Friday, September 7th 2018
NVIDIA's 20-series Could be Segregated via Lack of RTX Capabilities in Lower-tier Cards
NVIDIA's Turing-based RTX 20-series graphics cards have been announced to begin shipping on the 20th of September. Their most compelling argument for users to buy them is the leap in ray-tracing performance, enabled by the integration of hardware-based acceleration via RT cores that have been added to NVIDIA's core design. NVIDIA has been pretty bullish as to how this development reinvents graphics as we know it, and are quick to point out the benefits of this approach against other, shader-based approximations of real, physics-based lighting. In a Q&A at the Citi 2018 Global Technology Conference, NVIDIA's Colette Kress expounded on their new architecture's strengths - but also touched upon a possible segmentation of graphics cards by raytracing capabilities.
During that Q&A, NVIDIA's Colette Kress put Turing's performance at a cool 2x improvement over their 10-series graphics cards, discounting any raytracing performance uplift - and when raytracing is indeed brought into consideration, she said performance has increased by up to 6x compared to NVIDIA's last generation. There's some interesting wording when it comes to NVIDIA's 20-series lineup, though; as Kress puts it, "We'll start with the ray-tracing cards. We have the 2080 Ti, the 2080 and the 2070 overall coming to market," which, in context, seems to point out towards a lack of raytracing hardware in lower-tier graphics cards (apparently, those based on the potential TU106 silicon and lower-level variants).
This is just speculation - based on Kress's comments, though - but if that translates to reality, this would be a tremendous misstep for NVIDIA and raytracing in general. The majority of the market games on sub-**70 tier graphics cards (the 20-series has even seen a price hike up to $499 for the RTX 2070...), and failing to add RT hardware to lower-tier graphics would exclude a huge portion of the playerbase from raytracing effects. This would mean that developers adding NVIDIA's RTX technologies and implementing Microsoft's DXR would be spending development resources catering to the smallest portion of gamers - the ones with high-performance discrete solutions. And we've seen in the past what developers think of devoting their precious time to such features.
Additionally, if this graphics card segregation by RTX support (or lack of it) were to happen, what would be of NVIDIA's lineup? GTX graphics cards up to the GTX 2060 (and maybe 2060 Ti), and RTX upwards? Dilluting NVIDIA's branding through GTX and RTX doesn't seem like a sensible choice, but of course, if that were to happen, it would be much better than keeping the RTX prefix across the board.
It could also be a simple case of it not being feasible to include RT hardware on smaller, lower performance GPUs. As performance leaks and previews have been showing us, even NVIDIA's top of the line RTX 2080 Ti can only deliver 40-60 FPS at 1080p in games such as the upcoming Shadow of the Tomb Raider and Battlefield V (DICE has even said they had to tone down levels of raytracing to achieve playable performance levels). Performance improvements until release could bring FPS up to a point, but all signs point towards a needed decrease in rendering resolution for NVIDIA's new 20-series to be able to cope with the added raytracing compute. And if performance looks like this on NVIDIA's biggest (revelaed) Turing die, with its full complement of RT cores, we can only extrapolate what raytracing performance would look like in cut-down dies with lower number of RT execution units. Perhaps it really wouldn't make much sense to add the increased costs and per-die-area of this dedicated hardware, if raytracing could only be supported in playable levels at 720p.
All in all, it seems to this editor that segregation of graphics cards via RTX capabilities would be a mistake, not only because of userbase fracturing, but also because the highest amount of players game at **60 and lower levels. Developers wouldn't be so inclined to add RTX to their games to such a small userbase, and NVIDIA would be looking at dilluting its gaming brand via RTX and GTX - or risk confusing customers by branding a non-RTX card with the RTX branding. If any of these scenarios come to pass, I risk saying it might have been too soon for the raytracing push - even as I applaud NVIDIA for doing it, anyway, and pushing graphics rendering further. But perhaps timing and technology could have been better? But I guess we all just better wait for actual performance reviews, right?
Source:
NVIDIA via Seeking Alpha
During that Q&A, NVIDIA's Colette Kress put Turing's performance at a cool 2x improvement over their 10-series graphics cards, discounting any raytracing performance uplift - and when raytracing is indeed brought into consideration, she said performance has increased by up to 6x compared to NVIDIA's last generation. There's some interesting wording when it comes to NVIDIA's 20-series lineup, though; as Kress puts it, "We'll start with the ray-tracing cards. We have the 2080 Ti, the 2080 and the 2070 overall coming to market," which, in context, seems to point out towards a lack of raytracing hardware in lower-tier graphics cards (apparently, those based on the potential TU106 silicon and lower-level variants).
It could also be a simple case of it not being feasible to include RT hardware on smaller, lower performance GPUs. As performance leaks and previews have been showing us, even NVIDIA's top of the line RTX 2080 Ti can only deliver 40-60 FPS at 1080p in games such as the upcoming Shadow of the Tomb Raider and Battlefield V (DICE has even said they had to tone down levels of raytracing to achieve playable performance levels). Performance improvements until release could bring FPS up to a point, but all signs point towards a needed decrease in rendering resolution for NVIDIA's new 20-series to be able to cope with the added raytracing compute. And if performance looks like this on NVIDIA's biggest (revelaed) Turing die, with its full complement of RT cores, we can only extrapolate what raytracing performance would look like in cut-down dies with lower number of RT execution units. Perhaps it really wouldn't make much sense to add the increased costs and per-die-area of this dedicated hardware, if raytracing could only be supported in playable levels at 720p.
