{"id":15253,"date":"2026-06-07T03:45:56","date_gmt":"2026-06-07T02:45:56","guid":{"rendered":"https:\/\/innovatenews.site\/index.php\/2026\/06\/07\/amd-shipped-nvidias-new-ai-laptop-over-a-year-ago-and-the-software-is-finally\/"},"modified":"2026-06-07T03:46:01","modified_gmt":"2026-06-07T02:46:01","slug":"amd-shipped-nvidias-new-ai-laptop-over-a-year-ago-and-the-software-is-finally","status":"publish","type":"post","link":"https:\/\/innovatenews.site\/index.php\/2026\/06\/07\/amd-shipped-nvidias-new-ai-laptop-over-a-year-ago-and-the-software-is-finally\/","title":{"rendered":"AMD shipped Nvidia's new AI laptop over a year ago, and the software is finally\u2026"},"content":{"rendered":"
\n

The topic AMD shipped Nvidia’s new AI laptop over a year ago, and the software is finally\u2026<\/strong> is currently the subject of lively discussion \u2014 readers and analysts are keeping a close eye on developments.<\/p>\n

This is taking place in a dynamic environment: companies\u2019 decisions and competitors\u2019 reactions can quickly change the picture.<\/p>\n

Following Nvidia’s reveal of its RTX Spark laptops, I attended an AMD and HP roundtable at Computex. A fellow reporter asked Rahul Tikoo of AMD and Jim Nottingham of HP, two Vice Presidents at their respective companies, whether they welcomed the new competitor. After all, both companies have been pitching small machines that run big AI models locally for quite a while now. And Nvidia had just tossed its own hat into the ring.<\/p>\n

In response, Tikoo stood and gestured toward the HP-made and AMD-powered products scattered across the desk in front of us. He picked up the HP Strix Halo mini PC on the table, held it out for the room to see, and turned to Nottingham to ask a simple question: “Jim, when did you launch this system?”<\/p>\n

Nottingham’s reply was short, though accompanied by a slight grin \u2014 “CES 2025.” Tikoo, still standing, slowly repeated Nottingham’s answer, before looking around and grabbing another HP-made laptop from the table. Turning back to Nottingham, Tikoo was now mirroring his grin: “And this product, Jim \u2014 when did you launch it?”<\/p>\n

At this stage it was fairly clear what Tikoo’s point was going to be.<\/p>\n

Tikoo, handing the laptop over to us at the table before returning to his seat, circled back to the original question, answering it with a smile. “We have 35 products with Strix Halo in market,” he said. “Welcome, Nvidia, to the modern compute journey.”<\/p>\n

AMD, clearly, has confidence, and it’s not hard to see why from the outside looking in. After all, we don’t have pricing details for the RTX Spark, hands-on experiences have been guided, and the chip itself won’t ship until later this year. Tikoo, to his credit, also said he’s genuinely interested in seeing what Nvidia has specifically worked on and what it has achieved, but he was confident that Gorgon Halo (a Strix Halo refresh) would be a better product when it arrives in Q3.<\/p>\n

That same reporter who asked whether AMD welcomed a new competitor asked about the one thing most would say Nvidia has a clear advantage in: the software stack. After all, CUDA is the reason “just buy Nvidia” is a genuinely successful strategy, and ROCm, AMD’s alternative, has historically lacked many of the features developers need for local AI deployment and development. AMD has come a long way, and while it still isn’t CUDA, saying that you can’t do local AI on AMD would be inaccurate these days. The response to that stack question was more nuanced, but the short form is this: AMD is working on it, and as someone who’s been running ROCm for a while on a 7900 XTX, the old assumptions about AMD and local AI are quickly going stale.<\/p>\n

Nvidia’s RTX Spark is the GB10 essentially relaunched for laptops and small Windows PCs. It already powers the DGX Spark desktop, which is something that company CEO Jensen Huang himself confirmed when he tied the consumer N1 and N1X chips to the same design. It pairs a 20-core Arm “Grace” CPU with a Blackwell GPU carrying 6,144 CUDA cores, scales from 16GB up to 128GB of unified memory, and quotes up to 300 GB\/s of bandwidth with a claimed 1 PFLOP of FP4 compute. It was announced at Computex, it ships in the fall, and Nvidia hasn’t put a price on it beyond saying it targets the premium end.<\/p>\n

