Fulcrum29
Honorary Master
I see NVIDIA has allowed their board partners to go crazy with the 5090. MSI and Gigabyte now also have halo products built on that, knowing that the ASUS card is sold out, and see them selling like hotcakes too.
-
Join the MyBroadband community
South Africa’s biggest forum. Discuss, discover, and connect with thousands of members.
-
Question of the Day: Which is the most hair-raising South African mountain pass you've driven?
Vorastra
Honorary Master
Nice thread @Fulcrum29 
All the announcements in one place.
Saw a guy already reporting performance metrics for DLSS 4.5 on non-FP8 accelerated cards (pre-40 series).
15-25% drop in performance compared to TransformerV1 on 3080 Ti.
All the announcements in one place.
Saw a guy already reporting performance metrics for DLSS 4.5 on non-FP8 accelerated cards (pre-40 series).
15-25% drop in performance compared to TransformerV1 on 3080 Ti.
Fulcrum29
Honorary Master
Some here might need this though:
There is also the Lepro AMI that is an AI soulmate, yes, soulmate.
Fulcrum29
Honorary Master
For those interested in the whole "Community Update".
There is not much there, but do expect more games to adopt Ray... sorry Path Tracing. This will lean more towards DXR 1.2 and Vulkan (new tech).
Fulcrum29
Honorary Master
Nice thread @Fulcrum29
All the announcements in one place.
Saw a guy already reporting performance metrics for DLSS 4.5 on non-FP8 accelerated cards (pre-40 series).
15-25% drop in performance compared to TransformerV1 on 3080 Ti.
I am sure that there will be improvements in time, also as newer technologies are being adopted. It won't be radical, but RTX 20, and 30 could still be in an okay spot. RDNA 1, 2, and 3... they will all be challenged by path tracing. RDNA 1 and 2 should be avoided, and RDNA 3 is in a grey area.
Also when I said improvements, it will only really be applicalbe to new games that are adopting the new tech stacks, as well as those that could be updated. All I know is, some GPUs will start to rapidly age.
Vorastra
Honorary Master
It's worse. This is the frame time table from the official Nvidia DLSS programming guide so devs can gauge what to expect.
New DLSS 4.5 is doubly slow on 20 and 30 series. Not looking good. The issue with this is that there are only two ways to achieve higher quality with new transformer models. Bigger model, or smaller quants. There's no real way for performance to improve on 20 and 30 series, and soon 40 series when Nvidia naturally moves DLSS down to FP4.
The 40 series will be the big loser. FP4 is going to be a sort of big wall for a while. FP1 and FP2 are just not there yet. Nvidia's plans out to 2027 (Vera Rubin) are still leaning on FP4, and the 50 series is the first consumer generation to implement acceleration for it.
Fulcrum29
Honorary Master
Those are rough estimations that aren't dependent on a game engine and its integrations. Not all engine implementations work the same. The guide you see there is NVIDIA running the DLSS library without a 3D renderer. Though I can agree that it is still indicative as to what execution times can be expected, varying GPU to GPU, this can still be optimised per architecture. It is not a best case scenario, other tech stacks will also have its impacts coming with pros and cons. There is also documented usage to reduce the allocated memory which could improve latencies.
It is also worth noting that Blackwell has bespoke hardware that its predecessors don't have. Naturally it will be excel at some OptiX applications.
RTX 20 is now more than 6 years old, I think it released in late 2018. RTX 30 is 2020.
Six years... architecture does have its limits, and there is only that much you can do to optimise pipelines on the driver end. Is the 2080 Ti even a 1440p capable path tracing GPU? I don't think so.
Vorastra
Honorary Master
The whole MFG thing, which I very much like.
Thing is, even with Preset K... I mean...it's pretty damn good. A 20/30 user really shouldn't be dooming. Sure it's not the latest but who cares.RTX 20 is now more than 6 years old, I think it released in late 2018. RTX 30 is 2020.
