The demand for high bandwidth memory (HBM) chips has boomed with the growing popularity of generative AI. But Samsung, unlike peers SK Hynix and Micron Technology, has been conspicuous by its absence in any deal making with AI chip leader Nvidia to supply latest HBM chips.
One of the reasons Samsung has fallen behind is its decision to stick with chip making technology called non-conductive film (NCF) that causes some production issues, while Hynix switched to the mass reflow moulded underfill (MR-MUF) method to address NCF’s weakness, according to analysts and industry watchers.
Samsung, however, has recently issued purchase orders for chip-making equipment designed to handle MUF technique, three sources with direct knowledge of the matter said.
02:18
From K-pop to salesgirls: AI goes mainstream in South Korea
From K-pop to salesgirls: AI goes mainstream in South Korea
“Samsung had to do something to ramp up its HBM (production) yields … adopting MUF technique is a little bit of swallow-your-pride type thing for Samsung, because it ended up following the technique first used by SK Hynix,” one of the sources said.
Samsung’s HBM3 chip production yields stand at about 10-20 per cent while SK Hynix has secured about 60-70 per cent yield rates for its HBM3 production, according to several analysts.
The HBM3 and HBM3E, the newest versions of HBM chips, are in hot demand. They are bundled with core microprocessor chips to help process vast amounts of data in generative AI.
Samsung is also in talks with material manufacturers, including Japan’s Nagase, to source MUF materials, one source said, adding mass production of the high-end chips using MUF is unlikely to be ready until next year at the earliest, as Samsung needs to run more tests.
The three sources also said Samsung plans to use both NCF and MUF techniques for its latest HBM chip.
Samsung said its internally developed NCF technology is an “optimal solution” for HBM products and would be used in its new HBM3E chips. “We are carrying out our HBM3E product business as planned,” Samsung said in a statement.
Nvidia and Nagase declined to comment.
All sources spoke on condition of anonymity as the information is not public.
Samsung’s plan to use MUF underscores growing pressure it faces in the AI chip race, with the HBM chip market, according to research firm TrendForce, seen more than doubling this year to nearly US$9 billion amid AI-related demand.
NCF VERSUS MUF
The non-conductive film chip manufacturing technology has been widely used by chip makers to stack multiple layers of chips in a compact high bandwidth memory chipset, as using thermally compressed thin film helps minimise space between stacked chips.
But there are often problems linked to adhesive materials as manufacturing gets complicated as more layers are added. Samsung says its latest HBM3E chip has 12 chip layers. Chip makers have been looking for alternatives to address such weaknesses.
SK Hynix successfully switched to the mass reflow moulded underfill technique ahead of others, becoming the first vendor to supply HBM3 chips to Nvidia.
SK Hynix’s market share in HBM3 and more advanced HBM products for Nvidia is estimated at above 80 per cent this year, according to Jeff Kim, an analyst at KB Securities.
Huawei’s AI chip prowess under scrutiny after Nvidia taps it as likely rival
Huawei’s AI chip prowess under scrutiny after Nvidia taps it as likely rival
Micron joined the high bandwidth memory chip race last month, announcing that its latest HBM3E chip will be adopted by Nvidia to power the latter’s H200 Tensor chips which will begin shipping in the second quarter.
Samsung’s HBM3 series have not yet passed Nvidia’s qualification for supply deals, according to one of the four sources and another person with knowledge of the discussion.
Its setback in the AI chip race has also been noticed by investors, with its shares falling 7 per cent this year, lagging SK Hynix and Micron which are up 17 per cent and 14 per cent, respectively.