Google releases TurboQuant and shakes up the chip sector

The release of TurboQuantthe new compression algorithm per large language models presented by Google Researchpesa sul semiconductor sector because it aims to reduce one of the main ones infrastructure costs dell’generative artificial intelligence: the memory needed during inferencethat is, in the phase in which the model produces responses. TurboQuantwe read in the note from the Mountain View group, is a system designed to make theCheaper and faster AI.

The software intervenes on what can be described as the short-term memory of the model, i.e. the set of information that the system must temporarily retain to maintain the context and respond coherently. So far this process has taken advanced chipslarge amount of memory and a high energy consumption. According to Google, TurboQuant allows you to compress this information six times without compromising the accuracy of the results. This paves the way for the execution of complex models are less expensive infrastructure and allows you to manage longer texts with greater efficiency.

In industrial terms, the novelty indicates the possibility of obtaining the same performance with a much smaller amount of hardware resourceswith potential effects on operating costssui energy consumption and on scalability of artificial intelligence. The move of Google is immediately reflected on market and is read as a new signal of the push of the adults cloud providers towards a greater one vertical integrationwith the aim of contain costs and make model execution more efficient.

Reaction from the Asian market

Alone reaction it was heavy: the Kospi it closed down 3.35%, weighed down above all by the two main ones sector stocks. SK Hynixleader in memorie Hbmlost 6.23%, while Samsung Electronics it lost 4.71%. On the sector weighs heavily on the fear that thesoftware optimizationalong with the development of proprietary chips come Axioncan reduce the dependence on purely quantitative growth over time hardware capabilities. In other words, the market fears one slowdown in demand for advanced components by the adults technology groups.

Impact on the US market

The climate of caution it is also reflected on the US semiconductor sectorwhere in the premarket especially i titles more linked to memory and to production chain: Micron loses 3.39%, Sandisk il 3,49%, Western Digital l’1,70%, Lam Research il 2,27%, KLA 1.39% e Applied Materials 1.21%. However, it is moving against the trend Arm Holdingswhich gains 16.37%, supported by expectations on royalties related to the adoption of architecture Neoverse V2 on the part of Google.