Ahead-looking: As AI workloads reshape computing, AMD is exploring a devoted neural processing unit to enhance or exchange GPUs in AI PCs. This transfer displays rising trade momentum towards specialised accelerators that promise sooner efficiency and better vitality effectivity – key elements as PC makers race to ship smarter, leaner machines.
AMD is exploring whether or not PCs may benefit from a brand new form of accelerator: a discrete neural processing unit. The corporate has lengthy relied on GPUs for demanding workloads, however the rise of AI-specific {hardware} opens the door to one thing extra environment friendly and specialised.
Rahul Tikoo, head of AMD’s shopper CPU enterprise, advised CRN that the chipmaker is in early talks with clients about what such a chip may appear to be and the place it may match.
“We’re speaking to clients about use instances and potential alternatives for a devoted accelerator chip that’s not a GPU however could possibly be a neural processing unit,” Tikoo stated throughout a briefing earlier than AMD’s Advancing AI occasion final month.
The thought arrives as PC makers like Lenovo, Dell Applied sciences, and HP search methods to dump AI processing from conventional CPUs and GPUs. Dell has already taken that step with its new Professional Max Plus laptop computer, which contains a Qualcomm AI 100 inference card – touted as the primary enterprise-grade discrete NPU for PCs.
Tikoo declined to disclose when AMD may launch such a chip, stressing future plans stay underneath an NDA. Nonetheless, he prompt the corporate has the items in place to maneuver shortly if it decides to proceed, making the leap to a discrete NPU believable.
AMD’s efforts to embed AI capabilities into Ryzen processors may present the muse. The corporate has used AI engine know-how from its Xilinx acquisition as the premise for NPU blocks in its newest chips – a transfer that might scale into stand-alone merchandise.
Christopher Cyr, CTO of Sterling Computer systems, stated the know-how roadmap is already clear.
“If this explicit NPU tile creates 50 TOPS [trillion operations per second], tack on two of them, make it 100 TOPS,” Cyr stated.
He emphasised that any discrete NPU from AMD should ship significant efficiency good points with out consuming the form of energy or producing the warmth typical of a stand-alone GPU. Effectivity is vital for PC makers striving to keep up skinny designs and lengthy battery life whereas including AI functionality. With out these vitality financial savings, a discrete NPU dangers changing into simply one other cumbersome, heat-producing element fairly than a real various to at the moment’s GPU-driven options.
Additionally learn: Opinion: The quickly evolving world of AI PCs
Cyr cited AMD’s Gaia open-source venture, designed to run giant language fashions regionally on Ryzen-based Home windows PCs, as proof that the corporate is laying the groundwork for a broader AI push.
“They’re making actually good inroads in the direction of leveraging that complete ecosystem,” he famous.
Whereas GPUs have been the default accelerator for years – and Nvidia wish to maintain it that manner – NPUs are reshaping the panorama. Intel, AMD, and Qualcomm have built-in NPUs into their newest processors. Nonetheless, there may be rising momentum for discrete variations that ship greater efficiency with out the warmth and energy draw of GPUs.
A few of the first makes an attempt got here from Intel, which geared up a 2023 Floor Laptop computer with a Movidius VPU earlier than its Core Extremely chips had onboard NPUs. Dell’s newest workstation takes issues additional with a Qualcomm card pushing 450 TOPS in a 75-watt envelope. Startups like Encharge AI are getting into the fray too, promising NPU add-ons with GPU-level compute capability at a fraction of the price and energy consumption.
AMD’s discrete NPU would broaden its product lineup past CPUs, GPUs, and built-in accelerators. This addition would provide OEMs a brand new choice to combine AI capabilities into PCs, probably offering a leaner and extra energy-efficient various to at the moment’s GPU-heavy setups.
