AMD signals push for discrete NPUs to rival GPUs in AI-powered PCs

[Cal Jeffrey]

Forward-looking: As AI workloads reshape computing, AMD is exploring a dedicated neural processing unit to complement or replace GPUs in AI PCs. This move reflects growing industry momentum toward specialized accelerators that promise faster performance and greater energy efficiency – key factors as PC makers race to deliver smarter, leaner machines.

AMD is exploring whether PCs could benefit from a new kind of accelerator: a discrete neural processing unit. The company has long relied on GPUs for demanding workloads, but the rise of AI-specific hardware opens the door to something more efficient and specialized.

Rahul Tikoo, head of AMD's client CPU business, told CRN that the chipmaker is in early talks with customers about what such a chip might look like and where it could fit.

"We're talking to customers about use cases and potential opportunities for a dedicated accelerator chip that is not a GPU but could be a neural processing unit," Tikoo said during a briefing before AMD's Advancing AI event last month.

The idea arrives as PC makers like Lenovo, Dell Technologies, and HP seek ways to offload AI processing from traditional CPUs and GPUs. Dell has already taken that step with its new Pro Max Plus laptop, which features a Qualcomm AI 100 inference card – touted as the first enterprise-grade discrete NPU for PCs.

Tikoo declined to reveal when AMD might launch such a chip, stressing future plans remain under an NDA. However, he suggested the company has the pieces in place to move quickly if it decides to proceed, making the leap to a discrete NPU plausible.

AMD's efforts to embed AI capabilities into Ryzen processors could provide the foundation. The company has used AI engine technology from its Xilinx acquisition as the basis for NPU blocks in its latest chips – a move that could scale into stand-alone products.

Christopher Cyr, CTO of Sterling Computers, said the technology roadmap is already clear.

"If this particular NPU tile creates 50 TOPS [trillion operations per second], tack on two of them, make it 100 TOPS," Cyr said.

He emphasized that any discrete NPU from AMD must deliver meaningful performance gains without consuming the kind of power or generating the heat typical of a stand-alone GPU. Efficiency is critical for PC makers striving to maintain thin designs and long battery life while adding AI capability. Without those energy savings, a discrete NPU risks becoming just another bulky, heat-producing component rather than a genuine alternative to today's GPU-driven solutions.

Also read: Opinion: The rapidly evolving world of AI PCs

Cyr cited AMD's Gaia open-source project, designed to run large language models locally on Ryzen-based Windows PCs, as evidence that the company is laying the groundwork for a broader AI push.

"They're making really good inroads towards leveraging that whole ecosystem," he noted.

While GPUs have been the default accelerator for years – and Nvidia would like to keep it that way – NPUs are reshaping the landscape. Intel, AMD, and Qualcomm have integrated NPUs into their latest processors. Still, there is growing momentum for discrete versions that deliver higher performance without the heat and power draw of GPUs.

Some of the first attempts came from Intel, which equipped a 2023 Surface Laptop with a Movidius VPU before its Core Ultra chips had onboard NPUs. Dell's latest workstation takes things further with a Qualcomm card pushing 450 TOPS in a 75-watt envelope. Startups like Encharge AI are entering the fray too, promising NPU add-ons with GPU-level compute capacity at a fraction of the cost and power consumption.

AMD's discrete NPU would broaden its product lineup beyond CPUs, GPUs, and integrated accelerators. This addition would offer OEMs a new option to integrate AI capabilities into PCs, potentially providing a leaner and more energy-efficient alternative to today's GPU-heavy setups.

Permalink to story:

AMD signals push for discrete NPUs to rival GPUs in AI-powered PCs

 

[yRaz]

I think with DDR6 right around the corner, the cost of GPUs, how little RT actually contributes and general dissatisfaction in the market that this absolutely could be AMDs GPU moment. What they need to do is stop putting only top teir GPUs with the fastest and most expensive mobile processors. They need 890Ms paired with 9600 class CPUs or they are going to price themselves out of the market. Very few people who want an APU for gaming are doing it because thet want to, it's because it's their only option. Only pairing your top tier mobile GPUs in 16 core APUs does nothing to help the market. know who your target market is and make products for them, you'll build brand loyalty and likely people who can afford better Dedicated GPUs in the future. Their NPUs are basically just rebranded APUs so they're probably going to brand their RDNA5/XDNA whatever they're going to call it as an NPU.

