Google has just unveiled its Jima 3. It will be able to get the same results near the Dispesic using just a GPU. Practical to run AI locally on a mobile.
“Who says better?” In the game of? The actor of the “tech”iaia The most dynamic right now. Earlier in the year, the Dipic AI shook American tanners in this sector with its IA as much as it is violent in the needs of the resources. Today, GoogleGoogle It seems that the answer is thinking that its latest open source model is close to JEMA 3 R1 DepressicDepressic Define only one part of its computing power.
According to Google, only one H100 NVIDIA GPU chip for JEMA 3 will be necessary that R1’s performance will be almost. Could reach According to Google, equal results will use 32 of these CPUs as the latter. In its blog note, the Internet takes the opportunity to highlight its TPU chip (processing unit processing unit. TenseTense ) It can be used on its model to achieve performance like NVIDIA GPU. It should be noted that the number of GPUs indicated in comparison by Google is based on its own estimates. For example, the Dippic AI does not use H100 GPU, but uses very little powerful H800 chips.
A smartphone model
But with JEMA 3, Google specifically tries to show that its model can be exploited directly by user terminals, rather than that By In terms of the number of data centers parameters, there are four different structures in the Jima 3 code, of which 1 billion, 4 billion, 12 billion and 27 billion. Compared to what is currently being done, these are very few parameters. In this way, DPCAC also manages 671 billion parameters. It is certainly possible to reduce this number by 37 billion by disabled parts of the network.
Google, Google, used AI techniques called “Anti -Anti -” AI technique to limit AI’s weight and make it more efficient. It needs to start with a very large model, so that it can remove its substance with some parameters to enhance the performance of AI. This is exactly what DiPsic has done Chat GPTChat GPT Designing its model. To move forward, it is also important to improve learning firmly. For this, the return of “man” is essential. The results of its results, if Jima 3 is less than the previous models, which is based on a small amount of parameters, which is based on the other “closed” gymnasium model. Thus, in the case of “token”, Jima 3 can accept 128,000 as compared to 8,000 for Jima 2. It can also prepare the text as well as deal with images. Who says better now?
