IBM Telum, a chip with integrated AI against fraud

At the conference Hot chips 33 IBM unveiled its new processor, Telum. This is a processor made with a 7 nanometer production process that allows you to speed up the calculations of artificial intelligence related to inference. A product intended for companies operating in sectors such as banking, finance or insurance, which enables real-time analysis of transactions in search of attempted fraud.

Telum, IBM’s solution to detect fraud in real time


Artificial intelligence has been used for some time in many industries to prevent and detect attempted fraud in banking, financial and online transactions in general. However, the systems currently in use are not very efficient, so that in 2020, a good 3.3 billion dollars “evaporated” because of fraud (for example by making purchases with stolen credit cards), depending on the Federal Trade Commission Consumer Sentinel 2020 Network Data Book. The cause, according to IBM, is to be found in the high latencies of the current solutions adopted to fight against this type of scam, which often prevent the analysis of transactions in real time and end up not discovering the fraud until after it has been analyzed. has been carried out. . In short, they are very effective AI systems for detecting fraud, but not for preventing them, which will on the contrary be possible thanks to Telum..

IBM Telum, 8 cores at 5 GHz to prevent fraud and accelerate risk analysis

Unlike the complex artificial intelligence systems currently adopted, Telum – an 8-core processor clocked at 5 GHz and equipped with 32 MB of L2 cache – integrates an inference computation accelerator directly on the chip, thus reducing latencies for this type. calculations, widely used in the financial industry. This will greatly speed up the relative calculations risk analysis of credit card transactions, loan processing, anti-money laundering analyzes, since the data will be processed directly by the chip and not, as generally happens by separate systems, often geographically distant, which cannot guarantee the same answer times.