I've been saying for awhile people will likely have a second GPU(or NPU) for run local AI.

 

[opckieran]

TOPS are one thing, workspace is another. We’re overdue for a radical change our in cache/memory architecture, as shifting from DDR4>5>6 every 6-7 years, along with the on-die SRAM density bottleneck no longer appears to bode well in the long-term. Will 3D RAM ever make it to the mainstream? Or will we move away from RAM entirely?

 

[Dreadcthulhu]

I think with DDR6 right around the corner, the cost of GPUs, how little RT actually contributes and general dissatisfaction in the market that this absolutely could be AMDs GPU moment. What they need to do is stop putting only top teir GPUs with the fastest and most expensive mobile processors. They need 890Ms paired with 9600 class CPUs or they are going to price themselves out of the market. Very few people who want an APU for gaming are doing it because thet want to, it's because it's their only option. Only pairing your top tier mobile GPUs in 16 core APUs does nothing to help the market. know who your target market is and make products for them, you'll build brand loyalty and likely people who can afford better Dedicated GPUs in the future. Their NPUs are basically just rebranded APUs so they're probably going to brand their RDNA5/XDNA whatever they're going to call it as an NPU.

I've been saying for awhile people will likely have a second GPU(or NPU) for run local AI.

One issue with making an APU with a lower to mid CPU & good iGPU from AMD's perspective is that such a chip would only be slightly cheaper to produce than high end CPU + good iGPU; CPU cores take up a lot less die space than a decent GPU, so skimping on the CPU portion does not save much from the manufacturing costs. So cheaper APUs that still have decent iGPU power would be pretty low margin products.

 

[yRaz]

One issue with making an APU with a lower to mid CPU & good iGPU from AMD's perspective is that such a chip would only be slightly cheaper to produce than high end CPU + good iGPU; CPU cores take up a lot less die space than a decent GPU, so skimping on the CPU portion does not save much from the manufacturing costs. So cheaper APUs that still have decent iGPU power would be pretty low margin products.

And AMD wonders why their APUs aren't selling well... They are over valuing what the market will pay. People aren't going to buy a 16 core CPU just to get an 890M with it. Most people in this market are looking for a mobile platform, so the 16 core uses more than an 8 core, a 16 core CPU is pointless for gaming and finally, it increases the cost to consumers in a market segment that is VERY cost sensitive. Very few people out there are currently looking for 12 or 16 core laptop CPUs regardless of price. If I'm going to have to be stuck with a mobile system that is going to use more power I want that power to go to the GPU so I will buy an 8core system with a dedicated GPU in it.

 

[daffy duck]

I think with DDR6 right around the corner, the cost of GPUs, how little RT actually contributes and general dissatisfaction in the market that this absolutely could be AMDs GPU moment. What they need to do is stop putting only top teir GPUs with the fastest and most expensive mobile processors. They need 890Ms paired with 9600 class CPUs or they are going to price themselves out of the market. Very few people who want an APU for gaming are doing it because thet want to, it's because it's their only option. Only pairing your top tier mobile GPUs in 16 core APUs does nothing to help the market. know who your target market is and make products for them, you'll build brand loyalty and likely people who can afford better Dedicated GPUs in the future. Their NPUs are basically just rebranded APUs so they're probably going to brand their RDNA5/XDNA whatever they're going to call it as an NPU.

I've been saying for awhile people will likely have a second GPU(or NPU) for run local AI.

You mean like in all apu's released in last few years. Panther Lake will offer 100TOPS alone from the NPU, probably similar in Medusa Point. Intel will add even more powerful NPU in Nova Lake, but AMD still has no NPU in desktop. Current gen GPU's have TOPS >> 500, ~1400 in 9070XT, 3300 in 5090. Sure NPU will be far far more efficient and some AI software can leverage cpu+npu+gpu but for now npu seems far more important for mobile devices that rely on battery.

 

[human7]

There are plenty of dedicated "NPUs" in the server space - GPUs that aren't really GPUs at all since they only really have the compute engine.

For consumer devices, the hardware will probably come along. The enterprise market (outside of graphic design and such) doesn't need powerful GPUs, but AI has a lot of productivity potential, so I can see it there, but I also see a lot of roadblocks.