Six years... architecture does have its limits, and there is only that much you can do to optimise pipelines on the driver end. Is the 2080 Ti even a 1440p capable path tracing GPU? I don't think so.
And now we have rumours that Nvidia are restarting 3060 production....?
Speaking of VRAM usage, Nvidia still have neural texture compression in the works, as if this post it's at version 0.9.1 with the big hold up being Microsoft's cooperative vector release. That's expected to reduce VRAM usage by 4-6x. But again, that's even more load on an ever heavier tensor pipeline. Also slated for 40/50 series.
Fulcrum29
Honorary Master
Yes, but we have reached a point where rasterisation is not the be-all. Though newer GPUs are better at raster, logical reasoning is becoming ever more crucial.
They would only consider doing so where it is viable.
I have discussed that in the AMD Radeon thread. I don't think that it is MS holding them back, but that it is a cooperation between NVIDIA, Intel and AMD. Cooperative Vectors have only recently been introduced to RDNA 4. Let's say that NVIDIA are way ahead with the Shader Model. Intel and AMD don't have OMM, and only have SER workarounds. MS can't advance their Agility SDK without leaving Intel and AMD behind. Vulkan is now also getting there.
This is NVIDIA's RTX Kit technical blog,
Get Started with Neural Rendering Using NVIDIA RTX Kit | NVIDIA Technical Blog
Neural rendering is the next era of computer graphics. By integrating neural networks into the rendering process, we can take dramatic leaps forward in performance, image quality…
developer.nvidia.com
and their RTX NTC SDK:
GitHub - NVIDIA-RTX/RTXNTC: NVIDIA Neural Texture Compression SDK
NVIDIA Neural Texture Compression SDK. Contribute to NVIDIA-RTX/RTXNTC development by creating an account on GitHub.
github.com
The NTC system requirements are a interesting quote:
Nvidia also have non-Cooperative Vectors DX12 decompression. This will come with cons, it enables compatibility, but at a big cost. Reading their SDK's NVIDIA is transparent, and is not hiding anything and I hope that tech reviewers are sharper than they make themselves out to be.
Fulcrum29
Honorary Master
I am going OT now, and this to my knowledge isn't yet introduced in the main AMD Radeon branch. This is AMD's MS Agility SDK dev driver:
I have already made my remarks about this in the AMD Radeon thread. Looking at those highlights RDNA 3's shader cores should still have some legs. The question though, and I have asked AMD directly without receving a response is whether RDNA 3 could and will support Cooperative Vectors.
RDNA 3 being able to support all that, also having the ability to do SER (though with limitations) I can't see it not having FSR "Redstone" though it would have to be dependent on the INT8 model. That model can be improved, and I am sure AMD did.
There is also some things on that list some said RDNA 3 can't do. It is a dev driver...
I have already made my remarks about this in the AMD Radeon thread. Looking at those highlights RDNA 3's shader cores should still have some legs. The question though, and I have asked AMD directly without receving a response is whether RDNA 3 could and will support Cooperative Vectors.
RDNA 3 being able to support all that, also having the ability to do SER (though with limitations) I can't see it not having FSR "Redstone" though it would have to be dependent on the INT8 model. That model can be improved, and I am sure AMD did.
There is also some things on that list some said RDNA 3 can't do. It is a dev driver...
Fulcrum29
Honorary Master
I am also starting to pity the tech reviewers. They have long had an emphasis on gaming, and now gaming is the third wheel. Their reporting is akward. Taking digs at AI won't keep them relevant. Odd pivots coming up. At a time they hated upscaling, then they hated FG, and now it is all comparitive reviews. The market moves.
I am concerned with gaming becoming spoiled by AI assistants. I don't have an issue with it in general, but in competitive gaming it could potentially become pay to win. Now some would argue that there are better keyboards, mice, headsets, and hardware, but they do not provide in-game assistance, except those peripherals that enable scripting. Peripherals that allow scripting are disallowed at most esports events.
