Computex research: Where AMD, Nvidia and Intel are falling short

Contents

While last year's Computex IT trade fair was buzzing with new PC chips, this year was quieter than expected. c't and heise online went behind the scenes to find out about Intel's next generation of Panther Lake notebooks (probably Core Ultra 300), what AMD has up its sleeve and why Nvidia hasn't announced a PC processor after all.

In 2024, AMD and Intel hastily brought forward their Ryzen AI 300 and Core Ultra 200V in order to have a competing product with a strong AI unit against Qualcomm's Snapdragon X. The inherent disadvantage: If manufacturers launch a single chip generation earlier than planned, this does not also apply to their respective successors, which are still in earlier stages of development. In fact, they are more likely to be slowed down if resources are withdrawn.

Intel's current status

At Computex 2025, Intel presented a prototype system with Panther Lake, which inherits both Lunar Lake (Core Ultra 200V: strong NPU, small CPU) and Arrow Lake (Core Ultra 200H: big CPU, but no NPU for Copilot+) and uses Intel's brand-new 18A manufacturing process for the first time. However, even before the trade fair, it was already known that instead of the promised sales launch, only mass production will start in 2025. From then on, it will be several months before finished notebooks arrive in stores.

Perhaps some first notebooks will be on the shelves of some retailers in some regions shortly before Christmas, as was the case with Meteor Lake, but most will become available in 2026. The fact that Panther Lake, like Lunar Lake, is extremely unattractive for laptop manufacturers also plays a role here: Lunar Lake already requires special mainboards that cannot accommodate either the previous or the next generation and are therefore particularly expensive to develop. Panther Lake is yet another new platform that will not be suitable for its successor Nova Lake.

AMD: Refresh in sight

With Strix Point aka Ryzen AI 300, AMD achieved the notable success of selling the first x86 CPU for Copilot+ last year. However, there were only a few chips in the first few months, most of which Asus received. The situation improved in the fall – and repeated itself in the spring: the smaller offshoot Krackan Point (also Ryzen AI 300) with a maximum of eight instead of twelve Zen 5(c) cores and a smaller graphics unit made its debut at CES in January, but is only now being found in notebooks. According to notebook manufacturers, Krackan Point is simply too expensive, so they prefer to use the Hawk Point Ryzen 200 refresh.

At Computex 2025, AMD did not have a successor for the Strix Point, which is now a year old. Instead, the name Gorgon Point is circulating in the rumor mill. This is supposedly a rather unexciting refresh, which is to be fully exploited in terms of marketing with the new name Ryzen AI 400. AMD needs this for the 2026 product year because the next technical development is still a long way off: Only Medusa Point brings Zen 6 cores and a GPU with new RDNA4 architecture. The latter uses the recently released desktop graphics cards of the Radeon RX 9000 series and the former was specially designed for TSMC's N2 manufacturing process. This means that the entire chip is more likely to be something for the next Computex in a year's time ... or even later.

Qualcomm arouses curiosity about Snadragon X successor

By then, not only Intel will have something new in its portfolio, but certainly Apple and probably Qualcomm as well. The only noteworthy announcement made by company boss Cristiano Amon at his Computex keynote speech was that the successor to the Snapdragon X will be unveiled at the Snapdragon Summit in-house exhibition in September. And a press spokesperson promised that interested parties would then be able to buy such notebooks quickly – that sounds like the run-up to Christmas or the turn of the year.

Everything in the green at Nvidia?

Speaking of the end of the year and ARM processors: Nvidia already unveiled its DGX Spark mini workstation at CES in January, including its own GB10 processor, which consists of two chiplets. Nvidia developed the GPU part itself, while the CPU part, which also includes all other SoC functions from USB to RAM controller to network, comes from Mediatek.

Nvidia CEO Jensen Huang repeatedly emphasized at various events, which he attended as keynote or guest speaker, how great the cooperation was, that the complex chiplet network was already running in the first approach and that mass production would start.

Since journalists, as professional pessimists, are wary of the most effusive praise, and we are also generally skeptical about Nvidia, we asked around in Taiwan. It was suspicious that Nvidia exhibited numerous DGX Spark systems with GB10 at the GTC in-house exhibition, which took place parallel to Computex, but there was still no running system to be seen in the demo area – months after the unveiling.