Some of that potential is locked away behind the size of the models (you need more memory and compute than a typical corporate laptop would have), some of that potential is locked behind SaaS (meaning the AI needs to be provided by some software application, not running locally - whether that's you as a user of said software or your company is serving that software from servers (not laptops) to your company's clients), some potential is locked behind the lack of good endpoint software and use cases, and some potential is locked behind the difficulty in turning cool demos into reliable production-ready solutions.

There are plenty of open source (or at least "source available") models you can run locally, and they are getting better all the time. So yeah, discrete NPUs will probably be a thing as an enterprise product where the GPU doesn't make a lot of sense, but until the software catches up, it's going to be challenging to take that from niche to widespread adoption. I can see it in high power workstations for e.g., research tasks, but laptops for everyday corporate types (including people like myself who do a lot of coding, machine learning, and analytics) is a bit of a stretch - why wouldn't you just use cloud or other over-the-internet solutions and compute (instead of on device)?

And for everyday consumers? Forget it. Not until small models become reliable over complex agentic workflows paired with software that for some reason runs locally instead of over the internet do I see it being a particular value add.

Actually, there is one place I can see a dedicated NPU coming in handy en mass for the consumer: video games. It's been imagined and discussed before, but consider NPCs, strategy games, or procedurally generated (now AI generated) games having an intelligence that changes what's possible in those genres. Much easier to argue for dedicated hardware because gamers will spend money on hardware and the GPU is already busy with the graphics. Though, like with games needing to be designed with a common denominator for graphics settings in mind, AI would need to be designed the same way. A toggle to turn off the AI, a minimum NPU, a recommended NPU, etc.

 

[adaplect]

Prediction: Every PC will soon have a CPU, a GPU, an NPU and a QPU

 

[wiyosaya]

Good. Keep this NPU crap out of my CPU. Let me buy an NPU coprocessor IF, and ONLY IF, I want one - which I don't.

 

[WhiteLeaff]

I think with DDR6 right around the corner, the cost of GPUs, how little RT actually contributes and general dissatisfaction in the market that this absolutely could be AMDs GPU moment. What they need to do is stop putting only top teir GPUs with the fastest and most expensive mobile processors. They need 890Ms paired with 9600 class CPUs or they are going to price themselves out of the market. Very few people who want an APU for gaming are doing it because thet want to, it's because it's their only option. Only pairing your top tier mobile GPUs in 16 core APUs does nothing to help the market. know who your target market is and make products for them, you'll build brand loyalty and likely people who can afford better Dedicated GPUs in the future. Their NPUs are basically just rebranded APUs so they're probably going to brand their RDNA5/XDNA whatever they're going to call it as an NPU.

I've been saying for awhile people will likely have a second GPU(or NPU) for run local AI.

There's still no bandwidth to power 16CU properly even with dual channel LPDDR5X 8533, but I agree that AMD shouldn't release horrendous products with only quad-cores with 2-4CU iGPUs, and sell them in $700-1000 laptops. It's so dumb lol

The minimum CU number should be 8, just as the minimum core number should be 6.

 

[yRaz]

There's still no bandwidth to power 16CU properly even with dual channel LPDDR5X 8533, but I agree that AMD shouldn't release horrendous products with only quad-cores with 2-4CU iGPUs, and sell them in $700-1000 laptops. It's so dumb lol

The minimum CU number should be 8, just as the minimum core number should be 6.

Too many people forget what gaming in a budget is like. People aren't expecting to play at all 4k or with RT on with these things. The people interested in these products would be happy with 60FPS low run at 1080p upscaled from 480p. These products should be the solution for low end gamers, not $300 5050's. They should be able to do a budget build for $500 and then throw a GPU in later.

The tech elitists need to stay out of budget build conversations. Budget builds are all about choosing a bottleneck you can live with and fixing it later

 

[WhiteLeaff]

Too many people forget what gaming in a budget is like. People aren't expecting to play at all 4k or with RT on with these things. The people interested in these products would be happy with 60FPS low run at 1080p upscaled from 480p. These products should be the solution for low end gamers, not $300 5050's. They should be able to do a budget build for $500 and then throw a GPU in later.

The tech elitists need to stay out of budget build conversations. Budget builds are all about choosing a bottleneck you can live with and fixing it later

Exactly, that's why super old things like the RX580 and 1060, old Xeons, Ryzen 2xxx/3xxxG still sell so well in low- and even high-income countries.