A little closer and more in focus for the FPGA. By the way, there are four of them. They're all working together and that's why you see so many lights flashing. Market data running across the wire.

BOOM. Now people can start to see what's possible.

“The barrier to neuromorphic adoption is not the absence of a neuromorphic-specific methodology. The barrier is the continued treatment of neuromorphic hardware as a laboratory curiosity needing its own world or as a helper device for other compute. Neuromorphic silicon should be treated as a distinct class of processor standing on its own, as my demos demonstrate, with a first-class place in the heterogeneous compute ontology of the enterprise. A complete paradigm change is in order that both research institutions and commercial houses need to adopt.”

The piece in the screenshot below is ad copy wearing a lab coat. I won't even link to it and it can't get the basics right. Neuromorphic computing is not defined by analog circuitry. BrainChip's Akida, the most commercially mature neuromorphic processor on the market, is fully digital. The chips are not "largely research-stage" either. AKD1500 entered commercial availability with production shipments to customers three months before this article was published. And, the AKD1000 has been available now for years. The claim that the brain is "largely feedforward" is simply false, since cortex is massively recurrent and feedback projections rival or exceed feedforward ones. The framing is wrong too. Neuromorphic silicon is not out to replace GPUs. Instead, brain-inspired compute has become a peer that takes the work that suits it right alongside other processor types. Judging neuromorphic computing on raw throughput is like faulting the cortex for not being the retina. Worse, the "speculative" future the article describes is already running. I've built and measured learning without a backward pass on Akida silicon, including a fleet of AKD1500s that learned a previously unseen radar emitter after deployment, beating its frozen control on 80 of 80 runs at 0.93 AUROC. The article's proposal of many small models behind a lightweight router with cheap forward passes is something I have in production form: a three-million-parameter state-space model served through an unmodified vLLM endpoint across 56 nodes at three sites, orchestrated with IBM Spectrum Symphony, assisted by GPFS, a hundred models hot-swapped across ten chips in 0.155 seconds, and a fleet of 131 Akida nodes on six continents brought up in less than fifteen minutes. The article invokes cortical columns without mentioning that Akida was designed around them. No citations, no author, three ads for an app builder, and a conclusion that the field is years away from things that shipped this summer. Eighty-two documented builds at kevindjohnson.org say otherwise. I would be less direct here about the state of neuromorphic computing and the false information in an article like this. But, these kind of falsehoods lead the unaware astray and perpetuate further misunderstandings about the true state of neuromorphic computing, obviously making it into self-perpetuating LLM training runs, chat conversations, and more bad writing all around.