• Welcome to TechPowerUp Forums, Guest! Please check out our forum guidelines for info related to our community.
  • The forums have been upgraded with support for dark mode. By default it will follow the setting on your system/browser. You may override it by scrolling to the end of the page and clicking the gears icon.

NVIDIA Shares Details About Ampere Founders Edition Cooling & Power Design - 12-pin Confirmed

Lets see. On the one hand, there's Chrispy_, a dude who has been discussing things reasonably with over at TechReport for over a decade and has proven himself to me (multiple times) to have a sharp mind.

On the other hand, there's Jinxed, with ~50 posts in history. And literally every single one of them is about Nvidia vs AMD bullshit I don't give a care about. Someone who gets flustered over the simple mention of ambient occlusion and... floating point numbers.

Clean up your posting history Jinxed. Start posting about other topics, and prove yourself to me if you expect me to take yourself seriously.
Says the guy who was trying to make a point comparing light propagation in corners of rooms with completely different materials on walls. Yeah right. I take you sooooo seriously after that.
 
These are repeating chunks of ordered noise, tiles - as I have called them - that cannot be true raytracing. It doesn't match gaussian/quantization noise that you'd get from a non-infinite number of rays and it has no place being there.

I've actually looked at lots of RNG patterns before. A repeating pattern like that could be explained by a PRNG, like XORShift. Honestly, to me, it looks like its "well randomized", but with a relatively small tile.

Most GPUs these days are only fast with 32-bit numbers. I'm going to guess that they just have a small PRNG state, so the "tiles" will be small. They're using something higher quality than your typical Unix LCGRNG for sure (see here for LCGRNG patterns: http://www.reedbeta.com/blog/quick-and-easy-gpu-random-numbers-in-d3d11/).

32-bit cycles go fast (4-billion is pretty small, all else considered), and many PRNGs fail at actually randomizing their bit differences... so you'll get repeating patterns over the 32-bit cycles. Especially if you're seeing the same pattern over and over again? Seems like the seed has just "cycled over" like an odometer.

EDIT: 1920 x 1080 x 8000 samples == 16-Billion samples. Enough to overflow a 32-bit PRNG four times. Just for some napkin math (and each sample may use more than one RNG value in the raytracing). Throw down some low-quality lower bits, and a "pattern" could very well emerge.

Just my opinion on that matter though.
 
Last edited:
I've actually looked at lots of RNG patterns before. A repeating pattern like that could be explained by a PRNG, like XORShift. Honestly, to me, it looks like its "well randomized", but with a relatively small tile.

Most GPUs these days are only fast with 32-bit numbers. I'm going to guess that they just have a small PRNG state, so the "tiles" will be small. They're using something higher quality than your typical Unix LCGRNG for sure (see here for LCGRNG patterns: http://www.reedbeta.com/blog/quick-and-easy-gpu-random-numbers-in-d3d11/).

32-bit cycles go fast (4-billion is pretty small, all else considered), and many PRNGs fail at actually randomizing their bit differences... so you'll get repeating patterns over the 32-bit cycles. Especially if you're seeing the same pattern over and over again? Seems like the seed has just "cycled over" like an odometer.

Just my opinion on that matter though.
Patterns can even be intentional, like here, for example:
"Ray tracing has two quality settings: high and ultra. Ultra setting traces up to one ray per pixel, with all the denoising and accumulation running in full. The high setting traces up to 0.5 rays per pixel, essentially in a checkerboard pattern, and one of the denoising passes runs as checkerboard. We recommend high for the best balance between image quality and performance, but please note that we are still experimenting a lot, so this information is valid only at the time of writing."
Interview on Metro Exodus with the developer's CTO and rendering programmer

But the pattern is about where the rays are cast, not about replacing or faking something with some kind of artificial tiles that Crispy is making up. It's not some made up fake thing. They rays still go to those locations, do their thing as normal. They devs are just reducing the number of rays to improve performance.
 
Those are no patterns. You are trying to conjure things where there are none.
There's obvious weird banding in those noise patterns, you just seem to like living within the Matrix :) .

