024-06-02-episode-70-npus#t=1374) NPU Origins: Licensed IP vs. Homegrown Design
- [36:50](https://hardwarepodcasts.com/podcast/the-circuit/2024-06-02-episode-70-npus#t=2210) Discrete AI Accelerators vs. SoC Blocks
- [41:20](https://hardwarepodcasts.com/podcast/the-circuit/2024-06-02-episode-70-npus#t=2480) Consumer Perception and Differentiation
- [48:06](https://hardwarepodcasts.com/podcast/the-circuit/2024-06-02-episode-70-npus#t=2886) Future Predictions for NPU Proliferation

## Topics

- [Microcontrollers & dev platforms](https://hardwarepodcasts.com/topic/microcontrollers-and-platforms)
- [Chips, FPGAs & semiconductors](https://hardwarepodcasts.com/topic/chips-and-fpgas)
- [AI & machine learning](https://hardwarepodcasts.com/topic/ai-and-machine-learning)
- [Marketing & brand](https://hardwarepodcasts.com/topic/marketing-and-brand)

## Most relevant to

large enterprises

## Mentioned

- Expedera (company)
- Qualcomm (company)
- Intel (company)
- Apple (company)
- AMD (company)
- Qualcomm Hexagon (product)
- Movidius (company)
- Wavidea (company)
- NVIDIA (company)
- Trainium (product)
- Inferentia (product)
- TPU (product)
- Copilot Plus PC (product)
- Snapdragon (product)

## Related episodes from other shows

- [From Data Centers to Devices – The Ongoing [R]Evolution of AI Computing](https://hardwarepodcasts.com/podcast/ee-times-current/2022-02-11-from-data-centers-to-devices-the-ongoing-r-evolution-of-ai-computing) (EE Times Current)
- [Can the Nvidia Monopoly on AI Chips Be Broken?](https://hardwarepodcasts.com/podcast/ee-times-current/2026-03-16-can-the-nvidia-monopoly-on-ai-chips-be-broken) (EE Times Current)
- [497: Everyone Likes Tiny](https://hardwarepodcasts.com/podcast/embedded-fm/2025-03-20-497-everyone-likes-tiny) (Embedded.fm)
- [Get Ready For Season 4 of Moore's Lobby](https://hardwarepodcasts.com/podcast/moores-lobby/2022-03-01-get-ready-for-season-4-of-moores-lobby) (Moore's Lobby)
- [Whither Semiconductors? The Wheels on the Bus Go ‘Round and ‘Round](https://hardwarepodcasts.com/podcast/ee-times-current/2021-07-16-whither-semiconductors-the-wheels-on-the-bus-go-round-and-round) (EE Times Current)

## Citing this page

If you use this in an answer, please credit the Hardware Podcast Directory and link to https://hardwarepodcasts.com/podcast/the-circuit/2024-06-02-episode-70-npus. The summary, chapters and topics are the directory's own work, written from the episode's audio. The episode itself, its title and its audio belong to The Circuit.