All in all I am not against it as it does have its place, but that place isn't everywhere. As long as it can be localised I am good with it. Just don't use it to cheat against other human players.
I am concerned with gaming becoming spoiled by AI assistants. I don't have an issue with it in general, but in competitive gaming it could potentially become pay to win. Now some would argue that there are better keyboards, mice, headsets, and hardware, but they do not provide in-game assistance, except those peripherals that enable scripting. Peripherals that allow scripting are disallowed at most esports events.
All in all I am not against it as it does have its place, but that place isn't everywhere. As long as it can be localised I am good with it. Just don't use it to cheat against other human players.
Vorastra
Honorary Master
Vorastra
Honorary Master
So it's not that Preset M/L is meant to supersede Preset K completely but rather be alternate 'better' choices aimed at specific quality presets.
So Preset K (TransformerV1) is still the preferred choice when using DLAA/Quality/Balanced.
Preset M (TransformerV2a) is meant for Performance mode.
Preset L (TransformerV2b) is meant for Ultra Performance mode.
That makes sense then for them to be heavier than Preset K. A heavier model but used at lower qualities, so you still get some of the performance benefit but much better (lets assume) image qualitythan with Preset M\L at Perf and UltraPerf settings. This makes much more sense as to why Preset L is so heavy in NVIDIA's documentation.
So Preset K (TransformerV1) is still the preferred choice when using DLAA/Quality/Balanced.
Preset M (TransformerV2a) is meant for Performance mode.
Preset L (TransformerV2b) is meant for Ultra Performance mode.
That makes sense then for them to be heavier than Preset K. A heavier model but used at lower qualities, so you still get some of the performance benefit but much better (lets assume) image quality
Last edited:
Fulcrum29
Honorary Master
ASUS's (non-ROG) launch event will be in an hour at 19h00.
Do expect AI and mobility. This is what CES 2026 is all about. Many things seen now at CES 2026 are things already seen at CES 2025, but improved. Nothing revolutionary, yet.
Do expect AI and mobility. This is what CES 2026 is all about. Many things seen now at CES 2026 are things already seen at CES 2025, but improved. Nothing revolutionary, yet.
Vorastra
Honorary Master
I would lol if they announced 1st-party DRAM.
Fulcrum29
Honorary Master
HyperX has partnered with Nuerable to make a gaming headset that is supposedly made to make a gamer to be "locked-in." Everyone who has reported on this has stated to have had their aim and reaction time improved by it. I am saying it now, it is not the headset, nor the exercise. It is because there is a pattern to the AimLabs demo.
It doesn't take two runs to improve. Stop being paid shills. That headset needs some time with its user, and the exercises need to be repeated.
It doesn't take two runs to improve. Stop being paid shills. That headset needs some time with its user, and the exercises need to be repeated.
Fulcrum29
Honorary Master
Lenovo's keynote is during the early morning hours.
There agenda seems to be all AI, since this is what they want their audience to see:
More hybrid AI talk, more agentic AI talk, and more AI to make people smarter every day...
From every vendor I have seen, it seems that almost everyone is partnering with Intel. This is a sign that AMD is pivoting towards enterprise. Intel targets the business consumer void. Time will tell.
There agenda seems to be all AI, since this is what they want their audience to see:
More hybrid AI talk, more agentic AI talk, and more AI to make people smarter every day...
From every vendor I have seen, it seems that almost everyone is partnering with Intel. This is a sign that AMD is pivoting towards enterprise. Intel targets the business consumer void. Time will tell.
Fulcrum29
Honorary Master
MSI debuts its MEG X QD-OLED AI gaming monitor, alongside Nvidia G-Sync Pulsar and 5K DFR displays | Club386
There's a variety of screen types and features to appreciate in MSI's 2026 gaming monitor lineup, it's almost difficult to pick a favourite.
club386.com
Oh, my...
Fulcrum29
Honorary Master
The Siemens keynote is slated to be a good one:
They are very much involved in everything.
They are very much involved in everything.