Discussion NV Re-Enter ARM PC market in 2025!

[poke01]

Well, Qualcomm should be shaking in their boots now.

It’s not Nvidia’s hardware they should be worried about but their software prowess.

 

[Ghostsonplanets]

QCOM problem is that they're at crossroads. Their iGP is extremely efficient and performant for the Mobile market. It's been custom tailored through the years for this market demands.

But now they have to scale up Adreno to Desktop Class graphics computation. And there, the direction GFX is going is totally different from Mobile.

https://twitter.com/x/status/1805126265277501577

Mobile is basically TBDR + small batch FP16 workloads, with careful shaders that don't go overload. Desktop is completely different, with focus on plenty of bandwidth, Compute Shaders and next-gen geometry.

QCOM will need to design a GPU arch that can still be extremely area and power efficient for mobile while also being able to serve Desktop current and future demands.

https://twitter.com/x/status/1805561685119492561

 

[Doug S]

Mobile is basically TBDR + small batch FP16 workloads, with careful shaders that don't go overload. Desktop is completely different, with focus on plenty of bandwidth, Compute Shaders and next-gen geometry.

So they have the same problem Apple does with its TBDR based GPU that doesn't perform/benchmark well in stuff (esp. games) designed for a traditional PC GPU pipeline.

 

[FlameTail]

So they have the same problem Apple does with its TBDR based GPU that doesn't perform/benchmark well in stuff (esp. games) designed for a traditional PC GPU pipeline.

So how will Qualcomm/Apple go about solving this ?

 

[FlameTail]

Switch 2 iGPU is 12 SMs with 12 RT Cores and 48 Tensor Cores, 1536 CUDA Cores/ALUs.

Nvidia iGP is supposed to be much bigger than this.

Radeon 780M is 768 ALUs/1536 with dual-issue. But much higher clocked. Faster than Switch 2 but should be slower than Nvidia WoA SoC.

RTX 4050M is 20 SMs.

 

[Doug S]

So how will Qualcomm/Apple go about solving this ?

Well Apple simply doesn't care, because they aren't concerned with running Windows stuff that requires high GPU performance. People will run native for productivity type tools, and they mostly don't game on Macs either.

Its only really a problem for Qualcomm.

 

[Shivansps]

I think all mobile gpus are like that, like the Mali G610MP4 was considered high end just 3 years ago. But in Linux running desktop apps it performs worse than a 2019 Vega 3 IGP.

Not sure what the problem is exactly, but i belive mobile just has other requierements, and these gpus are specially bad in FP ops. I belive it is because mobile gpus are still tailored for OpenGL ES.

 

[podspi]

Well Apple simply doesn't care, because they aren't concerned with running Windows stuff that requires high GPU performance. People will run native for productivity type tools, and they mostly don't game on Macs either.

Its only really a problem for Qualcomm.

I think people are being a bit myopic regarding GPU performance. It is important for some people in some workloads. Not everybody games, and with the NPU the GPU is prioritized for these non-graphic workloads as well. I am not sure who buys a T&L to play games locally, but GFN and Gamepass also work fine.

The thing that really interests me is how NV will be able to deliver on CPU performance and power efficiency. It will be interesting to see how differentiated the ARM SoCs from nVidia, MTK, (potentially) AMD actually are. If it's just the different GPUs, that will be disappointing. It is a shame (although not completely surprising) that Denver didn't work out. I love having a super-crowded and competitive CPU market again.

 

[FlameTail]

I wonder if Nvidia will funnel $$$ into ARM Cortex-X team, enabling them to bring large IPC increases every generation.

 

[soresu]

I wonder if Nvidia will funnel $$$ into ARM Cortex-X team, enabling them to bring large IPC increases every generation.

That's not how it works.

It's not a simple "more R&D money = better IPC" problem.

Best case scenario more money can allow you to have multiple competing µArch design teams working in parallel to get the best result.

 

[soresu]

I think people are being a bit myopic regarding GPU performance. It is important for some people in some workloads. Not everybody games

Not all GPU workloads are games these days.

