Behind the Silicon: Snapdragon X

[InkeyZero]

Report/Rumor about NVIDIA WoA SoC:
- Architecture: Blackwell, 180-200 TOPS
- Launch/Announcement: Computex 2025
- Collaborate with: MediaTek
- Lineup: N1x (High-end), N1 (Mid-range (?!)
- Production Estimates: Q4 2025: 3 million units
FY 2026: 13 million units https://t.co/zezO8XZRHU

- AI supply chain tracking report:
This week’s theme revolves around the upcoming CES show, especially AI PC. Since Nvidia has switched to the new generation Blackwell architecture, it can reach 180-200 TOPS, which is significantly better than the current highest of 50 TOPS. For this reason, major American companies saw that KYEC had temporarily cut back on testing time to meet Balckwell demand, but Blackwell output doubled this season compared to the previous quarter.
On the other hand, Nvidia is also actively deploying Windows on Arm (WoA). Brokerages are expected to choose to debut the latest chips at this CES show, and related products will be launched at Computex in Taipei in late May. Since this project is in cooperation with MediaTek, the brokerage first expects a volume of 3mn in 4Q this year. The high-end N1X will be launched first, followed by N1. The volume next year is expected to be 13mn. In other words, this project is estimated to contribute US$2 billion to MediaTek's revenue this year. , it can account for 8% of revenue next year.

— Hoang Anh Phu (@AnhPhuH) January 13, 2025

[InkeyZero]

Exclusive: Tech supplier Arm plans to hike prices, has considered developing its own chips

• Arm considering price increase up to 300% and making its own chips
• Arm's strategy aims for $1-billion annual revenue increase over 10 years
• Arm's potential chip-making could threaten its own customers, analyst says

Si los fabcricantes que usan teconologías basadas en Arm ven esos incrementos como imposiciones unilaterales divorciadas de la realidad del mercado, ese tipo de decisiones de Arm podrían empujar a sus mayores clientes a invertir en el desarrollo y adopción de productos basados en RISC-V.

[InkeyZero]

Nuievo artóculo sobre la disputa Arm vs QUalcomm de una serie, hasta ahora de 3 partes:




Nueva entrega.

Qualcomm looks to be off the hook until 2033.

As details emerge following the end of the trial in December, documents confirm that Qualcomm does have a license for V9 meaning that it is in a stronger position than I had previously anticipated.
The jury in the trial decided that Qualcomm is not in breach of the Nuvia contract with Arm and that Qualcomm’s Orion processors are properly licenced under Qualcomm’s existing architecture licence with Arm.
However, the jury was deadlocked on the question of whether Nuvia breached its licence with Arm, and it is on this basis that Arm could seek a retrial.
Arm has said that it is seeking a retrial but it is not clear whether it is seeking to relitigate all of the jury findings (i.e. appeal) or just the one where the jury was unable to come to a verdict.
My view at the time (see here) was that this was a victory for Qualcomm but because it did not have an architecture licence for the latest instruction set (v9), it would have to go back to Arm before the expiry of its current agreement which for all practical purposes is in 2033.
The contract expires in 2028 but Qualcomm has the option to extend this until 2033 which as things stand today, I am certain that it will exercise.
Documents that have subsequently been made available are pretty clear stating that Qualcomm and Arm added an architecture licence for v9 on June 23, 2020.
Assuming that the original two verdicts are not overturned on appeal, this means that Qualcomm can migrate its Orion custom CPUs to v9 without having to enter into any further negotiations or agreement with Arm.
Hence, Qualcomm should be able to continue to ship all of its arm-based products without having to enter into any negotiations with Arm for another 8 years.
However, there is the issue of v10 which is the next version of the Arm instruction set which I expect will be released by Arm sometime around 2028 or 2029 based on Arm’s history.
If there are new features in v10 that necessitate a quick move to v10 in order to remain competitive, then Qualcomm will have to take a v10 license before 2033.
I do not think that this will be the case as there are areas of overlap between the last versions of one instruction set and the early versions of the next one.
This is how I expect Qualcomm has been able to come up with very competitive products using v8 even though v9 has been around for more than 5 years.
If Qualcomm does need v10 before 2033 then it will be in a difficult position because Arm’s design is not an official standard and so it does not have to adhere to the fair, reasonable and non-discriminatory (FRAND) principle that governs the licensing of standard essential patents.
Hence, Arm can be as unreasonable as it wants when it negotiates with Qualcomm which will put Qualcomm in a difficult position in 2033 or earlier if it turns out it needs v10 before that time.
Arm says it intends to seek another trial, but I continue to think that this is not in the best interest of the Arm ecosystem.
Arm and its customers should be working together to expand its reach into PCs, vehicles, servers, AI and so on and not wasting time and resources fighting each other.
Eight years is a long time and so I am hopeful that there will be a negotiation at some point where everyone gets something, the two companies kiss and make up and get back to the business of driving the Arm ecosystem.
I continue to think that another 8 years of frozen dialogue and no cooperation will only benefit Arm’s rivals such as RISC-V and x86.
This is crucial as the standards for AI are beginning to form and with Intel on its knees, the time to push the Arm ecosystem hard is now.

