Discussion ARM Cortex/Neoverse IP + SoCs (no custom cores) Discussion

[mvprod123]

And it's out... heck, the X4 is interesting. But again Google has to choose an inferior GPU this time.

Why Google who can choose something good, makes the screwup to make an awul desicion?

Meanwhile... no announcements of future cores.

There is a suspicion that a driver supporting Vulkan 1.4 does not yet exist or is still in development. The Imagination website states that support for DXTP-48-1536 is only planned for 2025.

 

[DZero]

Imagination DXTP-64-2048 GPU - Imagination

www.imaginationtech.com

Seeing this.... this might be the next GPU from Google after Tensor G5:

 

[soresu]

Imagination DXTP-64-2048 GPU - Imagination

www.imaginationtech.com

Seeing this.... this might be the next GPU from Google after Tensor G5:
View attachment 129237

They must really be impressed with IMG TEC's numbers if they are that set on shifting away from Mali when it's finally doing well against Adreno.

As to "OGL 4.6 via Zink" I'll believe it when I see it - you need a fair chunk of VK extensions to make Zink work somewhat performantly.

Given they aren't even advertising VK 1.4 support nearly 8 months into the year they promised to deliver it I have my doubts.

 

[soresu]

heck, the X4 is interesting

2+ years old at this point tho.

Meanwhile... no announcements of future cores.

There is an "ARM unlocked" event on September 10th, and an AI Infrastructure event on Sept 9th-11th, either of which might be the IP announcement bonanza.

I'd be very surprised if they leave it any longer and risk Qualcomm announcing Oryon3 first.

 

[DZero]

They must really be impressed with IMG TEC's numbers if they are that set on shifting away from Mali when it's finally doing well against Adreno.

As to "OGL 4.6 via Zink" I'll believe it when I see it - you need a fair chunk of VK extensions to make Zink work somewhat performantly.

Given they aren't even advertising VK 1.4 support nearly 8 months into the year they promised to deliver it I have my doubts.

Unless Google is paying a LOT to both GPUs and allows to Imagination to update their drivers.

Mali might step up now since there will be support with Linux in a future and that could open the way to see a driver for Windows. Something Imagination lacks

 

[soresu]

Mali might step up now since there will be support with Linux in a future

PanVK already has 1.4 compliance for Mali v10+ (v10 = G710) according to mesamatrix.

API standard compliance =/= optimisation though, so I wouldn't expect great performance for the time being.

The better VK driver efforts seem to have their own special compiler like ACO for RADV and NAK for NVK, so hopefully we might see a similar effort for PanVK and/or Panfrost in the future.

Panfrost OGL support is much patchier, especially compared to Asahi (Apple M1/2 GPU) oddly enough.

It doesn't even yet have full OGL 3.2 compliance (geometry shader support still lacking), though it does have compliance for all 3.3 extensions, and for 4.0+:

4.0 = 48%%, 4.1 = 50%, 4.2 = 83%, 4.3 = 91%, 4.4 = 93%, 4.5 = 58% and 4.6 = 36%.

If the Panfrost effort had put as much time in as for Asahi we might already have 4.6 compliance, but as it is sadly not.

Zink will give some relief for OGL problems, if not very performantly for the time being until PanVK gets more optimisation TLC.

There is another driver effort called Tyr underway built on Rust language code, but I think that is still pretty early in its development.

 

[MS_AT]

Seems they have increased the supported mem speed from earlier announcements for GB10 https://www.servethehome.com/nvidia-outlines-gb10-soc-architecture-at-hot-chips-2025/ and they have confirmed if anyone had doubts that MediaTek is doing the CPU chiplet.

Here if anyone would be interested in Grace Superchip mem subsystem analysis.