Steam Deck - SteamOS, Zen2 4C/8T, RDNA2 1.0-1.6 TF, 16 GB LPDDR5 88 GB/s, starting at $399 [2021-12]

Bondrewd said:
Those things are more Intel GNA and less programmable ML blob and all.
Click to expand...

Well Intel claims the Ice Lake and Tiger Lake GNA blocks can indeed do image upscaling, but on real time applications they only mention voice/speech recognition (it's made for low-power operation, not performance). The throughput of Van Gogh's CMVL block should be significantly higher than Intel's GNA, though.
 
Bondrewd said:
Yeah but it's still made for the exact same workloads as Intel GNA which is always-on Windows Hello-ish garbage.
Click to expand...
Why so much throughput just for that, though? Are they planning always-on video processing for Windows Hello?
 
Malo said:
What does this have to do with Steam Deck?
Click to expand...

1 - Steam Deck's Van Gogh SoC is a chip that was originally designed for a Microsoft device according to several sources (surface xbox? xbox portable? we don't know as Panos Panay cancelled it in late 2020 / early 2021 apparently)


2 - Van Gogh (again: originally developed with/for Microsoft) has a CVML - Computer Vision and Machine Learning block that has dedicated matrix processors with a combined throughput of ~1.5 TMACs FP16 and 4 TMACs INT8 (8 TOPs) - if clocked at 1.6GHz.

https://twitter.com/i/web/status/1316838798823690241


3 - Microsoft has been developing Machine Learning upscaling algorythms. These could be only for the Series S/X's INT4/INT8 RPM capabilities in the shader processors. However, if they requested AMD to put a dedicated ML inference coprocessor in the SoC of their handheld gaming device, then it could be to accelerate a ML upscaling technique using lower power and without taking compute resources from the iGPU shader processors.


4 - I'm fully aware this is 90% wishful thinking on my part.
 
ToTTenTranz said:
1 - Steam Deck's Van Gogh SoC is a chip that was originally designed for a Microsoft device according to several sources (surface xbox? xbox portable? we don't know as Panos Panay cancelled it in late 2020 / early 2021 apparently)


2 - Van Gogh (again: originally developed with/for Microsoft) has a CVML - Computer Vision and Machine Learning block that has dedicated matrix processors with a combined throughput of ~1.5 TMACs FP16 and 4 TMACs INT8 (8 TOPs) - if clocked at 1.6GHz.

https://twitter.com/i/web/status/1316838798823690241


3 - Microsoft has been developing Machine Learning upscaling algorythms. These could be only for the Series S/X's INT4/INT8 RPM capabilities in the shader processors. However, if they requested AMD to put a dedicated ML inference coprocessor in the SoC of their handheld gaming device, then it could be to accelerate a ML upscaling technique using lower power and without taking compute resources from the iGPU shader processors.


4 - I'm fully aware this is 90% wishful thinking on my part.
Click to expand...

steam deck use vangogh right? then that's good news, valve could add an ML upscaling for steam deck for when its docked.
 
Bondrewd said:
Beats me.

I dunno what but MS wants ML garbage on die and a lot of it.
Click to expand...

Microsoft made a big deal about the additions to the arm cpu they used for the surface pro x. They showed off it doing fake eyes for people while on video so eye contact is maintained. So who knows what else they are working on.

I guess the real question is what can the ML in the chip can be repurposed to do in the steam deck.
 
eastmen said:
Microsoft made a big deal about the additions to the arm cpu they used for the surface pro x
Click to expand...
They didn't, it just used the big standard 8cx Hexagon.
eastmen said:
So who knows what else they are working on.
Click to expand...
ML is a priori gimmicky in client so obviously something useless.
eastmen said:
guess the real question is what can the ML in the chip can be repurposed to do in the steam deck
Click to expand...
Not much.
Gonna be a pile of dark silicon there just like Pluton.
 
Bondrewd said:
They didn't, it just used the big standard 8cx Hexagon.

ML is a priori gimmicky in client so obviously something useless.

Not much.
Gonna be a pile of dark silicon there just like Pluton.
Click to expand...


I dunno just going by what they said

https://blogs.windows.com/devices/2...ion-with-eye-contact-now-generally-available/

Using the dedicated AI silicon in the Microsoft SQ1 (more on that below) in Surface Pro X along with the 5.0MP front-facing camera that captures 1080p full HD video
Click to expand...
 
Just got my e-mail mentioning the delay too.
Nooooooo!

However my expected arrival is still Q1 2022.
I guess they were expecting to ship my Deck in January but it's now pushed to March.
 
You must log in or register to reply here.
Back
Top