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

Uh-oh they have the same amount of L3 actually.
(AMD L3 is NOT unified at all).
I meant less L3 than the 4500U which has 2×CCXs. I know they're not unified, but there's more L3 per core which results in less travels to the RAM.


You severely overestimate the amount of b/w any RDNA2 part needs.

Perhaps I do, but I'm pretty sure RDNA2's optimal balance without Infinity Cache is above 40GB/s per TFLOPs because that's how much there is on the RDNA2 consoles, and Van Gogh's iGPU has less than half of that..
 
Screenshot_20210719-143216-500.png

128b actually so 2 QDPs running full channels.
So 2WGP disabled, leaving 4WGP active
Oh noes RMB has full 6WGP configs for Ryzen 9 alive too.
That being said I don't think Valve had input into Van Gogh's design
Nope, it's a premium FF APU and not a handheld material.
 
I meant less L3 than the 4500U which has 2×CCXs.
Same L3 factually.
They're non-additive exclusive chunks.
Leave the CCX and you're going to the IMCs either way.
Perhaps I do, but I'm pretty sure RDNA2's optimal balance without Infinity Cache is above 40GB/s per TFLOPs because that's how much there is on the RDNA2 consoles, and Van Gogh's iGPU has less than half of that..
Good news it's 128b.
Mero's the one with 64b then.
Perhaps I do, but I'm pretty sure RDNA2's optimal balance without Infinity Cache is above 40GB/s per TFLOPs
Nah you can skewer it around by going wider on SAs and beefing up the L2 relative to CU count which is what RMB does.
 
Wow, I just watched the Gamer's Nexus video and the GPD Win3 and the Aya Neo seemed like the only ones close in stats...but they're both running at over a grand?!?

Why would you get one of those instead of a gaming laptop? I don't get it.

If the $400us model has an open M.2 slot I really think it could tempt me, but by the time its available in December who knows how the PC hardware price scene will be?

I'm not gonna pre-order it, but damn seeing the competition its exciting me a lot more now!
 
Hmmm, 128GB M.2 2230s are pretty easily available on Amazon for under $20, the 512GB is like $249. :| (Prices US, I'm hopelessly American. My apologies.)

I could handle a $400 with a 128GB drive for a bit, depending on how hard they make it to swap.
 
Because it has a sizeable CVML block that won't ever be touched here
This got me thinking.. if these are ML inference accelerators, couldn't they be used to accelerate Microsoft's ML upscaler without resorting to the GPU ALUs?
 
This got me thinking.. if these are ML inference accelerators, couldn't they be used to accelerate Microsoft's ML upscaler without resorting to the GPU ALUs?
Yes, theoretically. It would be very beneficial for this hardware if used correctly.

Do we know for sure CVML is integrated into Van Gogh? I've just seen it in 2020 roadmaps, but plans could change. Is there any evidence of CVML implemented into this chip the Steam Deck uses? And if yes, how powerful is it?

Could be a game changer.
 
Nah, it's just a GNA comp offering from AMD.

Is it one of these?

https://www.design-reuse-embedded.c..._family_Datasheet_TIP_PB_Vision_DSP_FINAL.pdf

I read somewhere that it was the combination of a P6 and a C5.


Here:


If so, the dual C5 alone is seemingly capable of 2 TMACs INT8 / GHz, or up to 3.2 TOPs if it uses the same clocks as the GPU. Not too shabby on paper.

Of course, I have no idea if the implementation is practical for super low-latency operations such as ML upscaling in real-time renders.
 
Last edited by a moderator:
Back
Top