GPUs & CPUs & Enthusiast hardware: Questions, Discussion and fanboy slap-fights - Nvidia & AMD & Intel - Separe but Equal. Intel rides in the back of the bus.

[The Ugly One]

Yes. 1, developers are now using upscaling and frame generation as a substitute for proper optimization. 2, segmentation means that unless you have the latest cards their poor optimization hurts you further as the latest DLSS versions will require newer cards. If you honestly support this you may have brain damage.

Gamers have been saying, "This new technology is just making it okay for lazy developers to not optimize their code" since Intel introduced superscalar CPU technology to the desktop in 1993. Probably earlier. That's not a good reason to halt technological advance.

 

[#lurker]

Gamers have been saying, "This new technology is just making it okay for lazy developers to not optimize their code" since Intel introduced superscalar CPU technology to the desktop in 1993. Probably earlier.

That's irrelevant. It's objectively true that optimization is getting worse when developers are using DLSS for recommended settings.

 

[The Ugly One]

That's irrelevant. It's objectively true that optimization is getting worse when developers are using DLSS for recommended settings.

It's objectively true that optimization is getting worse when developers rely on the compiler to optimize their C++ code instead of handwriting assembly.

Still not a reason to freeze technology in the past. Inferencing is the biggest single-generation technology leap we've seen since DX8 and SM 1.0.

 

[#lurker]

It's objectively true that optimization is getting worse when developers rely on the compiler to optimize their C++ code instead of handwriting assembly.

Still not a reason to freeze technology in the past. Inferencing is the biggest single-generation technology leap we've seen since DX8 and SM 1.0.

Upscaling and frame generation aren't pushing technology further. They are bandaid fixes.

 

[Dawdler]

It's objectively true that optimization is getting worse when developers rely on the compiler to optimize their C++ code instead of handwriting assembly.

Still not a reason to freeze technology in the past. Inferencing is the biggest single-generation technology leap we've seen since DX8 and SM 1.0.

Compilers have enough brainpower and tricks to do very good optimization themselves.
They can't optimize raw sewage of a code that we se nowdays with ported games

 

[The Ugly One]

Upscaling and frame generation aren't pushing technology further. They are bandaid fixes.

Tensor units and AI inferencing are literally new technology. I would say that what makes them different is they make gamers mad, but I've been around a long time, and the reaction to inferencing reminds me a lot of the reaction to T&L, hardware AA, superscalar execution, and multicore CPUs.

Real-time AI is able to turn this:


Into this:


Besides, the vast majority of games are developed using the PS5 as the lead platform, which has an AMD GPU with no inferencing ability. If anything, they're probably better optimized than they were back when devs just said "lol buy a new GPU" when their games ran like shit.

Compilers have enough brainpower and tricks to do very good optimization themselves.

My hand-optimized dot product is around 2x faster than what the compiler did. The original code is literally just this:

double result = 0.0; for(int i = 0; i < n; ++i) result += a[i] * b[i];

 

[#lurker]

Real-time AI is able to turn this:
https://kiwifarms.st/attachments/screenshot-2025-01-05-120608-unscale-png.6823631/

Into this:
https://kiwifarms.st/attachments/screenshot-2025-01-05-120812-png.6823634/

Yeah, now show actual motion. Anyone trying to shill DLSS always posts dishonest stills.
Enjoy your ghosting, smearing, shimmering, and increased input latency.

 

[The Ugly One]

Yeah, now show actual motion. Anyone trying to shill DLSS always posts dishonest stills.

There are lots and lots of DLSS videos on YouTube. This is a comprehensive look at the Ultra Performance setting, which reduces the resolution by 3x in each direction, rendering internally at just 360p when upscaling to 1080p:

Here's MW2 at every DLSS setting:

Enjoy your ghosting, smearing, shimmering, and increased input latency.

You're thinking of FSR.

 

[geckogoy]

@The Ugly One : "nooo you have to let technology advance"

>its just interpolation
yea great advancement cheers, thanks for that

 

[Mister Fister]

Tensor units and AI inferencing are literally new technology. I would say that what makes them different is they make gamers mad, but I've been around a long time, and the reaction to inferencing reminds me a lot of the reaction to T&L, hardware AA, superscalar execution, and multicore CPUs.

Real-time AI is able to turn this:
View attachment 6823631

Into this:
View attachment 6823634

Besides, the vast majority of games are developed using the PS5 as the lead platform, which has an AMD GPU with no inferencing ability. If anything, they're probably better optimized than they were back when devs just said "lol buy a new GPU" when their games ran like shit.



My hand-optimized dot product is around 2x faster than what the compiler did. The original code is literally just this:

double result = 0.0; for(int i = 0; i < n; ++i) result += a[i] * b[i];

Gamers get mad at every tech leap. We've had an unusually long stagnation period due to the PS3/XBOX 360 era being so long and the pandemic. Upscaling and Ray tracing both are serious advances. Ray tracing will cut down on coding time as well.

