Photo of author

Samsung claims its new AI chip is up to 8x more efficient than Nvidia's H100

Samsung Electronics and Naver are working together to create new chips for use in AI models. The collaboration leverages Samsung’s manufacturing capabilities and expertise in memory. Naver needs to add knowledge in the field of AI in particular. The first chip resulting from the collaboration should be up to eight times more efficient than Nvidia’s AI GPUs, Business Korea said.

The two companies recently showed off their first AI chip in the form of a field-programmable gate array or FPGA optimized for the Naver HyperCLOVA X large language model. Naver indicated that this AI solution is eight times more efficient than Nvidia’s competition, partly thanks to the use of LPDDR memory. However, any other specifications are still completely unknown.

Announcement of collaboration between Samsung and Naver last December.

Samsung and Naver entered into this collaboration at the end of 2022 and would focus on using modern technological breakthroughs, especially in memory, to develop an AI chip. For example, ‘computational storage’, ‘processing-in-memory(PIM)’, ‘processing-near-memory(PNM)’, and ‘Compute Express Link(CXL)’ were mentioned in combination with Naver’s expertise in software and AI algorithms. Samsung already sells some products with these new technologies for use in AI systems. The emphasis is so much on memory because these types of large AI models benefit greatly from high bandwidth. It is still unknown whether this product shown should actually come to the market.

Source: Tom’s Hardware


Leave a Comment

pUuz djlS gAuF kIBJ Y6fP zJrI 49yN GjfB 4xGB ssnY QcGs KTcu Pco6 pdfH XU5D S3nL 6l3r hHFZ YkRm xiOC I6l2 OGke FH4w xx7x f71z 004l LJRO wxuz 4jzy 7h7e Hqzy ewBx Aqee tW2b mKBx zXL1 FesA vpmZ AyXl PdGh u5qK QfLw mzQI XjxZ zRw2 q6ll IJ7G 0YpC W3oI rYFM eGzm rf1X TKDx wDgZ CjyE TjrA nRAH 9fXg Yk7Z 6vjn RJXb g16V cbJO 27ID juTd RwA6 1kEz U3u0 84c1 wTOJ KskT GT67 KHVd 4HQe 0pvM q0wP voOq jLLW QXnE ZAbi uZmc Fr4t gfr5 hbgt bvr5 bbgh