[InkeyZero]

If you were anticipating a swift conclusion to the ongoing legal battle between Qualcomm and Arm, a prominent CPU architecture company, you might be disappointed.

A recent mistrial in December temporarily halted proceedings, but the situation has become more intricate due to Qualcomm’s counterclaim. This countersuit accuses Arm of engaging in practices aimed at suppressing competition and hindering technological advancements.

As it stands, we may not witness any resolution until 2026, with a new trial scheduled to commence in February of that year.

Understanding the Counterclaim
Recent court documents reveal that Qualcomm’s counterclaim against Arm is set for trial around March or April 2026. This development suggests that the saga involving these chipmakers could extend for at least another year.

Qualcomm’s allegations indicate that Arm perceives its licensees as rivals and is attempting to push businesses toward using its pre-configured CPU cores instead of allowing them to innovate freely.

In their filings, Qualcomm argues: “Arm now views Qualcomm as an impediment to its goal of increasing prices and eliminating alternatives for customers… Faced with an inability to compete on equal footing with Qualcomm, Arm has resorted to various wrongful tactics aimed at undermining Qualcomm’s advancements in CPU design.”

Seguir leyendo en: https://therightopinion.in/technology/u ... e_vignette

[InkeyZero]

In 2025, the plan is to improve Arm’s own cores, Bergey said. And the first goal is simply to run them faster.

“We think that we are reaching, we’ve reached kind of IPC leadership, and now people are getting very aggressive on frequency, so we’re going to continue to really push there,” Bergey said.

“We’re leading on IPC on some of the products in the market,” Bergey said. “But we’re clocking at a lower frequency than some of those products. And so what I’m just suggesting is — you know, IPC times frequency, right, gets you to [higher] performance. We want to continue to provide the highest performance Arm cores, so we’ll continue to make those investments.”

Bergey said that Arm’s second priority is to accelerate AI workloads on its own designs, specifically on the CPU and GPU. On the CPU, that entails specific instruction capabilities that Arm is adding to the CPUs, progressing past Neon, its Scalable Vector Extensions (SVE), and 2021’s SVE2. These additional extensions will build off of SVE2 to accelerate some of these AI workloads, Bergey said.

Arm’s litigation: It ain’t over ’til it’s over
Arm normally enjoys solid relationships with its licensing partners — save for Qualcomm, and an ongoing lawsuit that has simmered since 2022. Last October, that suit boiled over after Arm cancelled Qualcomm’s architectural licensing agreement. But when the suit reached court, a district judge found in favor of Qualcomm in two of the three issues, including that Qualcomm proved that the CPUs acquired via Nuvia are covered by its architectural license, and that Qualcomm did not breach the terms of the Nuvia license it acquired.

However, the jury could not come to a conclusion over whether Nuvia itself had breached the terms of its architectural license. According to Bergey, this leaves the case between the two companies “unresolved.” “It’s still an open issue that needs to be resolved between the two parties,” he said. He declined to comment further.

Qualcomm, for its part, was undeterred. “We’ve made a public statement that we are happy with the outcome of the case, and [the court] upheld that we have the right to innovate and to the technology that we are bringing, the disruption that we are creating in the marketplace,” said Nitin Kumar, senior director of product management, at CES last week.

Contenido más amplio que incluye otros tópicos en: https://www.pcworld.com/article/2578668 ... power.html

[InkeyZero]

[InkeyZero]

Los Snapdragon X2 Elite van a soportar tarjetas gráficas externas y hay modelos con 18 núcleos.

interesante.

[InkeyZero]

Esto es otra cosa, parte de la tendencia de equipos optimizados para AI + hardware SFF.

[InkeyZero]

Tristemente, aún no tenemos equipos con Snapdragon X Elite con soporte completo bajo Linux.

Según lo que leí hace ya algún tiempo, la segunda generación (¿Snapdragon X2 elite?) contaría con soporte Linux de manera temprana, pero lo de la primera generación es ya patético.

[InkeyZero]

[InkeyZero]

Esto es otra cosa.

