Osvaldo Pinali Doederlein@opinali

Sci-Fi, Gaming, PC hardware, Tech | Unoptimized code is the root of all evil | bsky.app/profile/opinali.me | medium.com/@opinali

 New Jersey
Joined April 2009
2,721 Photos and videos
DLSS 5 is finally upon us (soon), and NVidia's marketing doesn't miss. The launch videos are great, showing the new teach it its best light yet (sorry). We have a more details about how it works and finally how it looks with proper integration. But let's take this apart... 🧵
5
17
2,648
I'm playing 007 First Light the way NVidia says we should: using a lot of DLSS to handle heavy games. This is at 4K max settings, but with an RTX 5080 which is the best GPU you can buy today without selling an important organ. How does it work?🧵 youtube.com/watch?v=gYKKMe9e…

007 First Light, DLSS/FG artifacts

RTX 5080, 4K, Max settings, DLSS/P + FG X6

youtube.com
1
2
215
more replies
Worst-case scenario: heavy game; my pre-FG perf is in the mid-40s or low-50s in those scenes; Path Tracing noise; artifacts of SR + FG + RR compound. But this is the marketing: we should max out DLSS and get 300fps for those settings on a 5090 (or 200 on a poorer gamer's 5080).
1
1
34
Even in the marketing videos you can pause at random and often see bad artifacts. I'm not hating on DLSS, the fact it works as well as it does is magic. But when they say "AI draws 23 out of every 24 pixels" there is a cost. We still need A LOT more power for real rendering.🔚
1
29
Osvaldo Pinali Doederlein retweeted
Bryan Cantrill
@bcantrill
23h
If I end up having to write a blog entry explaining that a computer program is not a person, I'm going to be very upset
Aran Nayebi
@aran_nayebi
Oct 2
Trying to locate the spark.
74
65
881
19,547
I will admit the era of ASI, singularity or whatever you call, is upon us, when AI can fix big engineering problems human intelligence clearly can't fix, like: - A laser printer that never jams - Credit cards that cannot be cloned - Feed 600W+ into a GPU with one cable safely
10
455
Did latest 007 updates improve anything important? Nope, still no Reflex without FG. This is bad. Everything else should be priority #819 after anti-lag tech for all GPUs, @IOInteractive. Not that I would pay at 40fps anyway, but this is ultra settings at 4K so I can tweak it.🧵
4
1
22
1,182
I'm also thinking those benchs showing latency barely changed from FG 2X to 6X are lying to me. In my admittedly limited test in this scene I see ~62ms at 2X and up-to ~72ms at 4X. "Up to" because latency is less stable at higher FG rate, goes up and down even standing still.🔚
9
172
Hot take: AI Coding + Monrepos = 🔥🔥🔥 At work when I ask AI to help with coding, it can scan every library that my code uses at the impl level if needed. For deep optimization, obscure bugs, poorly documented libs... it's often magic when AI can see source code for EVERYTHING.
1
265
I wanted to watch this movie all my life but for some reason, never did before. I had the CD of Philip Glass's soundtrack since I was a teenager and it's still in my playlists, I know that music by heart, probably listened to the whole thing a thousand times. The movie? 🧵
1
177
It's the masterpiece I expected. Music is a HUGE part of it. And I had minimal info about the movie or the story but... I swear, I knew most of it just from listening to Glass's instrumental-only music. The beauty, the art, the obsession, the craziness, the climax all there. 🔚
89
Gemini is absolutely back to SOTA, at least for coding.
223
And RDNA4 was not yet caught up in PT; and also in AI, which it's very competitive in many areas/benchs but for gaming-oriented tech such as RR it seems NVidia still has significant performance advantage. Maybe it's in part less mature/optimized SW, but only in part.
1
8
456
So if RDNA5 only comes in early 28 or even Q4'27 that's a disaster even if AMD pick ups the pace with stuff they can improve (FG) or get better adoption (RR). Six months from now RDNA4 will compete much less, even compared to Ada at similar price tier.🔚
3
7
401
This AI-induced crisis, causing the next GPU gen to come later than usual, will be bad for AMD IMO. Because NVidia's current lineup has more fine-wine potential. The rollout of software features for Blackwell has been slow. DLSS4.5+RR, MFG, RTX Kit, NRD, DLSS 5, Reflex 2 etc. 🧵
14
1
29
3,100
But these features sooner or later are being delivered. I don't buy into every new hype, but today already I rarely play any games with my 9070 XT because most new AAA games now support Path Tracing. Adoption of PT has improved a lot, and DLSS4.5+RR makes it viable more often.
3
8
459
The Adrenalin 26.9.2 package updates Vulkan support from 1.4.349 to 1.4.355, and it also fixes Vulkan/GPGPU performance: My GeekBench 7 score had collapsed from 238K in 26.6.2 to 67K in 26.7.1, now it's back to 196K. Not fully back, but now OpenCL & Vulkan are identical perf. 🧵
2
2
31
2,004
Perhaps the OpenCL & Vulkan drivers are more unified; OpenCL's perf actually loses some, 206K in 26.9.1 to 194K now (though GB7's run-to-run variance is relatively high so could be luck even with my best-of-3 rule). But no changes in Vulkan games, DOOM TDA bench the same. 🔚
1
8
411
So, more stuff that was new in this driver but not mentioned in the Release Notes👇 There are also some changes in the Radeon App, when you tweak things for a running game that only take effect if the game is restarted the app now warns you.
AMD GPUOpen
@GPUOpen
Sep 30
AMD Advanced Media Framework (AMF) SDK v1.5.3 is now available! 🆕 Color matrix support for VideoConverter & VideoDecoder. 🔄 Rotation support for VideoConverter, and more. 🔧 Requires AMD Software: Adrenalin Edition 26.9.2 or newer. Download now 👇 gpuopen.com/advanced-media-f…
4
174
The fantasy: The US will build all these AI datacenters using many Gigawatts and lead the next century The reality: My commuter train out of New York just stopped for a few minutes because it lost power somehow
5
369
Load more