Due to concerns with heat and power consumption, Nvidia has not yet approved Samsung Electronics' most recent high bandwidth memory (HBM) chips for use in the American company's AI processors, according to three persons briefed on the matter.
According to them, the issues impact both the fifth-generation HBM3E chips that the South Korean tech giant and its competitors are releasing this year, as well as Samsung's HBM3 chips, which are the fourth-generation HBM standard that are currently most commonly used in graphics processing units (GPUs) for artificial intelligence.
For the first time, the reasons why Samsung failed Nvidia's testing are being revealed.
Samsung told Reuters in a statement that HBM is a customised memory product that needs "optimisation processes in tandem with customers' needs," and that it is currently working closely with customers to optimise its products. It refuses to comment on any particular clients.
Samsung replied in separate comments that "claims of failing due to heat and power consumption are not true" and that the testing was "proceeding smoothly and as planned" following the publication of this article by Reuters.
Nvidia chose not to respond.
Massive volumes of data generated by sophisticated AI applications are processed with the aid of HBM, a form of dynamic random access memory, or DRAM standard, which was first developed in 2013 and uses vertical stacking of chips to conserve space and lower power consumption.
The generative AI boom has increased demand for both HBM and advanced GPUs.
HBM makers regard satisfying Nvidia, which controls around 80% of the worldwide GPU market for AI applications, as essential to their future growth in terms of both profit momentum and reputation.
On Thursday, South Korea revealed a significant $19 billion package for its semiconductor industries.
According to the three sources, Samsung has been attempting to clear Nvidia's HBM3 and HBM3E tests since last year. Two of the individuals said that the findings of a recent test that went wrong for Samsung's 8- and 12-layer HBM3E chips arrived in April.
It's unclear right now if the issues are readily fixable, but according to the three individuals, Samsung's inability to fulfil Nvidia's criteria has raised investor and industry fears that it may lag behind competitors SK Hynix and Micron Technology in HBM.
The individuals, who were briefed on the subject by Samsung representatives, agreed to remain anonymous because the material was private.
In early Friday trading, Samsung's stock dropped 2%, somewhat underperforming the overall market.
In contrast to Samsung, Nvidia's primary HBM chip provider is the domestic competitor SK Hynix, which has been providing HBM3 since June 2022. Additionally, it started providing HBM3E to an unnamed client in late March. Sources claim that shipments ended up at Nvidia.
The only other significant HBM producer, Micron, has also said that it will provide HBM3E to Nvidia.
Samsung last week changed the head of its semiconductor unit, citing the need for a new leader to manage what it called an industry "crisis." Analysts said this decision seemed to highlight Samsung's fears about its lagging position in HBM.
As the largest memory chip manufacturer in the world, the market had high hopes that Samsung would breeze through Nvidia's testing; but, Jeff Kim, head of research at KB Securities, noted that it's also normal for specialist goods like HBM to take some time to fulfil customers' performance assessments.
Samsung supplies clients including Advanced Micro Devices (AMD) and plans to start mass producing HBM3E chips in the second quarter, even though it hasn't yet started supplying HBM3 to Nvidia. In a statement to Reuters, Samsung stated that everything is going according to plan with their product timeline.
Analysts also claim that SK Hynix, which created the first HBM chip in 2013, has a technological advantage over Samsung since it has invested a significant amount of time and money on HBM research and development over the previous ten years.
Samsung stated in a statement that it has been investing in HBM ever since it created the first commercial HBM solution for high-performance computing in 2015.
The sources said that GPU makers such as Nvidia and AMD are eager for Samsung to refine their HBM chips in order to increase their vendor alternatives and lessen SK Hynix's pricing clout.
In a sign at Samsung's stand during the Nvidia AI conference in March, Nvidia CEO Jensen Huang wrote "Jensen approved" for Samsung's 12-layer HBM3E, indicating the company's excitement about the possibility of Samsung providing such chips.
Trendforce, a research group, predicts that HBM3E chips will emerge as the industry standard HBM product this year, with shipments centred in the second half of 2024.
In addition, SK Hynix predicts that until 2027, the market for HBM memory chips as a whole would grow at a pace of 82% annually.
Investors have taken notice of Samsung's comparatively poor position in HBM, according to analysts. As of now in the year, its shares are down 2%, but the stock of SK Hynix and Micron has increased by 42% and 48%, respectively.
