Lin Yang, chief scientist with Gyrfalcon, told eeNews Europe in a telephone interview that the company could extend to analog or mixed-signal implementations of machine learning processors within a couple of years.
Gyrfalcon, founded in 2017, is already shipping its digital ASICs to customers and the 28 in the part number signifies the manufacturing process node used for the chip.
The 2802 will be a non-volatile memory version of the 2801/2803 made using an embedded magnetic RAM manufacturing process from foundry TSMC. “That is the most advanced MRAM technology in the world right now,” Yang said.
It is not clear whether Gyrfalcon has made use of the non-volatility of MRAM to augment functionality or the significant area saving offered by MRAM to increase the memory budget. The 2801 and 2803 chips are substantially the same but with some additional circuitry in the 2803 to support multi-chip arrays for use on PCIe cards for data centers.
Gyrfalcon is one of the early adopters of TSMC’s embedded MRAM.
TSMC has lagged rival foundries Globalfoundries and Samsung in supplying embedded MRAM but said last year that it would offer embedded MRAM as a non-volatile memory option for SoCs in 2018 (see Report: TSMC to offer embedded ReRAM in 2019).
Samsung and Globalfoundries have been offering eMRAM on 28nm CMOS and 22nm FDSOI processes, respectively, for some time but could be leapfrogged by TSMC which is planning to offer ReRAM non-volatile memory in 2019 and then move both embedded MRAM and embedded ReRAM to 22nm FinFET process.
Gyrfalcon also has ambitions to go beyond digital. “Analog implementations is an option for power and speed,” said Yang. He said it was most likely to happen off the back of a successful digital machine learning processor. “We could then re-optimize in analog but there would be less flexibility.” Digital provides the flexibility but analog could happen for some applications, even within two years, concluded Yang.
Related links and articles: