<p class="canvas-atom canvas-text Mb(1.0em) Mb(0)–sm Mt(0.8em)–sm" type="text" content="Qualcomm (NASDAQ: QCOM) recently revealed a new AI chipset, the Cloud AI 100, for cloud data center servers. The company claims that the chip will be over 50 times more powerful than its top-end Snapdragon 855 mobile chipset, and will be optimized for making decisions based on voice and image recognition.” data-reactid=”11″>Qualcomm (NASDAQ: QCOM) recently revealed a new AI chipset, the Cloud AI 100, for cloud data center servers. The company claims that the chip will be over 50 times more powerful than its top-end Snapdragon 855 mobile chipset, and will be optimized for making decisions based on voice and image recognition.
Qualcomm will reveal more details before it sends out the first samples later this year, and will launch the chips commercially in 2020. This seems like a way for Qualcomm to diversify its business away from smartphones, but investors should recall that it already tried — and failed — to crack the data center market before.
Image source: Getty Images.
<h2 class="canvas-atom canvas-text Mb(1.0em) Mb(0)–sm Mt(0.8em)–sm" type="text" content="Qualcomm’s past blunders” data-reactid=”25″>Qualcomm’s past blunders
<p class="canvas-atom canvas-text Mb(1.0em) Mb(0)–sm Mt(0.8em)–sm" type="text" content="In late 2017, Qualcomm launched its Centriq data center chips. The initial benchmarks indicated that the ARM-based chips could be viable alternatives to Intel‘s widely used (NASDAQ: INTC) Xeon CPUs, and tech giants like Microsoft started testing Qualcomm’s chips in its data centers. It also supplied chips for a server joint venture in China.” data-reactid=”26″>In late 2017, Qualcomm launched its Centriq data center chips. The initial benchmarks indicated that the ARM-based chips could be viable alternatives to Intel‘s widely used (NASDAQ: INTC) Xeon CPUs, and tech giants like Microsoft started testing Qualcomm’s chips in its data centers. It also supplied chips for a server joint venture in China.
For a while, it seemed like Qualcomm might dent Intel’s near-monopoly in data center CPUs. But last May Qualcomm’s data center technologies chief Anand Chandrasekher, who previously worked at Intel, abruptly resigned.
<p class="canvas-atom canvas-text Mb(1.0em) Mb(0)–sm Mt(0.8em)–sm" type="text" content="Several rounds of layoffs subsequently reduced the size of the data center unit from 1,000 employees to about 50 by the end of 2018. Qualcomm insisted that its data center business wasn't dead yet, but it certainly looked like a failed experiment.” data-reactid=”28″>Several rounds of layoffs subsequently reduced the size of the data center unit from 1,000 employees to about 50 by the end of 2018. Qualcomm insisted that its data center business wasn’t dead yet, but it certainly looked like a failed experiment.
<p class="canvas-atom canvas-text Mb(1.0em) Mb(0)–sm Mt(0.8em)–sm" type="text" content="Qualcomm's data center business flopped for three simple reasons: Intel's Xeons were still the industry standard for servers, AMD‘s Epyc chips were considered better alternatives (since they used the same x86 architecture as Xeons instead of the Centriq’s ARM architecture), and Broadcom‘s hostile takeover attempt forced Qualcomm to cut costs from non-core projects like data center chips.” data-reactid=”29″>Qualcomm’s data center business flopped for three simple reasons: Intel’s Xeons were still the industry standard for servers, AMD‘s Epyc chips were considered better alternatives (since they used the same x86 architecture as Xeons instead of the Centriq’s ARM architecture), and Broadcom‘s hostile takeover attempt forced Qualcomm to cut costs from non-core projects like data center chips.
<p class="canvas-atom canvas-text Mb(1.0em) Mb(0)–sm Mt(0.8em)–sm" type="text" content="Several parties — including ARM's parent company SoftBank, Singapore’s Temasek Holdings, former Qualcomm CEO Paul Jacobs, and former Intel executive Renee James — all considered buying Qualcomm’s data center unit. However, none of those talks panned out, and the unit seemed destined to wither away.” data-reactid=”30″>Several parties — including ARM’s parent company SoftBank, Singapore’s Temasek Holdings, former Qualcomm CEO Paul Jacobs, and former Intel executive Renee James — all considered buying Qualcomm’s data center unit. However, none of those talks panned out, and the unit seemed destined to wither away.
Image source: Getty Images.
<h2 class="canvas-atom canvas-text Mb(1.0em) Mb(0)–sm Mt(0.8em)–sm" type="text" content="Will the Cloud AI 100 fare any better?” data-reactid=”43″>Will the Cloud AI 100 fare any better?
The Centriq chip’s harsh trajectory may not necessarily be repeated. The key difference between the Cloud AI 100 and the Centriq is that it’s a discrete AI accelerator instead of a CPU aimed at replacing Xeons. AI accelerators — which include GPUs, FPGAs (field-programmable gate arrays), and custom ASICs (application-specific integrated circuits) — are add-in boards that work alongside CPUs.