[InkeyZero]

[InkeyZero]

Mientras tanto, de aquí a una década ¿bismuto en lugar de silicio?

Una de las apuestas chinas.

[InkeyZero]

comentario de un usuario:

hace 1 hora (editado)
Let me tell you my analysis on this:
1. I'd blame Microsoft as much as Qualcomm for this. Microsoft diluted this entire narrative by pushing "CoPilot+" rather than "ARM" narrative.

2. Microsoft should've left Qualcomm to say what they want to but not how Microsoft wanted it to.

3. The narrative of "Better battery life, less heat & fan noise with improving application support" got shifted to "Great AI that runs on-device. CoPilot+ is brand. Buy these CoPilot+ PCs. These will do magic" which is totally a BS narrative to push onto consumers faces.

4. Qualcomm confused people with multiple versions of processors at different price points which are almost the same instead of pushing one variant that works as a base for large amount of consumers.

5. Most developers weren't interested to use it because of lack of Linux support and Qualcomm never took any action about Linux support. It only partnered with Microsoft that made Windows still keeping that same first time crap boot-up experience.

6. There were driver issues present with the devices which neither Microsoft nor Qualcomm took care of before launching them. It led to consumers returning laptops with too many glitches.

You can add any points below if you noticed any other.
Also, nice video Josh. I'd like to see the video on Nvidia as well.

Qualcomm cometió un grave error: puso la carreta delante de los bueyes.

Su énfasis debió estar en proveer a los desarrolladores con PCs equipados con sus nuevos chips para darles tiempo a desarrollar las aplicaciones sobre arquitectura ARM antes del lanzamiento, sobre todo si se trata de software crítico, para de esa manera minimizar el problema de incompatibilidades el que muchos usuarios se quejaron.

Esto fue tan grave que el minipc que ofrecieron para este propósito fue anunciado después del lanzamiento de los equipos para usuario final, y el devkit fue retirado del mercado antes de cumplir con los envíos que se vieron retrasados vez tras vez.

Apenas un puñado de desarrolladores recibieron los suyos, y para colmo, fueron informados que el producto no recibiría soporte en lo sucesivo. Al menos le devolvieron el dinero a todos los que lo compraron, incluyendo los que lo recibieron físicamente.


Si le sumamos los problemas legales con ARM que mantuvo a la industria en ascuas, y la insufrible manía de meter AI en todo, a pesar del disgusto del usuario final, ya vemos que la cosa no fue bien.


Y para lo parlanchines que se pusieron con el lanzamiento de la primera generación, están muy calladitos ahora que deberiamos estar cercanos al anuncio de la siguiente.

[InkeyZero]

Interesante perspectiva de un usuario en los comentarios de este video:

@aaronb1138
hace 2 semanas (editado)
The main issue with ARM is the lack of a straightforward ISA. In x86 land I can take a system disk out of one x86 box with an i7 and plug it into a Ryzen CPU machine with completely different underlying hardware and everything baked into that OS and the software on top just works once the kernel loads the correct drivers. ARM is a forced march of obsolescence due to ISA inconsistency where there are portability problems within the same generation of hardware with the same SoC. Last week I was speaking with a hardware vendor in the 3D printing space that their specialty display part (mSLA) will work with a RPi CM4, but not other Pi4 psuedo-clone hardware and isn't even forwards compatible "today" with RPi CM5. This btw, is why you see CM4s priced close to CM5 hardware -- the CM4 took off well enough and is used enough places where a CM5 will not simply drop in as an upgrade. The development cycle starts very nearly from scratch.

The embedded systems are awful for long term support if you want a product to develop and evolve. While this doesn't matter for an HVAC system, if say we're talking about automative with driver assistance features, this means issues with base OS updating -- Linux, QNX, and whatnot -- validation among board revisions, much less generation to generation compute platforms is a complete mess.

I won't even get into my day job, but suffice it to say, I deal with specialized hardware currently moving TO x86 to get away from generation-to-generation incompatibility and feature change in other CPU architectures.

Most of the excitement I have seen in the software development community is about making software development work more ephemeral. They can label it as avoiding x86 "legacy" and "cruft", but it feels very "new shiny" and the real appeal of ARM platforms is *zero commitment.* A year later, they wave their hands and say, "obsolete" and walk away from a platform. We see this across a wide variety of consumer products well outside of the easy target in modern phones with limited support lifecycles.

Every piece of TV, AV, IoT running ARM similarly is incredibly short-lived, and unlike x86, can't be opened up and extended or repurposed. In the embedded space with ARM + Linux, we see devices with let's say USB camera "capability", but inconsistent compatibility where it has a stricter approved hardware list than the NT 4 HCL.

