Chinese firm Intellifusion is set to launch “DeepEyes” AI containers boasting an alleged AI performance of 48 TOPs for 1000 yuan, approximately $140. This move is targeted towards challenging the dominance of high-end hardware in the AI industry, which has led to China-specific GPU restrictions imposed by the US. To circumvent potential repercussions and uphold its position in the AI sector, China may opt for an older 14-n network and likely an ASIC.
The initial release, scheduled for 2024, will feature the “Strong Eyes” AI package integrating a DeepEdge10Max SoC delivering 48 TOPs in int8 teaching efficiency. Subsequent launches include the 2024 H2 Deep Eyes field with a potential DeepEdge10Pro offering up to 24 TOPS, and the 2025 H1 Deep Eyes field targeting a significant performance boost with the DeepEdge10Ultra achieving a score of up to 96 TOPS. While pricing details for these advanced models remain undisclosed, if Intellifusion can sustain the initial ~1000 yuan price point in the long run, they could realize their vision of providing “90% cheaper AI hardware” covering a wide range of scenarios.
All the aforementioned domestically-produced hardware capitalizes on Intellifusion’s proprietary NNP400T neural networking chip. This specialized chip features a 1.8 GHz 2+8 cores RISC CPU, complemented by a GPU of up to 800 MHz in DeepEdge 10, alongside other anticipated SoC components, making it a compelling choice in the market.
)
To comply with Microsoft’s criteria for an “AI PC,” contemporary PCs must deliver a minimum of 40 TOPs of NPU performance. Intellifusion’s rapid progress suggests that their offerings will soon cater to a broad spectrum of AI workloads, especially considering that most existing NPUs are limited to 16 TOPs. In contrast, Snapdragon’s X Elite chips are poised to introduce 40 TOPS alongside cutting-edge iGPU performance later this year.
Dr. Chen Ning, Intellifusion’s chairman, predicts, “In the next three years, 80% of companies globally will adopt large models.” While the assertion about widespread AI adoption may raise doubts, the statement underscores the substantial costs associated with leveraging AI effectively, particularly in model development. To facilitate extensive model training and deployment, the DeepEdge chips leverage “independent and controllable domestic technology” along with a RISC-V core.
