Share this page!

Enter URL or ID to Unroll

You can paste full URL like: https://x.com/threadreaderapp/status/1644127596119195649
or just the ID like: 1644127596119195649

How to get URL link on X (Twitter) App

  1. On the Twitter thread, click on or icon on the bottom
  2. Click again on or Share Via icon
  3. Click on Copy Link to Tweet
  4. Paste it above and click "Unroll Thread"!
  5. More info at Twitter Help
NerdTech Profile picture
NerdTech
@nerdtechgasm
Software Dev & Hardware Tech Enthusiast. Partakes in Geopolitical & other interesting discussions.
Aug 18, 2023 • 8 tweets • 2 min read
It appears that TSMC's attempt to expand in Arizona has been one giant shitshow. From US construction & trade Union causing problems to cost overruns, to site accidents due to the rushed pacing (or as one Taiwanese manger puts it, "US workers are retarded").. Officially it's delayed by 1 year, but unofficially, the managers know that's just PR. They can't find skilled engineers and tried to import some from overseas, but now local gov is playing politics, claiming it's some of the CHIPS act money being spent, should "hire Americans!"
Nov 15, 2022 • 4 tweets • 2 min read
It is good to see more western countries no longer recommend vaccines for <50 yr olds. As safety signals raise alarms, policies needs to be update.

"Safe & effective", we now know "effective" is very short duration that wanes into negative. "Safe", we'll find out more in time. ImageImageImage The Danes had this change in policy months ago. Other EU nations quickly followed. Australia finally updated their guidelines, no longer recommend vaccines for <30 yr old.

It should have been only for the vulnerable population from the start! Image
Nov 15, 2022 • 5 tweets • 2 min read
So looking at the slides AMD gave on RDNA 3 2nd gen RT (RA hw), it is obviously going to need dev support, because its hw that requires new instructions to access; ray flags, box sorting, and bvh stacking.

Not brute force that would have gained more perf across the board. They look the path of least xtors.. not surprised at all tbh given their focus on density & GCD size.

It's just won't be pretty when games don't come optimized for RDNA 3 RT hw, ie. RTX sponsored games.
Nov 13, 2022 • 5 tweets • 2 min read
Another new research paper, from Israeli scientists, on vaccine effectiveness waning. They found, even 3rd booster, 4th booster, wanes quickly. After ~3 months, into negative vs unvaccinated.

Those who had prior covid, their immunity lasted much longer.

nature.com/articles/s4146… Consider this data. The waning of protection already occurs on day 10+.. ie, if you had 3 shots, you are no better protected than unvaccinated who caught covid first time.

"Natural immunity" (prior infection) provides more effective & longer protection, over 2 years!
Nov 5, 2022 • 6 tweets • 2 min read
Re: Raster vs RT

It seems we have this misunderstanding every time a new GPU gen launches. Last time, RTX 3090 launched, people also said raster wasn't important, because it crushed old games at 4K.

New games come out to humble it. Cycle repeats, it's inevitable.. When you give devs extra compute & raster perf, they will surely take advantage of it.

Game engines evolve, they get more complex, better fidelity.
Nov 4, 2022 • 4 tweets • 1 min read
If AMD released a 450mm2 + 6 MCD, 450W chip, with tensor cores (lol why) and better RT, matching NV in perf and price it at $1399, would they gain marketshare?

Jensen would just drop the price a bit, and most would buy NV anyway, as they hoped? The only justification for tensor and XMX on gaming GPU thus far is upscaling.

FSR 2.1 is already decent. FSR 2.2 claims to improve ghosting further, we shall see.