 

[The Ugly One]

Gamers get mad at every tech leap. We've had an unusually long stagnation period due to the PS3/XBOX 360 era being so long and the pandemic. Upscaling and Ray tracing both are serious advances. Ray tracing will cut down on coding time as well.

A few I remember fondly over the years:

• 1990s:
  • Gamers rage at out-of-order execution because "you should just make the clockspeed faster and optimize your code better"
  • Hardware anti-aliasing is bullshit, it's just blurring, just run at a higher resolution
  • Mip-mapping is stupid, it's just blurring textures for no reason
  • SIMD is pointless, it doesn't make Quake go faster
• 2000s
  • Hardware T&L is a waste of silicon, just write better CPU code so we can have more TMUs
  • Normal mapping isn't real geometry (same would later be said of parallax mapping)
  • Multi-core CPUs are stupid, all game code is serial, what's needed are faster single-core CPUs
  • Thread parallelism is completely useless; threaded code is impossible to write, debug, or maintain (okay this came from developers)

Now gamers are mad that inferencing technology is used to make games prettier and faster because those aren't real pixels. They're fake pixels. The strange thing is we don't have to wonder what GPUs would be like if they didn't waste silicon on those silly, useless tensor units. We can look at AMD to see what that buys you, and it turns out you just get an inferior experience.

Anyway, the next place I'd like to see inferencing is in assets themselves. Why waste 100 GB on ultra-res textures if you can just inference your medium textures for results that are 99% as good?

 

[Mister Fister]

A few I remember fondly over the years:

• 1990s:
  • Gamers rage at out-of-order execution because "you should just make the clockspeed faster and optimize your code better"
  • Hardware anti-aliasing is bullshit, it's just blurring, just run at a higher resolution
  • Mip-mapping is stupid, it's just blurring textures for no reason
  • SIMD is pointless, it doesn't make Quake go faster
• 2000s
  • Hardware T&L is a waste of silicon, just write better CPU code so we can have more TMUs
  • Normal mapping isn't real geometry (same would later be said of parallax mapping)
  • Multi-core CPUs are stupid, all game code is serial, what's needed are faster single-core CPUs
  • Thread parallelism is completely useless; threaded code is impossible to write, debug, or maintain (okay this came from developers)

Now gamers are mad that inferencing technology is used to make games prettier and faster because those aren't real pixels. They're fake pixels. The strange thing is we don't have to wonder what GPUs would be like if they didn't waste silicon on those silly, useless tensor units. We can look at AMD to see what that buys you, and it turns out you just get an inferior experience.

Anyway, the next place I'd like to see inferencing is in assets themselves. Why waste 100 GB on ultra-res textures if you can just inference your medium textures for results that are 99% as good?

What's next raging because instead of using sin(x), you use 1-x^2/2 for the value range that produces the identical curve?

 

[warhammerautist]

I shared some of this information on another thread and I am assuming there might be some people that will be skeptical about their photonic computers like their cancer vaccine claim. I dont know if Asia, US or Europe will go down the same path but here we go.

Based on the experimental photonic processor Russia claims by 2030 they will have a computer that works 10^21 operations per second or rather 1 zettaflop which is what intel and the Japanese are trying to achieve with a bunch of supercomputers by 2030 as well.

https://www.rbc.ru/rbcfreenews/66b615749a79472cd0cb3a87

So, I am assuming CPU's and GPUs that run on photonics will be off the charts in video gaming.
This website states that PIC's can give you wireless 1000 Gbps speeds and their planned PIC production is in 2024.

https://technomoscow.ru/event/tekhnoklub-proizvodstvo-fotonnykh-integralnykh-skhem-fis/?ysclid=m3tkigd9c6107939088

Made their promise and the PIC production started 2 months ago.

https://iz.ru/1786402/maria-neduk/nahodka-dla-fotona-v-rf-nacnut-proizvodit-pribory-svazi-novogo-pokolenia

https://tass.com/science/1879543

Sergeyev also noted that, unlike modern computers with their universal processors and graphics processing units, developed by the leading international companies, specialized computing system can significantly speed up the solution of a number of tasks - by 100 to 1000 times - possibly reaching the exaop level by 2030, all while significantly increasing the energy efficiency.

If I am reading this right how would video gaming be like with GPUs and CPUs that operate with 100 to 1000 times the processing operations?

 

[Betonhaus]

If I am reading this right how would video gaming be like with GPUs and CPUs that operate with 100 to 1000 times the processing operations?

Probably very bloated code. With everything dynamically generated by AI. You wouldn't even program and build a game, just describe what kind of game you want and it builds itself.

 

[#lurker]

There are lots and lots of DLSS videos on YouTube.

Yes there are. And there are lots of them that show exactly what I'm talking about. Do you deny that DLSS has problems with ghosting, smearing, noise, etc.?

