Intel reveals Hala Point, a brain-inspired computer in a microwave-sized box

Intel just built the world’s largest brain-inspired computer — and it looks remarkably like a microwave oven.

The system, dubbed Hala Point, packs 1,152 experimental Loihi 2 chips that mimic the human brain. Instead of shuttling data between separate compute chips and memory pools like traditional PCs and server racks do, it processes and stores info inside artificial neurons at the same time.

That architectural shift means Hala Point houses 1.15 billion artificial neurons and pulls off 380 trillion synaptic operations per second, taking up roughly the same space as your kitchen microwave. Intel claims it can tackle complex optimization tasks using 100 times less energy than standard hardware.

Why settle on 1.15 billion neurons? Intel’s Mike Davies admitted there was no hard technical barrier stopping them from going bigger. A billion was just a nice round number.

For comparison, the upcoming DeepSouth neuromorphic rig scheduled for later this year tops out at an estimated 228 trillion synaptic operations per second.

Yet, nobody is throwing away their Nvidia H100s just yet. The software side is basically in the Stone Age.

Davies noted that the actual bottleneck isn't packing more silicon together, but writing the software layers needed to translate messy real-world problems into something a brain-like chip can actually parse. Building anything bigger right now would just be wasting metal.

There is also a massive dispute over training. Intel floated the idea that neuromorphic gear could eventually allow AI models to learn continuously without needing costly retraining from scratch.

James Knight at the University of Sussex called that pure hype. Because current frontier models rely on thousands of GPUs running parallel compute batches, trying to train an equivalent model from zero on neuromorphic hardware — which crunches inputs sequentially — could take decades.

Davies conceded the point, clarifying that future neuromorphic systems won't train massive models from scratch. Instead, they might take existing pre-trained weights and handle continuous learning tasks far more efficiently.

Why it matters

Neuromorphic computing won't fix your cloud bills tomorrow, but it takes aim at AI's dirtiest secret: the catastrophic power wall.

Brute-forcing Transformer models on conventional chips is running out of economic and environmental runway. If general artificial intelligence requires true real-time adaptation and efficiency, imitating biological neural pathways looks like a much more viable endgame than building dedicated nuclear plants for server farms.

For everyday developers, Hala Point is strictly a spectator sport right now. The toolchains don't exist for mainstream production, and you cannot easily deploy your existing models to it.

For Intel, however, maintaining a dominant footprint in non-von Neumann architecture is a smart hedge. While it trails Nvidia in conventional AI accelerators, keeping a working foot in next-gen biological computing guarantees it won't be left behind if the industry finally breaks away from standard GPUs.