Major tech companies have been competing to build data centers, facilities where AI models are trained and operated globally and facilities where Nvidia AI chips dominate. However, with the rapid development of AI technology, the superficial competition of data centers (and NVIDIA).
It all boils down to the process of generating answers from AI models (called “inference”). Currently, reasoning is largely done in data centers.
But in the future, mobile chip expert Qualcomm’s powerful bargaining chips may move out reasoning from data centers, smartphones and personal computers.
The thriving AI reasoning market is staked high – participation already accounts for 40% of NVIDIA’s data center revenues and is growing rapidly. Soon after, it surpassed the training model as the main source of AI revenue for CHIP players.
“Here, there is a high-level battle between NVIDIA and Qualcomm,” wrote Dean Bubley, a founder of disruptive analysis.
this The biggest change This year’s AI is introducing the so-called inference model, a new technology for reasoning. According to NVIDIA CEO Jensen Huang, these models gradually solve the problem of using more computing resources and using more computing resources, which is what NVIDIA CEO Jensen Huang said.
At the Mobile World Congress at the world’s largest mobile and communications trade show in Barcelona this month, Qualcomm and Micron provided reasons for reasoning where reasoning goes from here, and Nvidia showed how it plans to maintain its lead.
Qualcomm said the need for continuous availability, fast response time, privacy and security, and lower costs mean that AI reasoning will inevitably be transferred to users’ devices.
“It makes sense to run it on a device for many different reasons,” Qualcomm Chief Financial Officer Akash Palkhiwala told Barron’s in an interview with MWC.
However, reasoning about mobile devices can quickly consume battery life. Another problem is that AI chips can process data faster than memory systems, thus limiting inference performance (so-called “memory walls”), which can create frustrating lag for users.
Memory-Chip expert Micron Technology is working on a fix. Its latest high-end smartphone chip saves up to 15% compared to previous generations. Micron’s potential solution to “memory walls” might be semiconductor architectures that can perform some inference operations directly inside the memory chip (called processing in memory).
Qualcomm and Micron hope consumers will buy Advanced smartphone Equipped with chips to handle AI reasoning. However, intakes may be slow, according to International Data Corporation – smartphone shipments are only expected to grow by 2.3% this year, according to International Data Corporation.
Meanwhile, NVIDIA says its ready-to-do solution can be implemented immediately. Its purpose is to sell its chips to telecom companies, believing that local wireless infrastructure is the right place to do AI inference. The company said it is close enough to users to reduce lag while also leveraging existing power supplies.
This is difficult. Telecom companies are cautious about making large investments after they use large amounts of money on 5G infrastructure for insignificant returns. However, NVIDIA has plans for multiple recipients.
Last week, South Korea’s Samsung Electronics said it was integrating NVIDIA’s hardware to make its wireless network look like a data center. Verizon Communications announced its “AI Connect” suite of enterprise products earlier this year, which includes a partnership with Nvidia.
Most notably, Nvidia has established close relationships with the Japanese Telecommunications Company SoftBank is adhering to this concept. The two companies jointly operate technology trials and jointly estimate that telecom operators can earn about $5 in inference revenue each time they invest in combined AI and wireless infrastructure.
Mauro Filho, director of Softbank AI-Ran America, told Barron’s that going beyond text-based AI to other media, such as sound and video, is a key key, including saving device batteries, including saving device batteries, Mauro Filho, director of SoftBank AI-Ran America, told Barron’s.
“Have the ability to transfer some of this workload to the network [AI] At the same time, there is no need to choose. ” Filho said.
However, there are doubts about the idea that wireless networks can play an important role in AI processing.
“Whether training or reasoning, most overall computing will be concentrated in large data centers or devices,” Bubley of Inspriage Analysis wrote.
Through wireless infrastructure and devices, some mix of reasoning will inevitably occur, but the exact proportion will determine which companies become winners.
At present, Nvidia still has obvious advantages.
Large tech companies are locked in using their chips Huge investment In the data center, it will be desirable to use the same hardware as training an AI model. ChIP sales to telecom operators will further consolidate NVIDIA’s position on Qualcomm and other possible challenges to AI.
