Nvidia and Intel’s RTX SoCs could pose an existential threat to AMD’s APUs — if two companies can actually pull it off

For years, AMD’s most dangerous weapon hasn’t been its killer desktop CPUs or its most efficient laptop chips. It’s been the APU, a quiet killer of discrete GPUs in gaming consoles, handhelds, and notebooks alike. With the Ryzen AI 300 series, AMD proved it could pair 16-core Zen 5 CPUs with RDNA 3.5 GPUs and a 50 TOPS NPU into a single power-efficient die. It was a pitch neither Intel nor Nvidia could match. Until now.

In a surprise move last week, Intel and Nvidia jointly confirmed that the companies will co-develop x86 system-on-chips that package Intel CPUs and Nvidia RTX GPU chiplets together, bound by NVLink and unified memory. These “RTX SoCs” are being positioned as premium integrated chips that can power laptops, desktops, and consumer AI systems without a discrete GPU.

A packaged counterattack

The idea is simple. Intel will supply the CPU tile, and Nvidia will supply the GPU tile. Both are joined in a single multi-die package using Intel’s advanced Foveros and EMIB interconnects, the same technologies Intel uses to stack and stitch together chiplets on its own high-end platforms. The critical link is Nvidia’s proprietary NVLink, which offers up to 900 GB/s of on-package bandwidth, more than ten times faster than PCIe 4.0.

There are parallels with AMD’s APUs: one package, one memory pool, CPU and GPU working in lockstep. The difference is scale. Where AMD’s Strix Halo tops out at 40 RDNA 3.5 compute units and a 50 TOPS NPU, an Intel/Nvidia SoC could pack a full RTX-class GPU with ray-tracing cores, tensor units, and CUDA throughput tuned for both gaming and AI.

Nvidia CEO Jensen Huang called it “a new class of integrated graphics laptops the world has never seen before,” in a joint press conference following the announcement. What Nvidia and Intel are proposing is essentially a top-down rethink of what a CPU package can include. Instead of discrete compute domains with duplicated memory, we’re looking at a system that acts more like a single accelerated core — x86 and RTX fused into one thermal envelope, designed to scale from thin-and-lights to AI-enabled workstations.

How the silicon stacks up

AMD’s Strix Halo APU (featured in systems like the Framework Desktop) is no slouch. Built on TSMC’s 4nm process, it pairs up to 16 Zen 5 cores with a 40-CU RDNA 3.5 GPU, supports 128GB of LPDDR5X-8000, and houses a dedicated NPU delivering 50 TOPS. Bandwidth is supplied by a massive LPDDR5X interface, backed by 32MB of cache. Performance is competitive with an RTX 4060, but in a tighter, more power-efficient design. AMD sells it as a no-GPU-needed package. And for many workloads, that’s exactly what it is. However, it comes with clear power and bandwidth constraints, as well as the limitations of an integrated GPU die.

Nvidia and Intel’s blueprint attacks those constraints directly. The GPU tile going into these SoCs isn’t some low-end mobile salvage part. According to Huang, it’s an RTX chiplet, which should mean full ray-tracing, CUDA, and Tensor cores — in other words, a fully-fledged GPU. Then there’s the inclusion of NVLink, which provides far more bandwidth and superior latency than PCIe, allowing the CPU and GPU to share memory with limited bandwidth and latency penalties. Meanwhile, Foveros 3D packaging means each die can be built on its optimal process node, rather than compromised in a monolithic slab.

In comparison, AMD’s integrated GPU has to share die area with the CPU and NPU. The Nvidia-Intel package will not be vertically integrated; instead, it will be asymmetric. “Nvidia’s GPU technology is based on TSMC’s foundry… connecting Nvidia’s GPU die chiplet with Intel’s CPUs in a multi-technology packaging capability… it’s really a fabulous way of mixing and matching technology,” Huang said during the press conference.

But silicon isn’t built in press releases. The challenges for the duo will include coordinating dies from different fabs, tuning thermals in laptop envelopes, and persuading OEMs to bet on yet another premium part.

A solid business case for Nvidia and Intel

There’s a genuine business case for RTX SoCs that the duo could tap into. Laptops that currently rely on a Core Ultra CPU and a discrete RTX 4050/70 could instead ship with a single chip. That means a smaller motherboard, a simpler power subsystem, and less thermal juggling between CPU and GPU. Fewer parts may also result in a lower failure rate and potentially lower costs, but that depends on many unknown factors.

And then there’s branding. AMD has struggled to convince laptop buyers that its integrated graphics are serious gaming tools, even when they match or exceed performance expectations. Nvidia doesn’t have that problem. If the sticker says “RTX,” the assumption is that it’s fast. OEMs will lean into that, even if it means higher chip costs. Nvidia owns more than 90% of the discrete GPU market, and its RTX logo is a selling point that AMD has never matched in consumer mindshare. If Intel can price these SoCs competitively, we’ll likely see major brands like Asus and MSI lining up to buy them.

The flip side to all this is cost. Combining Intel’s leading-edge CPUs with Nvidia’s premium GPU chiplets won’t come cheap. AMD’s APUs may lose out in raw power, but their vertical integration makes them cheaper to fab and easier to bin. For price-sensitive segments, such as budget gaming laptops and mainstream productivity notebooks, AMD will likely maintain its foothold. But that’s fine, because Nvidia and Intel are aiming higher up the stack, where OEMs already spend freely for discrete GPUs.

The specter of the unknown

Any response from AMD is probably going to emphasize control. Its APU roadmap is in-house with seamless integration. Nvidia and Intel, by contrast, are looking to run chiplets from two vendors on different process nodes and different fabs. But that might not matter from an OEM’s point of view. If the Nvidia-Intel SoC is cheaper than a GPU+CPU combo and faster than an APU, it’s not a tough decision.

Of course, none of this exists yet. The silicon from the partnership is expected no sooner than 2027, and Intel has said it remains committed to Arc graphics. However, the reality is that Intel and Nvidia will be building a whole new stack, which could represent an existential threat in one of AMD’s fastest-growing categories. Strix Halo was designed to expand the company’s APU footprint beyond cheap laptops, into gaming notebooks, and even small desktop systems. Our own testing shows that it largely succeeded — an AMD APU that can trade blows with low-end discrete GPUs — but Intel and Nvidia are promising a leap beyond that, with IP that AMD cannot easily match.

The most important word in this entire partnership is “if.” If Intel can deliver leading-edge x86 cores on time. If Nvidia can make its GPU chiplets play nice with Intel’s CPUs. If unified memory over NVLink works in practice as well as in theory. And if OEMs buy into the concept at scale.

For now, the roadmap for the first launch could conceivably stretch into 2027, which gives AMD time to respond and leaves plenty of room for execution risk. Remember Kaby Lake-G? Intel’s ill-fated 2017 experiment fusing its CPU with AMD graphics didn’t survive a single generation. Still, the stakes are higher this time, as AI workloads balloon and gaming laptops represent a core growth market, and both Intel and Nvidia need each other to counterbalance AMD’s momentum.

If they can pull it off, AMD could face its toughest competition yet in the market it helped create, and Intel, long the laggard in GPUs, will suddenly have RTX inside.

Follow Tom's Hardware on Google News, or add us as a preferred source, to get our up-to-date news, analysis, and reviews in your feeds. Make sure to click the Follow button!