PHison democratizes the training of major language models

With a piece of his own middleware as a secret ingredient. For example, the PHison controller manufacturer wants to gain efficiency in the use of GPUs. However, Result: ‘The training of an AI is no longer reserved for a happy few.’

An increasing number of companies and institutions seek to develop artificial intelligence within their own infrastructure. Therefore, Especially in Europe. Meanwhile, under the combined effect of the need to keep sensitive data in the walls of the company and the desire to preserve its sovereignty. For example, But the reality is different. However, Think of the prohibitive costs of GPUs, whose availability is also limited. Consequently, With Aidaptiv+, Phison, a Taiwanese manufacturer in Hardware specializing in controllers, offers a pragmatic solution. Moreover, By deploying economic flash storage as an extension of the costly VRAM. Furthermore, it puts the phison democratizes training major language training of large language models within the reach of a much wider audience. Consequently, ‘The training of an AI is no longer reserved for a Happy Few’. However, said Michael Wu, General Manager and President of Phison Technology Inc. However, (USA), during a visit to the IT Press Tour in California in which Data News participated.

‘Crossing the wall of AI. Therefore, memory’ – Phison democratizes training major language

Innovations that are hidden in a small corner can sometimes be completely disruptive. Nevertheless, And Phison hopes to provide new proof. Meanwhile, In this case. crossing the wall of AI and memory, as explained Brian Cox (Phison): ‘In terms of parameters, the complexity and size of AI models increase a 410 factor every two years. The capacity of HBM (High Bandwidth Memory). other phison democratizes training major language GDDR Memoirs (Graphics Double Data Rate, an ultra -performing, editor’s note), two extremely fast memories, but very expensive by professional GPU, has only progressed by one factor two over the same period. ‘

This situation often forces companies to use costly rescue solutions. ‘I spoke to customers who had to buy several GPU cards not because they needed additional calculation power. but only for the memory capacity that these cards conceal,’ continues Brian Cox. ‘A large part of the GPUs remain under-exploited.’

Developed by necessity – Phison democratizes training major language

This challenge will become all the more pressing as IA applications on site will catch up. then exceed public cloud models. This is at least the Phison hypothesis, which is based on a Market Research dimension analysis, a market research company.

But the real reason why Phison developed a solution? Obviously, necessity has law, again. ‘Our CEO. KS Pua wanted to call on phison democratizes training major language AI to improve our own design and manufacturing processes, but the solutions offered to him cost several million dollars’, explains Brian Cox. ‘The response of our CEO was clear:’ I don’t have the money. Find a more effective way. ‘ And that’s what we have done, “smiles Brian Cox.

The SSD in life buoy

This more effective method therefore responds to the name of Aidaptiv+. The concept is as simple as it is effective: instead of loading a full AI model in the limited. expensive GPU true, the model is placed on a much larger flash storage module et cheaper. The secret ingredient is the AidapTivlink Middleware, a piece of software that acts as an intelligent traffic regulator. This software transfers the ‘slices’ of the necessary model from the SSD to the GPU at the precise moment when it needs it for the treatment. makes the GPU pipeline phison democratizes training major language remains constantly powered.

The hardware component, called aidaptivcache, is not a standard SSD. As the training of AI models is a process with a high writing intensity. Phison has developed a specialized disc capable of supporting up to 100 ‘complete entries per day’ (Drive Writes Per Day, DWPD). Above all, remember that this figure is much higher than the most robust corporate SSD disks.

Time and money

Phison is however realistic: ‘It is a technological compromise. And sorry, we gave priority to the costs’ recognized Brian Cox during the presentation. The training of a model takes more time than when you have an unlimited budget for a farm of. servers equipped with the most expensive GPUs. An illustration of the Phison laboratory shows that phison democratizes training major language the training of a model at 70 billion parameters at a four GPU workstation without Aidaptiv+ fails purely. simply for lack of memory. With AidapTiv+, it takes a little more than four hours. ‘But you can press on start And go for lunch: everything will be ready when you return ‘. says Brian Cox:’ For a large number of medium-sized companies, universities or R&D departments which should not re-use certain models every hour-but perhaps every day or every week-, it is compromised more than acceptable. The reduction in the promised costs of a factor 8 to 10 is then decisive ‘, he concludes.