SK Hynix Inc., the world’s No. 2 memory chipmaker, began mass-producing HBM3E, the best-performing DRAM chip for AI applications, for the first time in the industry to supply Nvidia Corp., the leading global semiconductor designer.
SK Hynix said on Tuesday it plans to begin delivery of HBM3E, the extended version of HBM3, or high bandwidth memory 3, later this month. The chip is expected to be used with Nvidia’s graphics processing unit set to launch in the second quarter.
HBM3E is a fifth-generation DRAM memory, succeeding the previous generations — HBM, HBM2, HBM2E and HMB3. The South Korean tech giant, which pioneered mass-production of HBM3, announced its successful development of HBM3E in August 2023.
“With the success story of the HBM business and the strong partnership with customers that it has built for years, SK Hynix will cement its position as the complete AI memory provider,” said Ryu Sungsoo, head of SK Hynix’s HBM business, in a statement.
HBM is high-value, high-performance memory that vertically interconnects multiple DRAM chips, dramatically increasing data processing speed compared with earlier DRAM products.
The company, which dominates the global HBM market with a 90% share, was known to be the industry's first to mass-produce HBM3E, sources said, although US-based Micron Technology announced last month it had started production of its latest chip for Nvidia’s H200.
FAST-GROWING HBM MARKET
The global HBM market has been growing rapidly thanks to the robust growth of AI technology.
HBM sales were forecast to account for 20.1% of total global DRAM revenue this year, up sharply from 8.4% in 2023 and 2.6% in 2022, according to industry tracker TrendForce. Global DRAM revenue was expected to soar 62.3% to $84.2 billion in 2024 from the previous year.
The global DRAM industry was expected to allocate about 14% of its total capacity to producing HBM using Through Silicon Vias (TSV) technology with an estimated annual supply growth of 260%, TrendForce said in a note on Monday.
SK Hynix and its larger rival Samsung Electronics Co., which each had an HBM TSV capacity of 45,000 units per month in 2023, have aggressive HBM production plans for the rest of this year, TrendForce said.
Samsung’s total HBM capacity was expected to nearly triple to 130,000 units per month by year's end, while SK Hynix’s capacity was forecast to more than double to 120,000-125,000 units, the research firm said.
TO PROCESS 230 FULL-HD MOVIES IN ONE SECOND
SK Hynix said its latest HBM3E processes up to 1.18 terabytes of data per second — the equivalent to processing more than 230 full-HD movies of 5 gigabytes each in one second.
The chipmaker said the product comes with 10% improved heat dissipation compared with previous models by adopting advanced mass reflow molded underfill (MR-MUF) technology. It is vital to control heat in AI memories given its extremely high-speed operations.
The MR-MUF technology is critical to securing stable mass production on the supply side of the HBM ecosystem as it can reduce pressure on stacked chips while better controlling warping.
Separately, SK Hynix said it is on track to commercialize 12-layer HBM3E to meet customers’ timelines.
“We aim to make the 12-layer product as slim as eight-layer models following JEDEC standards,” said a company official, referring to the JEDEC Solid State Technology Association, an independent semiconductor engineering trade and standardization body.
Last month, SK Hynix unveiled the industry’s first 16-layer HBM3E technology at the International Solid-State Circuits Conference in San Francisco.