As a well-informed source told us, GB10 probably has a small, but subtle bug with piquant consequences. The chip is generally working, but the display controller is apparently screwed up: it is only supposed to only run one resolution. And this is not even one of the resolutions that monitors usually display.

In conversation, one of the Spark partners therefore brought a new narrative into play: he sees DGX Spark not so much as an over-engineered office PC, but rather as an additional device with a dedicated task – similar to a network-attached storage (NAS), which perhaps sits on the same desk as the work PC. This is why DGX Spark would not be delivered with a mouse and keyboard like other complete PCs. Does it click for you too? Exactly: A NAS doesn't need a monitor because you only access it via the network.

Bad news for Nvidia: GB10 is also to get closely related offshoots called N1X and N1, which will then be used in powerful gaming notebooks with Windows on ARM and take away Qualcomm's exclusivity. However, a functioning display controller would be needed here. The error mentioned above requires a step back to the drawing board and a revised chip design, which is likely to take at least six months. A delay of this magnitude for the N1X and N1 was reported to us by several sources without them knowing the exact background. We therefore believe that an official unveiling in this calendar year is now out of the question; however, it would fit in well with CES in early 2026.

For DGX Spark with GB10, all companies are still hoping for a faster launch (probably with the above restrictions). At Computex, it was sometimes said vaguely "in the third quarter", sometimes more specifically "we are planning for the end of August". However, the only promise that we trust 100 percent with years of Nvidia experience is "when Jensen allows it" – and yes, this is also a quote from a partner.

Almost notebook-compatible

If it weren't for the stupid bug, GB10 would already have everything that is relevant for notebooks in its current form. Nvidia confirmed integrated WLAN and Bluetooth, for example. We had to dig deeper into other aspects, as there is still no official block diagram that names all of GB10's components.

Without an NPU that is suitable for Copilot+, Nvidia would not even need to compete with high-priced Windows notebooks. Nvidia itself cleverly dodged the question with "we're not exposing an NPU on DGX Spark": Just because there is no programming interface, as DGX Spark pushes all AI calculations in the targeted market segment onto the much more powerful GPU, an NPU can still be slumbering in the chip.

We have been told by third parties that there is an NPU based on NVDLA. Nvidia has already used these in the past, for example in the Jetson Xavier processor. Alternatively, cooperation partner Mediatek could also have contributed an AI unit.

Mediatek already has plenty of experience with such computing units: Mediatek's smartphone processor Dimensity 9400(+) and its Chromebook offshoot Kompanio Ultra 910 are equipped with a 50 TOPS NPU. With one Cortex-X925, three Cortex-X4 and four Cortex-A720, the processor also has everything you need for everyday computing power in non-AI applications. The GPU (Mali Immortalis G925) is also one of the latest and most powerful that design supplier ARM currently has on offer.

Piquant side note: In the footnotes of the processor data sheet, Mediatek mentions that GPU features such as ray tracing cannot be used under ChromeOS. Well, we can immediately think of another operating system that has no problems with this and would also enjoy such a big NPU. The stand staff put a friendly, but firm stop to this in conversation: Kompanio Ultra is for Chromebooks, for Windows notebooks "only the one with Nvidia" is in the works for now.

Supply or get involved?

As is so often the case, external factors play a more important role than internal will. Mediatek does not develop the CPU and GPU architectures used in its chips itself, but takes what ARM provides to all its licensees. This means that the company is also dependent on their support when it comes to (graphics) drivers; and these are currently not available for Windows.

The emphasis in the last sentence must be on "currently", because as a birdie in Taiwan chirped to us, this will change: ARM is working on Windows drivers for the Mali GPUs. We are not aware of the current status, though. However, ARM's lack of an announcement to date explains Mediatek's current focus on Chromebooks.

2026 ahoy

With all this knowledge, the only question we were unable to answer with certainty during the trade fair is of a purely academic nature: it is unclear when exactly the exclusive agreement between Microsoft and Qualcomm on Windows on ARM will expire. The two companies have never publicly admitted it. However, it is an open secret that Windows on ARM is currently only allowed to run on devices with a Qualcomm processor. By 2026 at the latest, there will be more than just the three companies AMD, Intel and Qualcomm offering processors for notebooks.

(mue)