A meta announces the second generation of the MTIA (Meta-training inference accelerator), 5-nanometer chip and supports LPDDR5 memory. The novelty is focused on artificial intelligence (AI) and it is clear that Facebook's turn has the idea of venturing with appropriate hardware. Owner and owner.
Like the first generation MTIA, Meta's new chip is an AI inference accelerator. The idea of inferring an operation performed by a system based on neural networks is a prediction that can be used for various actions, such as multiple decisions.
![Module for implementing the Meta chip (image: disclosure/Meta)](https://files.tecnoblog.net/wp-content/uploads/2024/04/meta-mtia-modulo-700x500.jpg)
![Module for implementing the Meta chip (image: disclosure/Meta)](https://files.tecnoblog.net/wp-content/uploads/2024/04/meta-mtia-modulo-700x500.jpg)
The new MTIA was designed to perform the most efficient tasks. A meta explains that, for this, the chip was projected to be the most powerful of that of the ancestor. It was possible to do this in a group of travelers, among others:
- supports up to 128 GB of LPDDR5 memory, or related to the first generation;
- memory stripe length of 204.8 GB/s, compared to 176 GB/s for the previous generation;
- PCI Express 5.0 controller with a stripe length of 64 GB/s, compared to a PCI Express 4.0 controller with a stripe length of 32 GB/s with the first MTIA;
- TSMC manufacturing process of 5 nanometers versus 7 nanometers without a chip antecedent.
Changing the manufacturing technology makes it possible to increase the frequency of the new MTIA from 800 MHz to 1.35 GHz (up to 68.8%). The chip is also larger, going from 373mm² to 421mm².
The chips then follow a basic structure consisting of an 8 × 8 matrix composed of “processing elements” (PE). Each PE follows the RISC-V architecture and provides a core for growing workloads and other military expansions.
However, the advances of the new generation have increased or the TDP goes from 25 W to 90 W. This is the price that will be paid for a very expressive evolution between generations. The priority is everyone's. This is why, in its current state, the company has been putting together cabinets with 24 second-generation MTIA chips for you every day.
![Second generation MTIA chips are already in use (image: disclosure/Meta)](https://files.tecnoblog.net/wp-content/uploads/2024/04/meta-mtia-servidor-1060x522.jpg)
![Second generation MTIA chips are already in use (image: disclosure/Meta)](https://files.tecnoblog.net/wp-content/uploads/2024/04/meta-mtia-servidor-1060x522.jpg)
In the first place, because Meta failed to attract the attention of Google and Microsoft to this artificial intelligence. For this reason, the company is investing heavily in giant chips such as Nvidia and AMD to focus on their growing AI workloads. A Meta is one of the main corporate clients.
But, with feitos sob medida chips, a Meta can implement a infrastructure for AI but condizente com as suas requested. The logic is to adapt the chip to your applications, and not on the contrary, this is what tends to find third party solutions. A clean meta-contextualization:
The MTIA is a long-term tool for creating a more efficient architecture for Meta's proprietary operations. (…) We have sold positive results to this program, allowing us to dedicate and invest in greater computing power for our most intensive AI operations.
A meta-recognition that the MTIA must also help a company with lidar with generative AI. What ultimately felt was that this scene was dominated by ChatGPT and Google Gemini ferramentas, with almost zero meta-participation.