Arthur Hanson
Well-known member
Any information on what fab makes the IBM North Pole chip and if it is as powerful and energy efficient as they say it is. Thanks
?s on IBM North Pole chip, where is it now
- Thread starter Arthur Hanson
- Start date
Any small-model chip can be efficient, due to efficiency of SRAM especially for batch-1. Batch size 1 inference is dominated by reading the parameters if the parameters have to be in DRAM due to model size.
Alternative like resistive memory on chip will be interesting if density significantly exceeds what can be done with SRAM, or if significantly cheaper, assuming they are similar energy. An SRAM delivers a byte of data for much less than 1pJ.
Alternative like resistive memory on chip will be interesting if density significantly exceeds what can be done with SRAM, or if significantly cheaper, assuming they are similar energy. An SRAM delivers a byte of data for much less than 1pJ.
IBM Research’s new NorthPole AI chip
A new prototype chip from IBM Research could potentially bring energy-efficient AI to the edge.
IBM has used Samsung in the past. I will ask around. I will see them next week at DesignCon.
From this blog, IBM said 22 billion transistors in 800mm2 using 12nm nodes. The transistor density will be ~27.5Mt/mm2 for HPC application. I bet it more likely will be in GF.
256 cores with 768kB each, 192 MB. That compares poorly with the 900MB on a Graphcore Colossus 2 (which has 54G transistors in a similar size chip) which is pretty surprising considering IBM has access to embedded memory cells on 12nm. In AI model terms, this is indeed now small. Graphcore arguably bet too small and must work hard to figure out what comes next, and IBM appears to be placing a later, smaller bet.