Exclusive-Samsung’s HBM chips failing Nvidia tests due to heat and power consumption woes, sources say

SEOUL/SINGAPORE : Samsung Electronics’ most modern excessive bandwidth reminiscence (HBM) chips occupy but to pass Nvidia’s exams to be used within the U.S. firm’s AI processors due to heat and vitality consumption considerations, three folks briefed on the disorders stated.

The considerations occupy an affect on Samsung’s HBM3 chips, which can presumably well be the fourth-generation HBM long-established currently most feeble in graphics processing items (GPUs) for synthetic intelligence, moreover fifth-generation HBM3E chips that the South Korean tech large and its competitors are bringing to market this year, they stated.

The reasons for Samsung failing Nvidia’s exams are being reported for the first time.

Samsung stated in a assertion to Reuters that HBM is a personalised reminiscence product requiring “optimization processes in tandem with customers’ needs,” adding that it is within the midst of of optimizing its products thru close collaboration with customers. It declined to observation on explicit customers.

Nvidia declined to observation.

HBM – a compose of dynamic random fetch admission to reminiscence or DRAM long-established first produced in 2013 thru which chips are vertically stacked to connect condo and within the reduction of vitality consumption – helps route of big quantities of files produced by complex AI applications. As quiz of for cultured GPUs has soared amid the generative AI growth, so has quiz of for HBM.

Pleasurable Nvidia – which commands round 80 per cent of the global GPU marketplace for AI applications – is viewed as key to future growth for HBM manufacturers – every reputationally and by components of profit momentum.

Samsung has been seeking to pass Nvidia’s exams for HBM3 and HBM3E since final year, the three sources stated. Per two of the folks, the consequences of a most modern failed take a look at for Samsung’s 8-layer and 12-layer HBM3E chips got here in April.

It became no longer without extend determined if the considerations can even fair also be with out say addressed, however the three sources stated that the failures to meet Nvidia’s requirements occupy elevated concerns within the alternate and amongst investors that Samsung can even fall further slack competitors SK Hynix and Micron Abilities in HBM.

The sources, two of whom were briefed on the topic by Samsung officers, spoke on situation of anonymity because the records became confidential.

NEW CHIP UNIT HEAD

Unlike Samsung, domestic rival SK Hynix is the principle vendor of HBM chips to Nvidia and has been supplying HBM3 since June 2022. It also started supplying HBM3E in stupid March to a buyer it declined to name. Shipments went to Nvidia, sources occupy stated.

Micron, the supreme other principal manufacturer of HBM, has also stated this can even fair provide Nvidia with HBM3E.

In a transfer that analysts stated perceived to underscore Samsung’s concerns about its laggard position in HBM, Samsung this week replaced the head of its semiconductor unit announcing a brand recent particular person at the cease became mandatory to navigate what it called a “disaster” affecting the alternate.

Market expectations were excessive that Samsung, because the realm’s supreme maker of reminiscence chips, would mercurial pass Nvidia’s exams, but or no longer it is on the complete natural that specialised products worship HBM will rob a whereas to meet customers’ efficiency opinions, stated Jeff Kim, head of compare at KB Securities.

Even though Samsung has but to turn correct into a vendor of HBM3 to Nvidia, it does provide customers corresponding to Advanced Micro Units (AMD) and is aiming to initiate up mass manufacturing of HBM3E chips within the 2nd quarter. Samsung stated in its assertion to Reuters that its product agenda is persevering with as deliberate.

Analysts also tell that SK Hynix, which developed the first HBM chip in 2013, has spent method more time and sources on HBM compare and pattern than Samsung over the last decade, accounting for its technological edge.

Samsung stated in its assertion that it developed the first business HBM resolution for excessive-efficiency computing in 2015 and has persevered to make investments in HBM since then.

GPU manufacturers worship Nvidia and AMD are alive to for Samsung to ideal its HBM chips so they’ll occupy more vendor choices and weaken SK Hynix’s pricing vitality, the sources also stated.

At an Nvidia AI convention in March, Nvidia CEO Jensen Huang signed a placard at Samsung’s booth that stated “Jensen popular” for Samsung’s 12-layer HBM3E – underscoring the corporate’s enthusiasm in regards to the likelihood of Samsung supplying those chips.

HBM3E chips are at risk of became the mainstream HBM product available within the market this year with shipments concentrated within the 2nd half of 2024, in step with compare firm Trendforce.

Individually, SK Hynix estimates that quiz of for HBM reminiscence chips on the complete can even enhance at an annual fee of 82 per cent thru 2027.

Analysts occupy stated that Samsung’s moderately former position in HBM has been noticed by investors. Its shares are flat for the year to this level, whereas SK Hynix’s stock is up 41 per cent and Micron’s shares are up forty eight per cent.

($1 = 1,362.8600 won)