\"\"<\/figure>\n

AMD’s equivalent has been on sale for over a year. The Ryzen AI Max+ 395, codename Strix Halo, packs 16 Zen 5 cores and 32 threads alongside a 40-CU RDNA 3.5 iGPU and up to 128GB of unified memory, with as much as 96GB of that addressable as VRAM. It turned up in laptops like HP’s ZBook Ultra G1a in early 2025 and in a wave of mini-PCs not long after. That’s the lineup Tikoo was pointing to when he asked Nottingham about those products.<\/p>\n

At this year’s Computex, AMD also launched a turnkey rival to Nvidia’s DGX Spark, the Ryzen AI Halo developer mini-PC. It’s a Strix Halo machine with 128GB of memory that runs models up to 200 billion parameters, boots both Windows 11 and Linux, and opens for pre-orders this month at a price of $3,999. The price is targeted as well, as it’s the same figure Nvidia charged for the DGX Spark before the company later increased it to $4,699.<\/p>\n

By AMD’s own telling, the hardware isn’t really the point of the Halo. Tikoo describes it as an exercise in making the software layer disappear, with ROCm, PyTorch, and a handful of models preinstalled and held in what he called a best-known configuration. AMD ships it as a Ryzen AI Developer Center, with AI playbooks on the machine and available for download, validated model packages so things run on first launch, and a commitment to re-qualify the whole stack every month so it keeps working as the underlying pieces shift. Setting up something like OpenClaw can eat a whole weekend even when you know what you’re doing, Tikoo said, and the box exists to give you that weekend back.<\/p>\n

On the spec sheet the two sit close. Both top out at 128GB of unified memory, and Nvidia’s 20-core Arm CPU lines up against AMD’s 16-core, 32-thread x86 part. AMD’s next step, Gorgon Halo, will push that to 192GB and 300-billion-parameter models in the third quarter. The bottom of Nvidia’s range is also relevant here, as its configurations will begin at 16GB of RAM. When asked whether AMD had anything for buyers who don’t want a $4,000 machine, the answer was pretty simple: you can already buy that today, thanks to the cheaper Ryzen AI 300 and 400 chips that already cover the lighter end.<\/p>\n

Nvidia does hold some paper advantages. Its 300 GB\/s of memory bandwidth just about surpasses AMD’s 256 GB\/s theoretical ceiling. As well, the Blackwell GPU with its FP4 tensor hardware should give it a lead on raw compute and prompt processing, which is the part of inference that decides how long you wait for the first token. There’s a caveat to that 300 GB\/s number, by the way: the GB10 was also said to be a chip with 300 GB\/s memory bandwidth at Hot Chips last year, though in actuality, it ended up being 273 GB\/s.<\/p>\n

On the desktop side, the math already favors AMD before the laptop fight even begins. The Ryzen AI Halo dev box is costly (though currently cheaper than the DGX Spark), but a 128GB Strix Halo mini-PC like GMKtec’s EVO-X2 sells for around $3,300, and Framework’s Desktop starts around the same, too. You pay the $700 additional premium for the validated, qualified software stack rather than the silicon, which is exactly the friction Tikoo says that box exists to remove.<\/p>\n

Tikoo answered a question on AMD’s stack with a proclamation that ROCm has come a long way, that open source is AMD’s “bet,” and that “open source is the way to go.” Tikoo is right on that point, as ROCm is no longer a simple project bolted onto the side of the AI ecosystem. PyTorch is probably the best demonstration of that. PyTorch 2.9 brought ROCm into its experimental wheel-variant work, making installation less awkward with compatible tooling, but the more important parts are what’s happened since. 2.11 added device-side assertions and TopK\/radix-select optimizations for AMD GPUs, and 2.12 added expandable memory segments, rocSHMEM symmetric memory collectives, and FlexAttention pipelining. The framework that sits under a huge amount of modern AI work now treats AMD as an actual accelerator target, rather than a community workaround.<\/p>\n

HIP is what closes some of the remaining distance. On a ROCm build of PyTorch, much of the familiar torch.cuda API still exists, but those calls execute through HIP on AMD GPUs rather than through CUDA on Nvidia hardware. In day-to-day use, it means a lot of PyTorch code written with CUDA assumptions can run on AMD hardware without being rewritten around a separate AMD-specific API.<\/p>\n

\"\"<\/figure>\n

The rest of the local inference toolkit has followed. llama.cpp has both Vulkan and ROCm backends, and Ollama, LM Studio, and ComfyUI all run on AMD now. On Strix Halo, Vulkan is often the easiest path and, in many llama.cpp-style setups, can be faster for token generation, while ROCm tends to matter more for prompt processing, long-context behavior, Flash Attention, rocWMMA, and anything that benefits from AMD’s HIP compute stack. Which backend wins depends on the model, context length, quantization, and app, but the important part is that there’s now a choice. Two years ago, this was a weekend of repeated compilations, testing, measurements, and hope. Now, much of it just installs and works.<\/p>\n