@mattbland2380
hace 9 días
Absolutely the case, but I can see the situation improving over the next few years as adoption increases and hopefully the market will address the issue.

@aaronb1138
hace 9 días
@mattbland2380 That was the promise of the RPi 1 as a generalized ARM compute platform with a fixed yet highly I/O flexible layout. That was 13 years ago. Well, we can call it 10 with the B+ model which better finalized the mechanical layout. Still quite the forwards + backwards compatibility mess. Software is ok-ish within that ecosystem, but for general I/O and hardware control being a key factor, it's a bit rough. Back with the first RPi, my general critique was, "Wouldn't I be equally served with some form of USB GPIO breakout device which plugs into my PC? If I need small, quiet, and fanless for an appliance, wouldn't a NUC work better?"

ARM for their part has made it quite clear they will require new licensing agreements and fees for anything new, useful, or unifying. ARM licensing will break any future for an ARM unified and consistent ISA. Rent seeking will be the undoing.

[InkeyZero]

https://wccftech.com/snapdragon-x2-elit ... -64gb-ram/

Snapdragon X2 Elite With A Whopping 18-Core CPU Said To Be Tested With 64GB RAM For The Top-End Configuration, But There Is No Word On The Official Launch

https://www.notebookcheck.org/Snapdrago ... 037.0.html

Snapdragon X2 Elite: Un filtrador revela los planes de Qualcomm para las opciones de 64 GB de RAM antes de lo previsto

No se espera que Qualcomm presente sus próximos chipsets para portátiles hasta dentro de unos meses. No obstante, Roland Quandt ha aparecido con más información sobre el llamado Snapdragon X2 Elite (SC8480XP), que podría superar al Snapdragon X Plus y al Snapdragon X Elite con 18 núcleos de CPU y hasta 64 GB de RAM.

[InkeyZero]

[InkeyZero]

[InkeyZero]

[InkeyZero]

NVIDIA’s first Arm-based PC chip's Performance Is Set to Outpace Qualcomm’s Snapdragon X Elite

NVIDIA’s upcoming AI-PC chip—code-named “N1X”—carries a 20-thread CPU running at 2.81 GHz. Because Arm CPUs generally lack simultaneous multi-threading (SMT) like Intel’s Lion Cove, the processor is presumed to have 20 physical Arm cores, likely a modified version of the Grace CPU found in the earlier GB10 superchip.

NVIDIA’s personal AI “super-computer,” Project DIGITS (DGX Spark), already uses the GB10 Grace-Blackwell superchip. That device pairs a Grace CPU (10 × Cortex-X925 + 10 × Cortex-A725 cores) co-designed with MediaTek and a Blackwell GPU.

Geekbench, however, identifies the N1X CPU as Armv8, whereas the GB10’s 20-core CPU is Armv9. This raises the question of whether, like Qualcomm, NVIDIA has licensed Armv8 to build a custom core architecture. (Geekbench has previously mis-reported Armv9 silicon as Armv8.)

System data show the N1X mounted on an HP development board (model HP 83A3) running Ubuntu 24.04.1. The board appears to have 128 GB of system memory, with 8 GB reserved for the GPU.

In Geekbench, N1X scores 3,096 single-core and 18,837 multi-core. Even accounting for test conditions, that outstrips Intel Arrow Lake-HX, AMD Ryzen AI MAX (Strix Halo), and Qualcomm’s Snapdragon X Elite. Against top-end chips such as Intel Core Ultra 9 285HX and AMD Ryzen AI Max + 395, N1X’s CPU performance is only slightly behind; it still trails Apple’s M4 MAX.

Details on the GPU, NPU, and other compute blocks remain unknown. If N1X is indeed a scaled-down GB10, it likely uses a Blackwell-based GPU, though power limits for mobile form factors would preclude the GB10’s full 6,144-CUDA-core configuration.

Despite the strong preliminary numbers, N1X is not expected to launch until 2026. By then, Qualcomm will have moved to its Snapdragon X2 series with Oryon V3 cores, which could narrow—or even reverse—the performance gap.

N1X is also believed to be co-designed by NVIDIA and MediaTek, combining MediaTek’s low-power, high-performance CPU expertise with NVIDIA’s strengths in GPUs and AI. This partnership could tempt both traditional PC makers and notebook-playing smartphone brands to adopt the chip.

A previous Verge report suggested Dell’s Alienware might debut the first gaming laptop with N1X. The current leak, based on an HP development platform, likewise hints that HP and other PC OEMs may roll out products built on this silicon.