news

according to the information on september 13, karan batta, senior vice president of software giant oracle cloud infrastructure, said that although nvidia is booming in selling artificial intelligence data center chips, some customers developing ai are turning to smaller rival amd.

oracle recognizes nvidia's dominant position in the chip market, but also believes that some competition is needed to control hardware development costs.

nvidia's products may increase in price, and many customers are worried about this.

oracle executives said amd hardware can effectively help the company guard against crises caused by nvidia chip supply shortages, which existed for most of last year.

apart fromoracle corporationin addition, cloud service providers such as microsoft have also reduced their spending on ai computing giants.nvidiadependence intention.

1. nvidia's supply cuts have become the norm, and customers are turning to amd

karan batta said: "we found that the achievement of chip inference does not necessarily have to be tied to a specific supplier." karan batta believes thatas nvidia leads the server chip market, it needs competition to reduce hardware costs.

whether it is developing ai applications independently or purchasing such applications from software vendors,increased competition will enable companies to control the cost of ai inference.

amd's latest flagship server chip, due in 2023, may not be as powerful as nvidia's latest flagship h100, due in 2022, but oracle and microsoft executives think it will be.they have no choice but to support amd to promote competition.karan batta added: "amd is putting a lot of effort and doing a good job in improving the performance of its chips."

karan batta believes thatoracle's use of amd hardware also helps it insure against potential future supply shortages from nvidia, which it experienced for much of last year.

batta declined to disclose specific details or name clients, but his comments suggest oracle is looking to reduce its reliance on ai computing giant nvidia corp., a move similar to other cloud service providers such as microsoft corp.

in addition, the cost of leasing nvidia chips for inference has been falling, but its next-generation flagship ai chip may change that. other nvidia customers have privately expressed similar concerns.they worry that if amd or other companies don't take market share from nvidia, nvidia products will continue to be expensive.

although so far, this has not really happened.

amd said its data center revenue grew 115% to $2.8 billion in the june quarter compared with the same period last year, while nvidia's data center revenue grew 154% to $26.3 billion in the july quarter.

▲comparison of nvidia and amd data center revenue (source: the information)

2. oracle is collaborating with openai and xai to increase the scale of its chip cluster

meanwhile, oracle and other cloud service providers are booming, renting out more nvidia servers to clients like openai and elon musk’s xai.

in the u.s. server rental market, oracle ranks fourth after amazon's amazon cloud service, microsoft azure cloud technology and google cloud.oracle generated $2.4 billion in revenue from this business, up 45% year-over-year.

oracle entered the cloud computing field relatively late, but was somewhat lucky because last year was a time when ai startups were scrambling to buy cloud servers. oracle did not develop a large language model like openai gpt-4, so it did not pose a competitive threat to ai development customers, which gave it an advantage in the market.

now, oracle is looking to capture a bigger share of the ai ​​model training market, co-founderlarry ellisonoracle corp. is designing a data center with more than 1 gigawatt of capacity, enough to power a city as big as san francisco, oracle ceo larry ellison said on monday.

karan batta declined to provide details about the data center or say when it will be operational. to support its expansion, oracle will increase the ability to connect clusters of ai chips in series so they can work like a single computer, an ideal setup for training new large language models, he said.

those improvements include changing the way oracle cools its heat-generating chips and building tunnels for cables to connect chips in different buildings so they don't all have to be in the same place.

he said,oracle plans to at least double the size of its chip cluster from 64,000 gpus to more than 128,000but he did not specify the specific time.

moreover, even if startups use other cloud services to develop their core models, oracle can still benefit from the demand for ai training, batta said.

for example, the companies also need the hardware for other purposes, such as fine-tuning, in which developers optimize models for specific tasks.

karan batta said some ai companies are buying the rights to access older chips to handle these workloads.

according to the information artificial intelligence data center database,oracle may have reached a deal with openai and microsoft to use a facility in texas next year, but karan batta declined to comment on the matter for any specific company.

conclusion: more courage with more setbacks? claiming to be among the best in ai computing power

recent,oracle suffers setback, because musk decided to build his own gpu cluster in memphis, usa, instead of renting it from oracle, because the company was unable to build data centers quickly.

karan batta believes it is too early to tell whether xai can complete the colossus cluster of 100,000 h100 chips faster than oracle.

“i don’t have a baseline for comparison. are we much faster than xai? are we much slower? … i think compared to other cloud service providers, we would feel like we are one of the fastest in terms of ai computing power.”

source: the information