BrainChip Second-Gen Architecture: Transformers, TENNs and STDP

Find the Latest SoC Solutions
for... Automotive... IoT... Security... Audio... Video

In this podcast, Sally interviews BrainChip CMO Nandan Nayampally about the company's second generation architecture for spiking neural network acceleration, which now includes hardware support for transformer networks like ViT.

www.eetimes.com/, Dec. 13, 2023 – 

SWF: Welcome to AI with Sally, a podcast that takes a closer look at some of the most interesting technology stories on artificial intelligence and machine learning. We'll hear about the latest in hardware and software that has a big impact on the world of AI. I'm your host, Sally Ward-Foxton.

Welcome back, everyone. In this episode, we'll be hearing from BrainChip about their spiking neural network accelerator, how it can spot transformers, and a new type of neutral network they've invented which combines spatial and temporal convolutions to efficiently process sequential spatio-temporal data. That interview with BrainChip's CMO, Nandan Nayampally, is later in this episode.

But first, here's a little roundup of some people I've been talking to lately.

Raja Koduri, Intel's former GPU chief architect, called me up to tell me about his new company, MIHIRA, and what it's going to do. MIHIRA will be a data center operator, a content creation company that will offer several layers of service. There will be a heterogeneous data center hardware architecture made up of CPUs for rendering, gaming GPUs, and various AI accelerators. Part of the secret sauce will be in the software orchestration layer, which handles distributing the workloads across heterogenous hardware. Roger has licensed some software from Intel's Project Endgame to build on. Endgame is a project that Intel recently abandoned which was going to be a unified gaming services layer for the cloud.

MIHIRA will also build a content creation studio in India to make digital content like movies and digital twins. Workloads from the content creation studio will be used to test MIHIRA'S hardware and software as it develops. Roger told me he hasn't completely ruled out working on new silicon IP going forward. My article has more details on MIHIRA and Roger's plans. I'll put a link to my exclusive interview on the podcast page at eetimes.com.

AI chip start-up Lemurian Labs has come up with an interesting idea using logarithmic number formats more commonly used in DSP maths for AI acceleration. Logarithmic formats have been used for ages in DSPs, where the algorithms have tons of multiplication operations, which is easy to do in hardware with a logarithmic format. Addition, not so much. But for AI, we obviously need multiplication and addition for matrix multiplication. So Lemurian has come up with a hardware solution for logarithmic addition. And the number formats it's invented are a bit special. They use multiple bases and multiple exponents, but they have much better coverage in the range around zero than FP8 does, which is where we need it for AI. So hopefully they can get away with lower precision for the same accuracy. The company is building a data center chip designed for its number formats, and they hope to tape out in Q3 24. Initial simulation results show it can outperform the H100.

click here to read more...

 Back