Qualcomm announces AI chips to compete with AMD and Nvidia

Qualcomm announced Monday that it volition merchandise caller artificial intelligence accelerator chips, marking caller contention for Nvidia, which has truthful acold dominated the marketplace for AI semiconductors.

The AI chips are a displacement from Qualcomm, which has frankincense acold focused connected semiconductors for wireless connectivity and mobile devices, not monolithic information centers.

Qualcomm said that some the AI200, which volition spell connected merchantability successful 2026, and the AI250, planned for 2027, tin travel successful a strategy that fills up a full, liquid-cooled server rack.

Qualcomm is matching Nvidia and AMD, which connection their graphics processing units, oregon GPUs, successful full-rack systems that let arsenic galore arsenic 72 chips to enactment arsenic 1 computer. AI labs request that computing powerfulness to tally the astir precocious models.

Qualcomm's information halfway chips are based connected the AI parts successful Qualcomm's smartphone chips called Hexagon neural processing units, oregon NPUs.

"We archetypal wanted to beryllium ourselves successful different domains, and erstwhile we built our spot implicit there, it was beauteous casual for america to spell up a notch into the information halfway level," Durga Malladi, Qualcomm's wide manager for information halfway and edge, said connected a telephone with reporters past week.

The introduction of Qualcomm into the information halfway satellite marks caller contention successful the fastest-growing marketplace successful technology: instrumentality for caller AI-focused server farms.

Nearly $6.7 trillion successful superior expenditures volition beryllium spent connected information centers done 2030, with the bulk going to systems based astir AI chips, according to a McKinsey estimate.

The manufacture has been dominated by Nvidia, whose GPUs person implicit 90% of the marketplace truthful acold and income of which person driven the institution to a marketplace headdress of implicit $4.5 trillion. Nvidia's chips were utilized to bid OpenAI's GPTs, the ample connection models utilized successful ChatGPT.

But companies specified arsenic OpenAI person been looking for alternatives, and earlier this period the startup announced plans to bargain chips from the second-place GPU maker, AMD, and perchance instrumentality a involvement successful the company. Other companies, specified arsenic Google, Amazon and Microsoft, are besides processing their ain AI accelerators for their unreality services.

Qualcomm said its chips are focusing connected inference, oregon moving AI models, alternatively of training, which is however labs specified arsenic OpenAI make caller AI capabilities by processing terabytes of data.

The chipmaker said that its rack-scale systems would yet outgo little to run for customers specified arsenic unreality work providers, and that a rack uses 160 kilowatts, which is comparable to the precocious powerfulness gully from immoderate Nvidia GPU racks.

Malladi said Qualcomm would besides merchantability its AI chips and different parts separately, particularly for clients specified arsenic hyperscalers that similar to plan their ain racks. He said different AI spot companies, specified arsenic Nvidia oregon AMD, could adjacent go clients for immoderate of Qualcomm's information halfway parts, specified arsenic its cardinal processing unit, oregon CPU.

"What we person tried to bash is marque definite that our customers are successful a presumption to either instrumentality each of it oregon say, 'I'm going to premix and match,'" Malladi said.

The institution declined to comment, the terms of the chips, cards oregon rack, and however galore NPUs could beryllium installed successful a azygous rack. In May, Qualcomm announced a partnership with Saudi Arabia's Humain to proviso information centers successful the portion with AI inferencing chips, and it volition beryllium Qualcomm's customer, committing to deploy up to arsenic galore systems arsenic tin usage 200 megawatts of power.

Qualcomm said its AI chips person advantages implicit different accelerators successful presumption of powerfulness consumption, outgo of ownership, and a caller attack to the mode representation is handled. It said its AI cards enactment 768 gigabytes of memory, which is higher than offerings from Nvidia and AMD.