D-RAM Market’s Gattteok Award with Chat GPT

D-RAM Market's Gattteok Award with Chat GPT

image text translation

(1)IBK Industrial Bank of Korea Financial Group
(2)IBK Investment & Securities Co., Ltd
(3)April 3rd, 2023
(4)Kim Woon-ho of IT Semiconductor Display
(5)AI will be the medium- to long-term growth engine for the DRAM market
(6)Demand for DRAM continues to be sluggish, but AI-related DRAM demand has a positive impact on the DRAM market. Omdia believes that this market change will not be a temporary event, and that it needs to expand the supply of coprocessor servers, including accelerators such as GPGPU
(7)HBM’s annual growth rate of more than 40 is expected to fall to the 30s due to macroeconomic factors, but demand for HBM is rising due to interest from ChatGPT in late 22 and early 23. HBM is expected to grow more than 50 DRAM products this year
(8)Demand for HBM will increase rapidly
(9)OMDIA SAYS DEMAND FOR COPROCESSOR SERVERS HAS RECORDED 10 OF THE OVERALL SERVER MARKET IN 2022 BUT LATERAL CHANGES ARE EXPECTED TO DOUBLE THE COPROCESSOR’S PORTION IN 25-26 YEARS TO GROW TO THE PORTION OF 20-30
(10)HBM demand is also forecast to account for 2-3 of the total server DRAM market in 2023, but is expected to more than double between 25 and 26 Total system memory demand, including high-density RDIMMs and HBMs, accounted for 10 percent of server DRAM this year, but is expected to increase to about 30 in 25-26
(11)More importantly, the industry’s efforts to overcome the memory wall, which was considered a limitation of computing power, are continuing due to the need for much more computing power when expanded to images and videos, and the core of CXL is memory expansion and memory pooling

D-RAM Market's Gattteok Award with Chat GPT

image text translation

(1)AI development is a new driver of DRAM market growth
(2)AI’s significant change is not an issue that ends with discussing the short-term supply and demand of the DRAM industry. AI’s development will change the IT market in the future and this will be a medium- to long-term growth momentum for the DRAM market
(3)There are many concerns surrounding the DRAM market. The weakening demand due to macroeconomic variables and the unprecedented high inventory of DRAM manufacturers can’t erase doubts about the market’s growth potential Judgment
(4)Data OMDIA IBK Investment & Securities Co., Ltd

The current idea of Chat GPT in the company’s marketing department is that Chat GPT is an increase in demand for semiconductors from customers, including ultra-high performance memory, but it is not yet clear whether this is temporary or continuous demand That’s how much it boils down to, because it’s a demand we haven’t experienced yet, so of course the forecast is conservative, but I think it feels like a deja vu that I’ve seen before, just before the 17-18 memory supercycle

Up until 16 years ago, almost half of the memory demand was for mobile and 30-valence PCs, and the demand for memory for servers that had never been considered an important market since late 2016 suddenly soared. It wasn’t Samjeon, but Amazon, Google, Microsoft, and big tech companies It’s impossible

There was a Google Deep Mind Challenge match between AlphaGo and Lee Sedol at the beginning of 2016, and we were all shocked, but we didn’t think it was the starting point of the real server memory demand explosion. It’s hard to know at the time. It’s still a weak brain

From this point of view, Omdia’s Memory Semiconductor Report, which came out this time, is very confident about the explosive demand growth of memory semiconductors, especially DRAM, led by AI such as Chat GPT, so I introduced the summary of the report as an attachment

The 17-18 memory supercycle is driven by NAND demand at inflection point where SSDs replace HDDs, so replacing storage from HDDs increases computing speed incomparably. At that time, storage was the most serious bottleneck of the computing structure phone Neumann structure

So if the channel bus bandwidth with the AI accelerator is the speed limit, the channel bus must significantly increase the number of lanes to resolve this bottleneck. PIM also removes some of the computational functions and moves them to memory. I’m not sure if the purpose of CXL HBM PIM is the same

Also, I think that HBM demand in the real market is much stronger than 3 levels of DRAM for servers this year, especially Hynix, which currently controls almost 7-80 HBM market, will have a much higher share of HBM sales than ODDIA predicted

Please rate the post

0 0

Your page rank:

Leave a Comment