<p class="canvas-atom canvas-text Mb(1.0em) Mb(0)–sm Mt(0.8em)–sm" type="text" content="AI accelerators can be applied to specific AI tasks. NVIDIA‘s (NASDAQ: NVDA) GPUs, for example, can speed up deep learning calculations at data centers. Programmable chips like Intel or Xilinx‘s FPGAs can be customized for specific AI tasks, while ASICs are created for specific tasks but can’t be reprogrammed.” data-reactid=”49″>AI accelerators can be applied to specific AI tasks. NVIDIA‘s (NASDAQ: NVDA) GPUs, for example, can speed up deep learning calculations at data centers. Programmable chips like Intel or Xilinx‘s FPGAs can be customized for specific AI tasks, while ASICs are created for specific tasks but can’t be reprogrammed.
<p class="canvas-atom canvas-text Mb(1.0em) Mb(0)–sm Mt(0.8em)–sm" type="text" content="Qualcomm hasn't revealed much about the Cloud AI 100's design, but AnandTech noted that the accelerator’s dedicated focus on AI "inference" tasks like voice and image processing suggests that it will be an ASIC accelerator.” data-reactid=”50″>Qualcomm hasn’t revealed much about the Cloud AI 100’s design, but AnandTech noted that the accelerator’s dedicated focus on AI “inference” tasks like voice and image processing suggests that it will be an ASIC accelerator.
<p class="canvas-atom canvas-text Mb(1.0em) Mb(0)–sm Mt(0.8em)–sm" type="text" content="Data center customers will likely stick with one type of accelerator instead of buying all three, so Qualcomm will be competing against NVIDIA, Intel, Xilinx, and even Alphabet‘s (NASDAQ: GOOG) (NASDAQ: GOOGL) Google in this market. Google’s TPUs (tensor processing units), which were launched three years ago, are arguably the most well-known ASIC accelerators.” data-reactid=”51″>Data center customers will likely stick with one type of accelerator instead of buying all three, so Qualcomm will be competing against NVIDIA, Intel, Xilinx, and even Alphabet‘s (NASDAQ: GOOG) (NASDAQ: GOOGL) Google in this market. Google’s TPUs (tensor processing units), which were launched three years ago, are arguably the most well-known ASIC accelerators.
<h2 class="canvas-atom canvas-text Mb(1.0em) Mb(0)–sm Mt(0.8em)–sm" type="text" content="A smart move if Qualcomm sticks with it” data-reactid=”52″>A smart move if Qualcomm sticks with it
Launching an AI accelerator instead of another CPU is a smart move, since this is still a nascent market with plenty of room for growth. Research and Markets estimates that the global data center accelerator market will grow from $2.5 billion in 2018 to $21.2 billion by 2023.
The market also seems undecided regarding the usage of GPUs, FPGAs, ASICs, or more powerful CPUs for AI acceleration tasks. Therefore, Qualcomm’s Cloud AI 100 probably won’t immediately cause problems for the Centriq. If it’s successful, it could diversify Qualcomm’s business away from its mobile business, which faces sluggish smartphone sales, competition from rival chipmakers, and ongoing lawsuits and probes over its licensing fees.
However, Qualcomm will likely need to stick with this project for years before it bears fruit. If Qualcomm doesn’t prematurely starve this business, as it did with the Centriq, it could eventually diversify its business and widen its moat against Intel and NVIDIA.
<p class="canvas-atom canvas-text Mb(1.0em) Mb(0)–sm Mt(0.8em)–sm" type="text" content=" More From The Motley Fool ” data-reactid=”56″> More From The Motley Fool
<p class="canvas-atom canvas-text Mb(1.0em) Mb(0)–sm Mt(0.8em)–sm" type="text" content="Suzanne Frey, an executive at Alphabet, is a member of The Motley Fool’s board of directors. Teresa Kersten, an employee of LinkedIn, a Microsoft subsidiary, is a member of The Motley Fool’s board of directors. Leo Sun has no position in any of the stocks mentioned. The Motley Fool owns shares of and recommends Alphabet (A shares), Alphabet (C shares), Microsoft, and NVIDIA. The Motley Fool owns shares of Qualcomm. The Motley Fool recommends Broadcom Ltd and Xilinx. The Motley Fool has a disclosure policy.” data-reactid=”65″>Suzanne Frey, an executive at Alphabet, is a member of The Motley Fool’s board of directors. Teresa Kersten, an employee of LinkedIn, a Microsoft subsidiary, is a member of The Motley Fool’s board of directors. Leo Sun has no position in any of the stocks mentioned. The Motley Fool owns shares of and recommends Alphabet (A shares), Alphabet (C shares), Microsoft, and NVIDIA. The Motley Fool owns shares of Qualcomm. The Motley Fool recommends Broadcom Ltd and Xilinx. The Motley Fool has a disclosure policy.
Add Comment