The old adage that everything is good in moderation threatens to be true in relation to the extremely rapid development of the artificial intelligence sector, McKinsey analysts warn. On the one hand, the market’s urgent need for specialized hardware and software is spurring investment and revitalizing the dynamics of almost the entire supply chain, from materials and raw materials for manufacturing processors to the creation of specialized servers on which increasingly “smart” AI models are trained. On the other hand, against the backdrop of a rather restrained macroeconomic forecast for the foreseeable future, there is no certainty that the necessary funds will be found in sufficient quantities and in the required time frame.
The main difference mexico whatsapp resource between generative AI and the algorithmic software with high system requirements that is familiar to the market is that insufficient hardware performance makes the use and especially the training of generative models essentially unprofitable. For example, if a model takes up 12 GB of video memory, it simply cannot be launched on a PC with a video card containing 8 GB of memory or less: it must fit entirely in RAM.
Theoretically, of course, it is possible to organize sequential loading and unloading of individual computation blocks into a smaller memory. However, generative AI is based on neural networks with tens and hundreds of billions of input parameters (in fact, the need to keep them all in memory at once determines its significant volume), and therefore the time spent on transferring huge data arrays between video RAM and other PC subsystems will be excessively large.
The execution speed of classical software based on algorithms is determined primarily by the ability of the central processor to perform fairly complex calculations in a small number of threads (usually single-threaded) in a limited time; memory volumes are secondary for such software. Generative AI, on the contrary, is based on extremely simple calculations in a huge number of parallel threads: the ability to place a titanic amount of data in memory with the highest possible access speed becomes critical in this case. If we are talking not about the execution of ready-made AI models, but about their training, the hardware requirements increase many times over.
What does this mean from the IT market point of view? Well, that the demand for both processors with as many simple cores as possible and for the memory with ultra-fast access connected to them has grown rapidly over the past year - and is unlikely to slow down in the coming years (unless, of course, generative AI itself as a concept disappoints the masses of customers for some reason). In other words, the market needs more high-performance video cards - after all, these computer components combine many simple computing nodes, and video memory with excellent performance, and fast data buses connecting them. Considering that the clouds that provide businesses and individuals with access to generative AI also use servers stuffed with video cards (much more powerful than consumer ones), the market is in dire need of new processors and video memory chips.