Source
- URL: https://x.com/NOTimothyLottes/status/2106566586819609002
- Author: NOTimothyLottes (@NOTimothyLottes)
- Posted: 2026-10-04 02:06:27
Thread
1/ @NOTimothyLottes
Thread: Serious talk on SteamDeck 2 speculation starting with numbers based on the nice summary here https://www.youtube.com/watch?v=QSg5pjkcg9E - TLDR, actually this is exciting if true.
Branches: 2026-10-04-SebAaltonen-funny-that-they-say-actual-fp32-compute-meaning
2/ @NOTimothyLottes
Lets talk 2 configurations based on Moore’s Law is Dead (MLID) and The Phawx speculation numbers,
Estimated SteamDeck 2/
RDNA3.5 3.2 GHz * 12 CU = 38.4
MLID estimate of another handheld
RDNA5 1.2-1.65 GHz * 16 CU = 19.2-26.4
Given that estimated comparison, I’d go Deck2
3/ @NOTimothyLottes
SteamDeck has historically had rather amazing power slosh ability - if they keep this with the Deck2 and focus on higher peak clocks, it will be easy to get the CPUs to idle and actually leverage what the GPU can do!
4/ @NOTimothyLottes
As for ALU:BANDWIDTH, I’ll be ALU-bound, so it’s ok to be thin in the bandwidth department IMO.
5/ @NOTimothyLottes
As for RDNA 3.5 vs RDNA 5 - no plublic RDNA 5 ISA guide yet. RDNA 3.5 already has the one interesting thing of scalar FP32 ops. MLID says 40-50% more “raster” for RDNA 5 vs 3.5, but I doubt it. 3.5 already has dual issue FP32. Where you going to find 50% more IPC?
6/ @NOTimothyLottes
Can easily push near 100% VALU issue on prior RDNA architectures before dual FP32 issue. With dual FP32 issue (3.5 has), get some amount above that of effective IPC. The next logical evolution is polishing some of the dual corner cases, but that won’t bring 50% IPC in practice.
7/ @NOTimothyLottes
RT is effective useless on bandwidth and cache constrained portable. And ML scaling is crap. So whatever RDNA5 brings to the table for RT and ML is simply more static parasitic power draw.
8/ @NOTimothyLottes
In the end if the rumored Deck 2 specs are true, in optimized code like what I’d do, you’d see a 3x realized perf increase. And that would be way better than any RDNA5 portable that didn’t have matching clock*CU count capability.
9/ @NOTimothyLottes
IMO Valve would be genius to just ship the same 800p OLED. A 3x ALU perf boost with no need to push 2x or more pixels, would make devs quite happy, something that actually feels quite a bit faster.
10/ @NOTimothyLottes
But I’d expect the obvious mistake of going to 1080p or 1440p. And I’d mitigate that with a CRT emulation scaling to simulate a lower resolution display. That would be a little perf tax, but nothing horrible like an ML TAA pass.
11/ @NOTimothyLottes
The real danger that this possible Deck 2 presents is this: it’s Linux, you don’t have to use Vulkan, you can go direct to the kernel driver. Someone could use AI to write a PS5 shader binary to RDNA 3.5 binary translator, and then do direct emulation of system libraries.
12/ @NOTimothyLottes
If Deck 2 would release with similar sales as Deck 1, it’s past the threshold of being interesting for a single indie for a to-the-metal non-portable game too, ie something going direct to the kernel drivers with assembly shaders.
13/ @NOTimothyLottes
Also one has to wonder this, if the AI slop-gramming tools are as good as people claim they are, why use C? why use Vulkan? Why not just slop-gram direct to assembly on GPU? Challenge someone to go there.