To understand the interest of this new project, let us remember that it is an architecture called “ Grace Blackwell », embedding up to 20 CPU cores based on ARM architecture and a state-of-the-art in-house GPU equipped with fifth generation Tensor cores. That is, the system supports large AI models, potentially up to 200 billion parameters on a single unit. This is a giant step for those who wish to develop or adjust LLM (Large Language Model) type language models without having to rent server farms.
The result? A box that resembles a mini-computer, still pluggable into a standard socket, without requiring a monstrous cooling system. During its official presentations, NVIDIA insisted that up to 1 petaflops of power is achievable with this configuration. This does not prevent the platform from upgrading for more demanding calculations, thus offering a wide range of use cases.
If NVIDIA already masters the field of GPUs and accelerated computing, MediaTek brings here its know-how in the design of ARM chips with reduced power consumption. We are talking about 20 ARM cores for the CPU part, all connected to the graphics chip via a very high speed interface. Result: significantly reduced communication latency and controlled consumption on a single SoC (System-on-Chip). This unified approach promotes memory consistency and avoids having to multiply storage modules, since each workstation can carry up to 128 GB of memory and several terabytes of NVMe storage.