Patterns can even be intentional, like here, for example:
I'm confused now, I thought there were no patterns.
 
I'm gonna sell my 2080Ti on Ebay and get as much as possible for it.
Definitely going for the 3090, as my PSU can handle it.
Gonna buy it on my card, get the Rewards Flyer points for it and then write the whole thing off as a business expense.
I do the same thing with a new or used laptop almost every year.
 
I've actually looked at lots of RNG patterns before. A repeating pattern like that could be explained by a PRNG, like XORShift. Honestly, to me, it looks like its "well randomized", but with a relatively small tile.

Most GPUs these days are only fast with 32-bit numbers. I'm going to guess that they just have a small PRNG state, so the "tiles" will be small. They're using something higher quality than your typical Unix LCGRNG for sure (see here for LCGRNG patterns: http://www.reedbeta.com/blog/quick-and-easy-gpu-random-numbers-in-d3d11/).

32-bit cycles go fast (4-billion is pretty small, all else considered), and many PRNGs fail at actually randomizing their bit differences... so you'll get repeating patterns over the 32-bit cycles. Especially if you're seeing the same pattern over and over again? Seems like the seed has just "cycled over" like an odometer.

EDIT: 1920 x 1080 x 8000 samples == 16-Billion samples. Enough to overflow a 32-bit PRNG four times. Just for some napkin math (and each sample may use more than one RNG value in the raytracing). Throw down some low-quality lower bits, and a "pattern" could very well emerge.

Just my opinion on that matter though.
Yeah, those tiles as I called them could easily be pseudorandom repetition from an interpolation algorithm too. I only said tiles because they looked like tiles, but I'm just saying what they look like. There's all sorts of ways that repeating pattern could be generated and PRNG is as good a hypothesis as any, probably better than tiles because that wouldn't require any additional VRAM to store the tile library, it'd be generated rapidly on the fly.

I'm familiar with pure CPU and GPU raytracers from a career perspective. Whilst I don't actively model and render myself, I'm responsible for budgeting and buying hardware farms for the teams that do, and in 20 years of doing this I've never seen repeating patterns like this, so I'm reasonably certain they're not generated by true raycasts. The various videos by devs on how they implemented DXR are all pretty cool, but I suspect the real secret sauce of what generates these patterns is another Nvidia black box that contains proprietary methods that Nvidia don't want to divulge to the competition. Unless someone finds more official detail, we're left to speculate with the evidence we can gather from the end result.
 
There's obvious weird banding in those noise patterns, you just seem to like living within the Matrix :) .
You are seeing things.

I'm confused now, I thought there were no patterns.
The pictures taken are from Quake II RTX. The interview is about Metro Exodus. Two different games, two different implementations.

I only said tiles because they looked like tiles, but I'm just saying what they look like. There's all sorts of ways that repeating pattern could be generated and PRNG is as good a hypothesis as any, probably better than tiles because that wouldn't require any additional VRAM to store the tile library, it'd be generated rapidly on the fly.
So now they are no longer tiles. Hmm. And it only took a few pages of debunking your "facts".
 
Last edited by a moderator:
Look at all that innovation:

My old XFX Fury (2015???) had a triple fan cooler with a short PCB like that so the third fan blew upwards, it resulted in lower CPU temperatures under combined load vs just CPU load in some situation.

I suspect a similar approach will be taken on may AIBs, does this mean the 3090 is going with HMB, surely you can't squash 20+ GB of GDDR6 into a short PCB like that?

Got a triple slot cooler on your new top end card? Welcome to 2008, you're going to love it!

Got a vapour chamber on your GPU? Welcome to 2006, you're going to love it.
That's kind of along the lines of why I think a section of the PCB being cutout as a air passage inlet/outlet between the top and bottom of the PCB with a twin blower setup would work well in principle just push pull push and the natural heat rising element will even do some of the work for you at the same time. The way it's been with GPU's for years is all the heat rises and any heat that doesn't get pushed out the rear of the case which even with a blower fan is probably relatively limited gets fairly stuck in place sure can kind of trickle out the sides, but I bet you a lot still remains overly stagnant and adversely impacts the PCB temps a fair bit. Really even if they had some holes like they use to mount M.2 drives like four to eight spaced out on the PCB to let some of heat escape upward it wouldn't hurt at all. They could do that and have two blowers on the top of the card with a heavily perforated hexagonal backplate to let air escape that pulls heat upward and pushes all that heat out the rear of the case perhaps with a pair of top down coolers down bottom. Basically four fans the card won't bend all the heat pretty much gets expelled out the rear of the case efficiently and could be run at low and quiet RPM's trivially they could even dynamically heat cycle between the top two fans and bottom fans.
 
