Intellifusion, a Chinese manufacturer, is introducing "DeepEyes" AI boxes targeting the high-end hardware segment in the AI market. These AI boxes boast impressive AI performance of 48 TOPS, priced at 1000 yuan, which translates to approximately $140. This move comes in response to China-specific GPU bans imposed by the US, prompting Chinese companies like Intellifusion to innovate using older 14nm technology and likely ASIC designs to circumvent sanctions and compete effectively in the AI market.
The latest "DeepEyes" AI box for the first half of 2024 is equipped with a DeepEdge10Max SoC, delivering 48 TOPS in int8 training performance. The upcoming 2024 second-half model will feature a DeepEdge10Pro offering up to 24 TOPS. Looking ahead to the first half of 2025, Intellifusion plans to launch the DeepEdge10Ultra model, promising an impressive performance boost of up to 96 TOPS. While pricing details for these advanced models remain undisclosed, if Intellifusion can sustain the initial 1000 yuan pricing strategy, they may achieve their goal of providing significantly cheaper AI hardware covering a wide range of scenarios.
Powered by Intellifusion's custom NNP400T neural networking chip, these domestically-produced AI boxes feature a specialized NPU alongside other essential components like SoCs, including a 1.8 GHz 2+8 cores RISC CPU and a GPU capable of up to 800 MHz in the DeepEdge10 series.
To put this performance into perspective, to meet Microsoft's criteria for an "AI PC," modern computers must deliver at least 40 TOPS of NPU performance. Intellifusion appears on track to cater to various AI workloads, outpacing many existing NPUs that typically offer speeds of only 16 TOPS. However, Qualcomm's Snapdragon X Elite chips are expected to introduce 40 TOPS performance along with top-tier iGPU capabilities later this year, posing potential competition in the market.
Dr. Chen Ning, Intellifusion's chairman, predicts a significant increase in the global adoption of large AI models within the next three years, highlighting the substantial costs associated with training such models. While the claim that 80% of companies worldwide will implement AI may be viewed skeptically, the discussion around the financial barriers for businesses to leverage AI effectively remains valid. Intellifusion's DeepEdge chips, leveraging independent domestic technology and a RISC-V core, aim to support extensive model training and inference deployment efficiently.