Why waste die space & power for matrix units? Just to look good on paper?
May 3, 2022 • 4 tweets • 1 min read
My view is that AMD lacks the 5nm supply at TSMC to hit their goals for both Zen 4 (EPYC demand is record levels from what I have heard, Ryzen will be a hit too), and their N32 and N31. They knew it, so they've gone N33 first on mature high supply node. This cadence works better for AMD, where margins on Zen 4 chiplets to consumer & enterprise market will be very high. Front load Zen 4. Then bring to market bigger RDNA3 in 2023.
Apr 25, 2022 • 4 tweets • 2 min read
World of Warcraft, 5800X3D blitz with big gains. Another nice advantage, it gets 99% of the potential perf with DDR4 3200, no need for expensive RAM kits.

This can be really nice for someone even on 1st gen Ryzen with DDR4 3200. They plug in 5800X3D and get a huge perf upgrade without their older memory bottlenecking perf.
Apr 22, 2022 • 4 tweets • 1 min read
This century is going to be wild. Image Literally, the Asian Century.
Apr 2, 2022 • 5 tweets • 2 min read
"FineWine" is really post-game release driver re-optimizations (replacing dev shaders). NV being the dominant PC GPU vendor, heavily incentivizes studios to optimize the PC original (especially), or PC port of consoles (that are usually RDNA2 optimized). Generally, PC original games tend to run much better on NVIDIA GPUs on release.

While neutral PC ports of console games, tend to perform well on AMD GPUs. Sometimes, performing poorly on NV.

Sponsored PC ports of console games, from NV, tend to run better on their hw (shocker).
Mar 31, 2022 • 4 tweets • 1 min read
Come on man, give @IntelGraphics a break! 😅 To be fair, the good results here in gaming perf, comes at the expense of an inferior media engine and lack of advanced AI/ML blocks on the Radeon design.

Engineering is all about compromises. :)
Mar 30, 2022 • 4 tweets • 1 min read
Proprietary solutions only benefit when you are the dominant player in a given market.

dGPU underdogs should focus to nullify the GPU leader's key feature, DLSS, by offering equivalent features that are both open source & widely hw compatible. Much more beneficial, to increase XeSS adoption first, with DP4a incentivizing studios to reach broader audience. Then update library later, add XMX path, for added perf on ARC.

This approach counters DLSS & FSR, while benefiting @IntelGraphics down the road.
Mar 30, 2022 • 4 tweets • 1 min read
Iris Xe with 96EU -> A370M with 8/128EU, slightly higher clocks. Per EU, some gains.

It looks like perf in the ballpark of AMD's recent RDNA2 iGPU? If anything this makes the iGPU on AMD look good, considering it shares system memory for similar perf.

Intel needs an ADL CPU + A370M, to compete vs just Ryzen APU. :/
Mar 30, 2022 • 5 tweets • 2 min read
Some assume just because rt is done via "standard API" DXR, therefore, it's performance is deterministic based on hw capability, leaving little room for optimization.

The major bottleneck in perf is SIMD lane occupancy ray testing, and later shading + denoising. Intel's ARC RT presentation covered some of the optimization principles.

I also mentioned that RDNA2 RT hw is not designed for DXR 1.0, in this mode, its perf is a fraction of what's possible because most of the potential perf is wasted.

Jun 9, 2021 • 4 tweets • 1 min read
The trend for graphics architecture is reduced ALUs per SM/CU, as it is difficult to keep utilization high. Shared registers & cache subsystem, and instruction fetch, decode, and warp schedulers all have to be expanded to utilize the extra ALUs. At which point it's more advantageous to just expand number of SM/CUs. Kepler -> Maxwell -> Turing etc went from higher ALUs per SM to 64. Its only Ampere where FP32 is doubled, but its really a datacenter compute focused design.
Mar 20, 2021 • 4 tweets • 2 min read
It's basically this: game engines aimed at high fidelity visuals test & optimize on NV if it's PC exclusive first, then AMD. Cross-platform high fidelity engines designed for AMD first, then NV. Intel GPU testing is rarely ever given the time it needs. You can understand why. Why test & optimize for Intel GPU architecture & driver stack when almost nobody is going to be AAA gaming on an iGPU? It cannot justify the time & manpower investment. Even some of the PC exclusive engines are poorly optimized for AMD due to low marketshare!