IBM NorthPole
A research chip with no separate memory at all. NorthPole claims 25 times the energy efficiency of a conventional design by refusing to move data, and IBM has already run a Granite model on it.
Almost every computer ever built keeps its memory in one place and its brain in another, and spends enormous effort shuttling data between them. That trip is where most of the electricity goes. NorthPole's answer is blunt: stop taking the trip.
Picture a kitchen where the ingredients are already laid out at each station instead of stored in a walk-in fridge down the corridor. Nobody walks anywhere. The cooking is not faster because the chefs got quicker, it is faster because the walking disappeared. NorthPole has no off-chip memory. The memory is woven through the compute.
The published numbers
In one demonstration the chip handled dense, high-resolution video and picked out multiple classes of object while drawing roughly 5 watts. That is less than a phone charger. In September 2024 IBM presented results at the IEEE High Performance Extreme Computing conference showing NorthPole running a 3 billion parameter Granite language model.
What it means for a Power buyer today
Nothing you can order. That is the honest answer and it should stay the honest answer until IBM says otherwise. What NorthPole tells you is the direction of travel, and the direction is efficiency rather than raw size. The Spyre card already reflects it: more than 300 TOPS inside a 75 watt envelope is a low-power design, not a brute-force one.
NorthPole has no announced product, no price, and no place on a Power roadmap. Do not let a vendor use it in a proposal. If you need AI acceleration on Power this year, the Spyre Accelerator and on-chip Matrix Math Acceleration are the two real options.
Related
IBM Artificial Intelligence Unit (AIU)
The research chip that became the card you can order. The AIU is why the Spyre Accelerator has 32 cores and why those cores look so much like the AI engine already inside an IBM Z processor.
IBM Spyre Accelerator for Power
A 75 watt PCIe card that adds more than 300 TOPS of AI inference to a Power 11 server. It is the only IBM AI chip an AS/400 shop can actually order as an add-on, and it comes with a short list of things it refuses to work without.
IBM NanoStack 0.7 nm Transistor
IBM says it has built the first transistor smaller than one nanometre, stacking them in three dimensions to fit nearly 100 billion onto a fingernail. It is the answer to a question every AI buyer keeps asking: does this keep getting cheaper?
Power 11 Matrix Math Acceleration
The AI acceleration most AS/400 shops already own and have never switched on. It is inside every core of the processor, it costs nothing extra, and it is the cheapest honest answer to a first AI pilot.