132 Comments on NVIDIA's 20-series Could be Segregated via Lack of RTX Capabilities in Lower-tier Cards
That still wouldn't bother most but needs more consideration going forward.
Oh and that was 20 years ago not 10;).
English is the only German I'm willing to know.
I should add that I'm not overclocking this 7820x I own as of yet. I'm happy with Intel's turbo boost these days. And haven't touched the Vega. Tbh, I don't know a thing about it... I heard they should be undervolted.
I don't see the point of OC anymore (both CPU and GPU) with all the automatic stuff going on. Of course I'm talking about performance, not just having fun from the process.
As for unlocking GPUs and so on... I don't think there's literally a single purely software feature of Quadro that I miss.
But sure: if it was possible to buy a non-RTX card and unlock tensor / RT ones, I would at least meditate on the idea.I guess people in Europe have different perspective. ;-)
And personally, since I've been working for 2 Austrian companies for the last 7 years and now moved to a German one... Knowing the language wouldn't hurt for sure. :-)
...and I do use the open source AMDGPU (not pro,) driver.
That bloated non-native Gallium AMD driver is better than the proprietary Windows one.
What's your problem with proprietary software? Have you ever considered a possibility that it might be the better way to go? :-)
I don't get this obsession with everything having to be open-source or free.
Thankfully, you're still allowing PC part manufacturers to make money! :-D
The vast majority work well either way. And AMD has their own handful of optimized games (mostly from console land.. like Forza).
FreeBSD is truly free imo. So free it'll let you commercialize and then close out your own source.
GNU demands you be a neckbearded freak in Stallman's cult.. and name your firstborn after him.
So, tl;dr: I don't have an issue with proprietary software so long that it is balanced by an appropriate amount of open documentation and code. That means not being completed closed and pulling the kind of shit that nVidia does.It's the free as in freedom or free as in beer thing.
I'd agree though that Nvidia is problematic. But honestly, not enough people give a shit.
Nowadays buyers really should pay attention to which games they play because it makes a MASSIVE difference in performance. For instance I play a TON of Battlefielld, and my other favorite games (on PC) of recent years were Wolfenstein II, Deus Ex, Far Cry, and Fallout 4. With my 20% overclock on Vega, I beat even aftermarket 1080 Ti's in almost every game I play! Even Fallout 4 typically runs very well on AMD if you simply turn godrays down (and it's an easy to run game now either way).
However if someone played a lot of games like Frostpunch or PUBG.... Even after overclocking my Vega would likely lose to a 1080! That is an incredibly weird variance in performance...
Tensor and RT cores are specialized on particular problems, but again - they can't be called "pro" just because they aren't designed for basic rendering.
The traditional approach was to make a homogeneous chip and push all instructions through the same core structure.
Nvidia slowly switches to a more effective idea of building a GPU from different elements specialized at different tasks.
We knew from the start that tensor cores could help with rendering and now look: with RTX some studios started to implement AA this way (and it's way faster).
The RT cores are still a mystery to me. It will be interesting to see what else they can do. I mean: even while gaming, if you decide not to use RTRT, they could do something else and boost performance instead of slowing it down.
Looking at the general idea of how ray tracing works, IMO there is a chance that RT cores turn out to be great for collision detection, which means taking over a huge chunk of CPU gaming load.But why is that bad? I could call that an issue, if Nvidia's "closed" software was rubbish, but it isn't.Well, I don't care about device compatibility which makes these choices a lot easier. :)
And it's not about being faster for computation. CUDA is just way nicer to work with and, as a result, much faster to code.
From what I've seen, it's also much easier to learn for people without straight programming background.
OpenCL ticks all the typical FOSS boxes - both good and bad. Generally speaking, I'm not a huge fan of the direction open-source is going lately, to be honest.
You can't update the system because the lazy Nvidia team still doesn't have a new driver for the most recent kernel, and since they don't provide any documentation, the kernel driver is shit.
Both AMD and Intel have great kernel drivers, sometimes even better than the Windows ones, while in Nvidia you are limited to geforce.com
Of course this point is moot if you only use Windows, where the only option is the Nvidia driver, support is faster, and the kernel gets updated a lot slower.
That said, I try not to use the latest kernels nor software versions, so maybe I just miss this kind of problems (that's the whole point of using "stable" releases, anyway).
There is no excuse to stay put on old software because Nvidia takes it's time to upgrade the drivers. Not everyone uses Ubuntu.
:D
1) a f*ckup happens from time to time
2) obviously there's some delay before packages get updated. You should be used to that.
Earlier I meant stable Linux releases. If you want everything to work, just stay with main LTS releases of whatever you use. It's that simple.
LTS kernels get the important security updates anyway, so there's little reason to update.
My Debian is still 4.9 and Mint is 4.15 (Ubuntu 18.04 LTS).
I mostly use Manjaro these days and, since it's Arch-based, it already got the latest kernel. But I use it in VMs only, so no GPU issues. :-)Exactly. Back in the day GPUs differed massively. 3D was young and each company had its own approach. Every major GPU launch meant new features rather than more fps, so a bit like with RTX now. Also the APIs were less high-level and were often updated to support new games and GPUs.
Honestly, it's a miracle that games actually were cross-platform in the 90s. :-D
Today you can literally change the card to the other brand and launch some games on generic drivers. It's great how things have changed. I mean: installing a game 15-20 years ago often was a day-long process of rebooting, fighting with DirectX compatibility and editing config files (especially if you didn't have Internet access). Today you click "install" in Steam or whatever and go for a walk.