Your 2x8-pin cable connects to the PSU with one 8-pin connector, not two.


nope two cables & two connectors one for each in psu...i don't know what psu you have but didn't saw yet the combination you mentioned...
 
@DuxCro
last time i checked, "you" aim for 1st, not second place..
so you want Nv to do less, just so amd can have something close in performance?
what other product on this planet gets artificially castrated, so competition can catch up?
----



funny how many ppl here "know" with how good one brand will be, and the other will be doing so bad, on cards that havent seen light of day yet,
nor been reviewed.

please make a sign and put it on your fridge:
it could/should/would..
the good thing? it will work for ANY future release of any hw. :D
 
Last edited:
nope two cables & two connectors one for each in psu...i don't know what psu you have but didn't saw yet the combination you mentioned...
I've seen both implementations. My prior EVGA PSU had a single 8-pin cable that then split into two 6+2 pins at the end. My current Corsair unit is 1 to 1 where the PSU end is 8pin and the other is 6+2pin. Both have their pros and cons, the pros of having the split is you only need to run one cable to the PSU to power a dual 8pin card, but the con is if you're using only one of them you've got a chonky 8pin dangling off to the side.
 
I've seen both implementations. My prior EVGA PSU had a single 8-pin cable that then split into two 6+2 pins at the end. My current Corsair unit is 1 to 1 where the PSU end is 8pin and the other is 6+2pin. Both have their pros and cons, the pros of having the split is you only need to run one cable to the PSU to power a dual 8pin card, but the con is if you're using only one of them you've got a chonky 8pin dangling off to the side.
this could explain why he insisted on his variant which i never saw but it seems that depend on manufacturer than ; thx for confirming!
 
Doubtless a nice piece of design, but necessitated by high power consumption. 8 layer PCB, back drilled VIAs, high compoent density/quality -> high BOM. They're explicit in the video that the way to get higher performance is through higher power use. Will be telling for efficiency.
From page 1. Correction, 12 layer PCB as per design guide. How do you like BOM now?
 
You know what frightens me the most? That those ridiculous prices are the new standard for graphics cards. Won't go down, and can only go up. :(
 
A company so rich cannot make a more polished video? This looks so amateurish, with meaningless graphics and horrible photography. There are million home produced channels on YouTube that have higher production values.
 
You know what frightens me the most? That those ridiculous prices are the new standard for graphics cards. Won't go down, and can only go up. :(
Well, if sales are going great (while prices are high), how can prices go down? All computer hardware sales in the last 6 months did awesomely. Prices might come down only when and if people buy less, but unfortunately that doesn't seem to be the trend.
 
You know what frightens me the most? That those ridiculous prices are the new standard for graphics cards. Won't go down, and can only go up. :(
if the consoles are priced at $500, the PC gaming market won't be able to maintain the ridiculous increasing Nvidia tax for too much longer. Despite the vocal DIY PC market, your average gamer is just going to be swayed towards one of the new consoles instead. The games are heavily optimised for the console hardware and controllers, they get earlier game releases and addtional exclusives and the online community for consoles is usually larger when it comes to matchmaking purposes and keeping servers alive.
 
Consoles going from 7850/7870 GPU to 2080/2080s levels, beating 95/98% of the PC market at GPU power and now targeting 4k resolutions will only make it worse for the PC market.
 
So a single 12-pin will pull power from a single 8-pin pci-e cable?

2 x 8-pin, same way when 8-pin connectors were new and weren't on psu's yet, people had to get by with 2 x 6-pin to 8-pin adapters.
 
Back
Top