Another player has placed a bet in the AI chip track.
ICY Tech and Samsung's SEMIFIVE have completed the tapeout of an edge AI chip based on Samsung's 8nm eMRAM process. This is Asia's first commercial deployment of 8nm embedded MRAM.
The name is a mouthful, but the core logic is simple: the bottleneck in traditional AI chips isn't the compute units, it's the data transfer between memory and compute — the so-called "memory wall." ICY Tech's approach mixes MRAM (magnetoresistive RAM) and SRAM, keeping data closer to computation.
What Makes MRAM Different
MRAM isn't a new concept, but commercialization has been difficult. Its advantages are non-volatility, fast write speeds, and long endurance — theoretically ideal for AI inference scenarios. The problem has been manufacturing difficulty, low yield, and high cost.
ICY Tech completing tapeout on Samsung's 8nm process shows that, at least on the engineering side, the path to using eMRAM for edge AI chips is feasible. "Feasible" and "good" are still some distance apart, but the first step is taken.
The Edge AI Chip Landscape
The edge AI chip space is crowded right now. NVIDIA's Jetson series, Qualcomm's AI Hub, Huawei's Ascend 310B — big players each have their own solutions. ICY Tech, as a startup, chose a relatively niche technical route in MRAM — essentially differentiated competition: big players compete on compute scale, they compete on architectural efficiency.
Whether it can break out depends on two factors: the actual power and performance data of this chip, and whether customers are willing to pay for this architecture. Tapeout success is just the starting point. Mass production and customer adoption are the real tests.
One Observation Point
What interests me is the specific implementation of ICY Tech's "MRAM+SRAM hybrid memory architecture." Pure MRAM would be theoretically optimal, but cost and yield are issues. The hybrid architecture is a pragmatic compromise — SRAM covers high-frequency access areas, MRAM covers large-capacity storage areas.
If this design approach can be validated in actual products, it might offer a new reference direction for edge AI chip architecture design.
We'll see the real value of this chip when mass production data and first customer feedback come out.
Main sources:
- IT Bear report
- ICY Tech official information