Gains In Generative AI With Dimensity 9400 Revealed In A New Leak

Gains In Generative AI With Dimensity 9400 Revealed In A New Leak

Models like Llama 2 and Stable Diffusion demonstrate the remarkable performance of the flagship-grade MediaTek Dimensity 9300 chipset for mobile devices. In comparison, the 9400, which is supposed to be its successor, is now said to be even more potent, producing text 50% faster and images up to roughly 17% faster by 2024.

Potentially even comparable to the upcoming Snapdragon 8 Gen 4, MediaTek’s new flagship SoC from the Dimensity series is now associated with “monstrous” processing and gaming performance. Additionally, the 9400 is currently being marketed as surpassing its predecessor in these areas by up to 18% while consuming 32% less energy (the second of which is reasonable given that the 9300 was introduced with promises of 33% generational efficiency gains in power). However, how good is its AI?

For the first time, the alleged benefits of the 9400 in this regard are allegedly described in a recent leak. This comprises a table that supplements other currently anticipated enhancements, specifically the transition to a 3nm architecture and the deliberate addition of a next-generation Cortex-X5 core (in lieu of one of the 9300’s quad X4s). Performance estimates are provided for one widely-used large language model (LLM) and one equally well-liked text-to-image generator, with comparisons to the Dimensity 9300 for the same tasks.

So, according to the table (which is purportedly based on estimates of official MediaTek press material), the 9400 can process 12 to 15 tokens (which are essentially the smallest textual unit for an LLM) per second using the 7-billion parameter Llama 2 model from Meta.

Likewise, if the 9300 requires 3.5 seconds to generate an image with Stable Diffusion 1.5, the 9400 is estimated to require 3 seconds. In light of this, it is now predicted that MediaTek’s APU 790 replacement will be 20–50% and up to 16.7% faster, respectively. In light of this, the OEM may promote it later in 2024 as an even greater “multi-modal generative AI” powerhouse.

error: Content is protected !!