By Heekyong Yang and Fanny Potkin
SEOUL/SINGAPORE (Reuters) – Samsung Electronics (KS:)’ newest excessive bandwidth reminiscence (HBM) chips have but to go Nvidia (NASDAQ:)’s checks to be used within the U.S. agency’s AI processors as a result of warmth and energy consumption issues, three individuals briefed on the problems mentioned.
The issues have an effect on Samsung’s HBM3 chips, that are the fourth-generation HBM normal at present most utilized in graphics processing models (GPUs) for synthetic intelligence, in addition to fifth-generation HBM3E chips that the South Korean tech big and its rivals are bringing to market this 12 months, they mentioned.
The explanations for Samsung failing Nvidia’s checks are being reported for the primary time.
Samsung mentioned in an announcement to Reuters that HBM is a custom-made reminiscence product requiring “optimization processes in tandem with prospects’ wants,” including that it’s within the technique of optimizing its merchandise by means of shut collaboration with prospects. It declined to touch upon particular prospects.
Nvidia declined to remark.
HBM – a kind of dynamic random entry reminiscence or DRAM normal first produced in 2013 through which chips are vertically stacked to save lots of area and scale back energy consumption – helps course of large quantities of information produced by advanced AI functions. As demand for classy GPUs has soared amid the generative AI increase, so has demand for HBM.
Satisfying Nvidia – which instructions round 80% of the worldwide GPU marketplace for AI functions – is seen as key to future development for HBM producers – each reputationally and when it comes to revenue momentum.
Samsung has been attempting to go Nvidia’s checks for HBM3 and HBM3E since final 12 months, the three sources mentioned. Based on two of the individuals, the outcomes of a current failed take a look at for Samsung’s 8-layer and 12-layer HBM3E chips got here in April.
It was not instantly clear if the issues may be simply addressed, however the three sources mentioned that the failures to fulfill Nvidia’s necessities have elevated considerations within the business and amongst buyers that Samsung might fall additional behind rivals SK Hynix and Micron Know-how (NASDAQ:) in HBM.
The sources, two of whom have been briefed on the matter by Samsung officers, spoke on situation of anonymity as the data was confidential.
NEW CHIP UNIT HEAD
In distinction to Samsung, home rival SK Hynix is the principle provider of HBM chips to Nvidia and has been supplying HBM3 since June 2022. It additionally started supplying HBM3E in late March to a buyer it declined to establish. Shipments went to Nvidia, sources have mentioned.
Micron, the one different main producer of HBM, has additionally mentioned it’s going to provide Nvidia with HBM3E.
In a transfer that analysts mentioned appeared to underscore Samsung’s considerations about its laggard place in HBM, Samsung this week changed the top of its semiconductor unit saying a brand new particular person on the prime was wanted to navigate what it referred to as a “disaster” affecting the business.
Market expectations have been excessive that Samsung, because the world’s largest maker of reminiscence chips, would rapidly go Nvidia’s checks, but it surely’s additionally pure that specialised merchandise like HBM will take a while to fulfill prospects’ efficiency evaluations, mentioned Jeff Kim, head of analysis at KB Securities.
Though Samsung has but to grow to be a provider of HBM3 to Nvidia, it does provide prospects equivalent to Superior Micro Gadgets (NASDAQ:) and is aiming to start mass manufacturing of HBM3E chips within the second quarter. Samsung mentioned in its assertion to Reuters that its product schedule is continuing as deliberate.
Analysts additionally say that SK Hynix, which developed the primary HBM chip in 2013, has spent much more time and assets on HBM analysis and growth than Samsung over the previous decade, accounting for its technological edge.
Samsung mentioned in its assertion that it developed the primary business HBM answer for high-performance computing in 2015 and has continued to spend money on HBM since then.
GPU producers like Nvidia and AMD are eager for Samsung to good its HBM chips to allow them to have extra vendor choices and weaken SK Hynix’s pricing energy, the sources additionally mentioned.
At an Nvidia AI convention in March, Nvidia CEO Jensen Huang signed a placard at Samsung’s sales space that mentioned “Jensen accepted” for Samsung’s 12-layer HBM3E – underscoring the corporate’s enthusiasm in regards to the prospect of Samsung supplying these chips.
HBM3E chips are more likely to grow to be the mainstream HBM product available in the market this 12 months with shipments concentrated within the second half of 2024, in response to analysis agency Trendforce.
Individually, SK Hynix estimates that demand for HBM reminiscence chips usually might enhance at an annual charge of 82% by means of 2027.
Analysts have mentioned that Samsung’s comparatively weak place in HBM has been observed by buyers. Its shares are flat for the 12 months so far, whereas SK Hynix’s inventory is up 41% and Micron’s shares are up 48%.
($1 = 1,362.8600 gained)