The version situation has improved too. ROCm 7.2.4 is the latest Linux-side quality release focused on inference performance and stability for Instinct GPUs, while AMD\u2019s Radeon and Ryzen documentation separately lists ROCm 7.2.1 support for Radeon RX 9000, select RX 7000 cards, and Ryzen AI Max, AI 300, and select AI 400 APUs. Windows is now part of AMD\u2019s ROCm story for consumer hardware, especially around PyTorch, even if Linux remains the broader and more mature target.<\/p>\n

I have been running ROCm long enough to remember when a 7900 XTX meant manually building half the stack and accepting that some of it would never work. That card and the hardware around it sit on the official support list today. “You need CUDA” is increasingly becoming less relevant as time goes on, and that’s fantastic to see.<\/p>\n

None of that makes ROCm equal to CUDA, but AMD is refreshingly aware of that. When I asked Tikoo what gaps still needed filling, the first thing he named was sandboxing for new agentic use cases. “That’s one of the things we want to address quickly,” he said.<\/p>\n

The bigger one, though, was what he called getting “day-zero on the endpoints.” The goal is to make ROCm development for AMD’s data-center Instinct cards \u201c100% usable\u201d on endpoint integrated-graphics solutions, meaning the iGPUs in machines like Strix Halo. “So we keep increasing that library,” he said. It is a candid answer, because it gets at the gap AMD still has to close: ROCm is much better than it used to be, but AMD still needs more of that data-center software work to carry cleanly down to consumer and integrated graphics.<\/p>\n

The NPU has the same problem, just with a different software stack. “When you want performance you go to the GPU, but when you want efficiency you go to the NPU,” Tikoo said. That means AMD needs an ISV library for NPU use cases too, and he called that “a big focus.”<\/p>\n

Historically, AMD has struggled with that kind of software parity. Flash Attention on ROCm has often lagged behind, and Strix Halo has already exposed the sort of edge case that still makes ROCm feel less mature, with PyTorch Flash Attention failing on gfx1151 in some builds. Quantization libraries like bitsandbytes are no longer CUDA-only in the way they once were, but they are still a good example of the problem: CUDA is the default path, and AMD support tends to arrive later, with more caveats, and after more work from AMD, maintainers, and users.<\/p>\n

Training is the weakest story of all, although that’s changing too. Zyphra\u2019s ZAYA1-8B was trained on AMD Instinct MI300X clusters, which is an important milestone for AMD to cross. The company describes it as the first large-scale MoE foundation model trained entirely on AMD Instinct MI300X GPUs, AMD Pensando networking, and ROCm. It’s impressive, but read between the lines: we’re several years into the LLM craze, and if that kind of run can still be newsworthy, it tells you how CUDA-first the training world remains.<\/p>\n

The broader ecosystem also moves CUDA-first. New models? CUDA first. AI software packages? CUDA first. ROCm eventually arrives, but usually with a delay, and often with more rough edges than the Nvidia path. Ollama, for example, still has a habit on Strix Halo of timing out while it hunts for the GPU and quietly dropping to the CPU. None of that’s a dealbreaker if you don’t mind tinkering, but you’ll hit problems a CUDA user never would.<\/p>\n

The hardware itself can be another hurdle for AMD. Nvidia\u2019s Tensor Cores and CUDA libraries remain the reference point for the low-precision matrix math these workloads lean on, especially once you get into FP8, FP4, and the fast prompt-processing path reviewers will be watching closely. If Nvidia\u2019s laptop turns out to be faster at prompt processing once consumers get hold of it by the end of the year, I won’t be too surprised.<\/p>\n

With all of that said, ROCm is good enough to stop being the reason not to buy AMD. It’s not good enough to call it CUDA, but for an enthusiast running local LLMs on a Strix Halo machine, PyTorch with llama.cpp or Ollama or LM Studio is a supported, working path. As are many of the local AI workloads you’d want to run, such as ComfyUI. The software gap that used to justify ignoring AMD has shrunk to a list much smaller than you’d think. While Nvidia may well ship something excellent in the fall, you can buy AMD’s version today and run it, which is more than you can say for Nvidia’s machine trapped behind glass with no promises when it comes to pricing.<\/p>\n