(Source:www.reuters.com)
According to them, the issues impact both the fifth-generation HBM3E chips that the South Korean tech giant and its competitors are releasing this year, as well as Samsung's HBM3 chips, which are the fourth-generation HBM standard that are currently most commonly used in graphics processing units (GPUs) for artificial intelligence.
For the first time, the reasons why Samsung failed Nvidia's testing are being revealed.
Samsung told Reuters in a statement that HBM is a customised memory product that needs "optimisation processes in tandem with customers' needs," and that it is currently working closely with customers to optimise its products. It refuses to comment on any particular clients.
Samsung replied in separate comments that "claims of failing due to heat and power consumption are not true" and that the testing was "proceeding smoothly and as planned" following the publication of this article by Reuters.
Nvidia chose not to respond.
Massive volumes of data generated by sophisticated AI applications are processed with the aid of HBM, a form of dynamic random access memory, or DRAM standard, which was first developed in 2013 and uses vertical stacking of chips to conserve space and lower power consumption.
The generative AI boom has increased demand for both HBM and advanced GPUs.
HBM makers regard satisfying Nvidia, which controls around 80% of the worldwide GPU market for AI applications, as essential to their future growth in terms of both profit momentum and reputation.
On Thursday, South Korea revealed a significant $19 billion package for its semiconductor industries.
According to the three sources, Samsung has been attempting to clear Nvidia's HBM3 and HBM3E tests since last year. Two of the individuals said that the findings of a recent test that went wrong for Samsung's 8- and 12-layer HBM3E chips arrived in April.
It's unclear right now if the issues are readily fixable, but according to the three individuals, Samsung's inability to fulfil Nvidia's criteria has raised investor and industry fears that it may lag behind competitors SK Hynix and Micron Technology in HBM.
The individuals, who were briefed on the subject by Samsung representatives, agreed to remain anonymous because the material was private.
In early Friday trading, Samsung's stock dropped 2%, somewhat underperforming the overall market.
In contrast to Samsung, Nvidia's primary HBM chip provider is the domestic competitor SK Hynix, which has been providing HBM3 since June 2022. Additionally, it started providing HBM3E to an unnamed client in late March. Sources claim that shipments ended up at Nvidia.
The only other significant HBM producer, Micron, has also said that it will provide HBM3E to Nvidia.
Samsung last week changed the head of its semiconductor unit, citing the need for a new leader to manage what it called an industry "crisis." Analysts said this decision seemed to highlight Samsung's fears about its lagging position in HBM.
As the largest memory chip manufacturer in the world, the market had high hopes that Samsung would breeze through Nvidia's testing; but, Jeff Kim, head of research at KB Securities, noted that it's also normal for specialist goods like HBM to take some time to fulfil customers' performance assessments.
Samsung supplies clients including Advanced Micro Devices (AMD) and plans to start mass producing HBM3E chips in the second quarter, even though it hasn't yet started supplying HBM3 to Nvidia. In a statement to Reuters, Samsung stated that everything is going according to plan with their product timeline.
Analysts also claim that SK Hynix, which created the first HBM chip in 2013, has a technological advantage over Samsung since it has invested a significant amount of time and money on HBM research and development over the previous ten years.
Samsung stated in a statement that it has been investing in HBM ever since it created the first commercial HBM solution for high-performance computing in 2015.
The sources said that GPU makers such as Nvidia and AMD are eager for Samsung to refine their HBM chips in order to increase their vendor alternatives and lessen SK Hynix's pricing clout.
In a sign at Samsung's stand during the Nvidia AI conference in March, Nvidia CEO Jensen Huang wrote "Jensen approved" for Samsung's 12-layer HBM3E, indicating the company's excitement about the possibility of Samsung providing such chips.
Trendforce, a research group, predicts that HBM3E chips will emerge as the industry standard HBM product this year, with shipments centred in the second half of 2024.
In addition, SK Hynix predicts that until 2027, the market for HBM memory chips as a whole would grow at a pace of 82% annually.
Investors have taken notice of Samsung's comparatively poor position in HBM, according to analysts. As of now in the year, its shares are down 2%, but the stock of SK Hynix and Micron has increased by 42% and 48%, respectively.
(Source:www.reuters.com)
Homepage
Send to a friend
Printable version
Share