 

[FlameTail]

Nvidia Windows On ARM SoC leak just dropped:

https://twitter.com/x/status/1809204562655346985

Cortex X925× 10
ARM Cortex A725× 4
32 MB L3 cache

GPU:
20SM-Blackwell architecture

NPU: Nvidia-DLA with 60+TOPS.

Memory: 128b-LPDDR5X @ 9600mhz

Only 60 TOPS for the NPU? I was expecting more, something that meets the bar for the second gen Copilot+ PC requirement.

Maybe that's for FP16. So INT8 is 120 TOPS.

 

[Thibsie]

If that's NPU only, it's fine. GPU can help there too.

 

[FlameTail]

Rumours abound that Nvidia's SoC is releasing in 2025H1.

It supposedly consists of two chiplets: A CPU tile and GPU tile.

CPU tile is designed with help from Mediatek. GPU tile is designed by Nvidia.

Both tiles are connected together via advanced packaging.

 

[igor_kavinski]

CPU tile is designed with help from Mediatek. GPU tile is designed by Nvidia.

They are screwed then. All my MediaTek phones randomly develop freezing issues.

 

[DAPUNISHER]

Not certain if this is news to you nerds -

https://www.reddit.com/r/pcmasterrace/comments/1ew9nnb/nvidia_makes_a_cpu_now_and_i_benchmarked_it

 

[FlameTail]

Nvidia ARM SoC will be much more desirable than Snapdragon SoCs for gamers and professionals alike, thanks to:

Nvidia GPU Architecture + Nvidia GPU Drivers

 

[soresu]

Nvidia ARM SoC will be much more desirable than Snapdragon SoCs for gamers and professionals alike, thanks to:

Nvidia GPU Architecture + Nvidia GPU Drivers

Until games start coming with native ports as standard I wouldn't get too excited on that score no matter whose SoC it is.

As for pros the majority of creative professionals seem to gravitate towards Macs anyway, where most of the DCC software packages have since been ported to ARM Mac, so I don't expect it to be any great driver for WoA.

IMHO the greatest benefit for WoA will be the casual use market for browsing and general office use.

Once cheaper SoC SKUs from Mediatek are out on the market it will boost market share a lot.

 

[FlameTail]

It's all about applications and gaming (more so) as Qualcomm is finding out with the bad press about that. There is a long ways to go with WoA, but imo once the most widely used apps are there, it’ll pick up momentum. It’ll be interesting to see how Nvidia can help out there, if/when they jump into the fore.

I am calling it now. Improving gaming will be one of the best things Nvidia ARM SoC will bring to Windows-on-ARM.

 

[Ghostsonplanets]

I am calling it now. Improving gaming will be one of the best things Nvidia ARM SoC will bring to Windows-on-ARM.

But will it benefit WoA as a whole? Or it will just be a side-effect of relying on a mature and actively supported GPU IP ecosystem?

 

[igor_kavinski]

Improving gaming

If by that, you mean making sure games don't run as well on other SoCs...

 

[FlameTail]

Rumour: Nvidia's ARM SoC to release in 2025 September.

Nvidia's Arm-based PC chips for consumers to launch in September 2025, commercial to follow in 2026: Report

Nvidia has ambitious plans for the PC market.

www.tomshardware.com

Looking forward to this one. For a product that's less than one year away, there have been any credible leaks about it's specs.

 

[FlameTail]

September 2025 release suggests that it is using this year's Cortex X925 core. So we can expect a Geekbench ST of 3000-3300, varying based on how it's clocked and the caches are configured.

There's also a possibility that it could be using a custom Nvidia ARM core, but that would be coming out if nowhere and a total wild card.

 

[igor_kavinski]

There's also a possibility that it could be using a custom Nvidia ARM core, but that would be coming out if nowhere and a total wild card.

Gotta have a GB6 ST score of 4100 or else it's a fail.

 

[jdubs03]

Gotta have a GB6 ST score of 4100 or else it's a fail.

Gotta go big or go home.