 

[The Mass Shooter Ron Soye]

Based on the experimental photonic processor Russia claims by 2030 they will have a computer that works 10^21 operations per second or rather 1 zettaflop which is what intel and the Japanese are trying to achieve with a bunch of supercomputers by 2030 as well.

We went through this hype cycle with Optalysys a decade ago. It's vaporware until it gets sold, and may not be a replacement for CPUs/GPUs but a decent accelerator at best.

If Russia wants to make a splash, an unconventional approach that sidesteps advanced lithography could be the best move.

 

[The Ugly One]

Yes there are. And there are lots of them that show exactly what I'm talking about. Do you deny that DLSS has problems with ghosting, smearing, noise, etc.?

The current iteration of DLSS 2 really doesn't. I've used all these technologies a lot over the last few years. I know you're not serious, but other people are, so here's the rundown:

• FSR 1 - It's basically garbage that is only slightly better than bilinear scaling. I never use it in a game. I'd rather lower settings if it comes to that. I only use it in things like CAD or viz software that has nothing else and runs like shit otherwise (a lot of this ancient software is written in OpenGL, which AMD really sucks at).
• FSR 2.0 - Decent at Ultra Quality, tolerable at Quality. At Balanced or below, the ghosting/shimmer was bad enough that I would rather lower settings to get good fps.
• FSR 2.1 - Significant update that makes it at least tolerable across the spectrum, but I wouldn't use it at Performance or Ultra Performance unless it was the only way to get 60 fps (it never is).
• XeSS DP4A - Superior to FSR 2.1 at any resolution and any setting level. It's almost as good as DLSS 2.0.
• DLSS 2.0 - Zero complaints at Balanced or better. Dodgy at Performance and below.
• DLSS 2.5 - Huge improvement, have yet to run into a scenario where I'd rather lower the settings than the DLSS base resolution.

Based on the experimental photonic processor Russia claims by 2030 they will have a computer that works 10^21 operations per second or rather 1 zettaflop which is what intel and the Japanese are trying to achieve with a bunch of supercomputers by 2030 as well.

Everyone is experimenting with photonics right now, including AMD, Intel and IBM.

If I am reading this right how would video gaming be like with GPUs and CPUs that operate with 100 to 1000 times the processing operations?

You're not reading it right. Photonics are much faster than silicon at transmitting data, but worse at logic gates, which makes them very good at specialized hardware. The only advantage they'd have over CPUs and GPUs, if you could even make a practical general-purpose computer out of them, is they don't get very hot. In the near term, we're likely to see data fabrics like memory buses and PCIe eventually get replaced with photonics.

 

[#lurker]

The current iteration of DLSS 2 really doesn't. I've used all these technologies a lot over the last few years. I know you're not serious, but other people are, so here's the rundown

You can like DLSS, you can defend DLSS, but you are in denial when it comes to its many negatives
You can watch this to see a number of examples.

All the polygons in modern games don't change the fact that games look worse than they did 10+ years ago thanks to a decline in motion clarity.

 

[The Mass Shooter Ron Soye]

AMD's spaghetti appears to be spilling early. What's on the menu?

Fire Range desktop CPUs in laptops. Not too exciting but a 16-core X3D flagship should be available soon, the 9955HX3D.

Strix Halo official. What's new? The 6-core, 16 CU model is "Pro" only, so it may be a niche (of a niche) part that some OEM wants. All are listed with a TDP range of 45-120W. AMD marketing will dare to compare the flagship Ryzen AI MAX+ 395 to the Apple M4 Pro.

RX 9070 XT and 9070 non-XT. The information is scarce showing AMD wanted to prevent this pre-briefing leak. However, it is apparently confirmed that FSR4 is exclusive to the RX 9070 series (what about 9060/9050/9040?) and only in games that already support FSR 3.1. One such game is Call of Duty: Black Ops 6.

Three new Z2 series APUs for gaming handhelds. These are:
Ryzen Z2 Extreme (Strix Point) = 3x Zen 5 cores, 5x Zen 5C cores, 16 RDNA3.5 CUs, 15-35W
Ryzen Z2 (Phoenix/Hawk) = 8x Zen 4 cores, 12 RDNA3 CUs, 15-30W
Ryzen Z2 Go (Rembrandt) = 4x Zen 3+ cores, 12 RDNA2 CUs, 15-30W, Lenovo exclusive
None of these have an enabled NPU.

Valve is confirmed to be a customer of the Ryzen Z2. But it might be a for a new Steam Machine and not the Steam Deck 2.

9950X3D and 9900X3D.

 

[4090 Chad]

Ryzen Z2 Extreme (Strix Point) = 3x Zen 5 cores, 5x Zen 5C cores, 16 RDNA3.5 CUs, 15-35W

I would be really curious to see the performance jump from the z1 extreme to this and how much more expensive this chip would make a new steam deck or ROG Ally.