MediaTek and Nvidia cooperating on a high-end Arm PC chip

[Artemis]

https://www.extremetech.com/computing/nvidia-and-mediatek-are-reportedly-co-designing-snapdragon-x-elite-competitor

https://x.com/dnystedt/status/1789825653501833463?s=46

CoWoS part isn’t true, it’s bad translation. It’s most likely an SoC. MediaTek using stock Arm (probably Cortex X5 that is coming) and some of their IP for other blocks, and then obviously Nvidia does the GPU IP.

Goal apparently is to tape out out in Q4 this year on TSMC N3, production in H1/early 2025. Built with efficiency and AI/gaming in mind per the article, cost ~ $300 (read: premium or probably the same price as the very top Snapdragon SKU, but they could bin some of it).

 

[Artemis]

This is different from previous rumors because MediaTek and Nvidia were rumored to do this last year, and then we learned it was just automotive and with chiplet packaging for Nvidia GPUs on MediaTek SoCs, nothing to do with collaborative PC SoCs.

Then we heard Nvidia doing their own SoC on Intel 3 for Windows without anyone else.

But this one suggests Nvidia is collaborating and letting them use GPU IP on the die basically, and that it’s for the PC market. Maybe they’ll do both.

Most dynamic time for laptop PCs arguably ever, honestly.

 

[dada_dave]

This is something I might get to replace my aging CUDA workstation.

 

[dada_dave]

AMD also reportedly working on an ARM chip and rumors are that RDNA4 will finally get proper ray tracing to actually be competitive. The article below also talks about RDNA5 but that’s far enough away I don’t pay much attention to rumors that far out:

AMD RDNA 5 a "Clean Sheet" Graphics Architecture, RDNA 4 Merely Corrects a Bug Over RDNA 3

AMD's future RDNA 5 graphics architecture will bear a "clean sheet" design, and may probably not even have the RDNA branding, says WJM47196, a source of AMD leaks on ChipHell. Two generations ahead of the current RDNA 3 architecture powering the Radeon RX 7000 series discrete GPUs, RDNA 5 could...

www.techpowerup.com

 

[Artemis]

AMD also reportedly working on an ARM chip and rumors are that RDNA4 will finally get proper ray tracing to actually be competitive. The article below also talks about RDNA5 but that’s far enough away I don’t pay much attention to rumors that far out:

AMD RDNA 5 a "Clean Sheet" Graphics Architecture, RDNA 4 Merely Corrects a Bug Over RDNA 3

AMD's future RDNA 5 graphics architecture will bear a "clean sheet" design, and may probably not even have the RDNA branding, says WJM47196, a source of AMD leaks on ChipHell. Two generations ahead of the current RDNA 3 architecture powering the Radeon RX 7000 series discrete GPUs, RDNA 5 could...

www.techpowerup.com

Nice. Overdue. That and general matrix HW in the GPU

 

[Cmaier]

AMD also reportedly working on an ARM chip and rumors are that RDNA4 will finally get proper ray tracing to actually be competitive. The article below also talks about RDNA5 but that’s far enough away I don’t pay much attention to rumors that far out:

AMD RDNA 5 a "Clean Sheet" Graphics Architecture, RDNA 4 Merely Corrects a Bug Over RDNA 3

AMD's future RDNA 5 graphics architecture will bear a "clean sheet" design, and may probably not even have the RDNA branding, says WJM47196, a source of AMD leaks on ChipHell. Two generations ahead of the current RDNA 3 architecture powering the Radeon RX 7000 series discrete GPUs, RDNA 5 could...

www.techpowerup.com

LOL. They were talking about an Arm chip all the way back around the time i left! I guess they weren’t in a rush.

 

[Artemis]

I really like AMD, they’re much better than Intel at actual engineering when they put their minds to it these days. More efficient about it in particular

But it would be cool if they put more effort into A) designing lower power parts B) their AI stack and gaming upscaling

Probably a lot of this comes from having limited resources, it just takes time coming back from the dead

 

[dada_dave]

I really like AMD, they’re much better than Intel at actual engineering when they put their minds to it these days. More efficient about it in particular

But it would be cool if they put more effort into A) designing lower power parts B) their AI stack and gaming upscaling

Probably a lot of this comes from having limited resources, it just takes time coming back from the dead

Yeah a major problem on the GPU front for AMD is their software stack. Nvidia built a huge lead in software. I mean hardware too, with specialized cores, but the combination is hard to compete against.

 

[KingOfPain]

LOL. They were talking about an Arm chip all the way back around the time i left! I guess they weren’t in a rush.

According to Wikipedia the K12 was announced almost exactly 10 years ago, May 2014 for a 2016 or 2017 release:

AMD K12 - Wikipedia

en.m.wikipedia.org

But AMD are still much better than some book publishers. I have a smallish dictionary for Middle High German. In the back of which is an advertisement for the index for a much larger dictionary for the same language. It only took them roughly over 100 years to release the index...

 

[dada_dave]

This seems about right for its performance since it is the same chip as the Linux mini-workstations released last year.

Nvidia N1X far behind Apple M5 Max: comparable chips show expected performance

The Nvidia RTX Spark is intended to compete directly with the Apple M5 Max, as the 80-watt ARM chip promises extreme performance with 20 processor cores and a "1 petaflop GPU". Benchmarks show that Nvidia's marketing should be viewed with caution.

www.notebookcheck.net

Of course there will be a smaller variants I believe and many things weren't tested since it was a Linux build meant exclusively for AI, while they are emphasizing its AI role for the Windows machines too, more variants with smaller RAM pools are going to be available as well. One small difference of opinion with the above link: the main Geekbench website also reports its compute score where it is pretty close to the 5070ti mobile - as such the largest difference with 5070 ti mobile isn't the TDP (different nodes), but the difference in bandwidth. While, I wouldn't be shocked if the gaming performance was equal to or better than the M5 Max due to Nvidia's drivers and most games being optimized for the Nvidia GPU, I suspect in bandwidth heavy loads (including inference), the GB10 will be constrained (especially relative to its potential compute performance). So for some 3D rendering and inference it'll be closer to M5 Pro depending on how much bandwidth matters - possibly for gaming in 4K resolution. That said, TDP will matter for some of the thinner laptops that have been introduced which simply won't be able to cool the GB10.