Hadlee Simons / Android Authority
TL;DR
- MediaTek has announced that its next-generation flagship chipset will better support on-device generative AI.
- The company is working with Meta to offer the Llama 2 large language model for on-device AI capabilities.
- It adds that the first phones with this chip will be available by the end of the year.
The Taiwanese chipmaker revealed that it’s working with Meta to better support the Llama 2 LLM on phones powered by the company’s next-generation flagship processor, which will presumably be called the Dimensity 9300. Llama 2 is an alternative to OpenAI’s GPT and Google’s PaLM 2 models.
The company outlined a few key features that will enable on-device generative AI capabilities in 2024’s phones:
MediaTek’s next-generation flagship chipset, to be introduced later this year, will feature a software stack optimized to run Llama 2, as well as an upgraded APU with Transformer backbone acceleration, reduced footprint access and use of DRAM bandwidth, further enhancing LLM and AIGC performance.
MediaTek says it expects Llama 2-enabled generative AI apps to be available for phones powered by its next-generation chipset. It adds that these phones are scheduled to hit the market by the end of 2023.
Either way, the fact that both MediaTek and Qualcomm have announced better support for on-device generative AI is a good thing. Being able to run generative AI tasks locally has privacy benefits, as your data isn’t sent to the cloud for inference. Local inference is also a boon if you’re roaming or in a part of the world with costly data fees. So we’re looking forward to seeing how 2024’s high-end smartphones will take advantage of this capability.
We’re also keen to see the Dimensity 9300 chipset in general as it could deliver a radical design. The company already confirmed that it will use Arm’s Cortex-X4 and A720 CPU cores, as well as an Immortalis-G720 GPU. But a leaker has claimed that this chipset will offer no little cores, using four big Cortex-X4 cores and four medium A720 cores instead.
