Today AMD formally revealed the next-generation Radeon GPUs powered by the RDNA 2 architecture and it looks like they’re going to thoroughly give NVIDIA a run for your money.
What was announced: Radeon RX 6900 XT, Radeon RX 6800 XT, Radeon 6800 with the Radeon RX 6800 XT looking like a very capable GPU that sits right next to NVIDIA's 3080 while seeming to use less power. All three of them will support Ray Tracing as expected with AMD adding a "high performance, fixed-function Ray Accelerator engine to each compute unit". However, we're still waiting on The Khronos Group to formally announce the proper release of the vendor-neutral Ray Tracing extensions for Vulkan which still aren't finished (provisional since March 2020) so for now DirectX RT was all they mentioned.
Part of the big improvement in RDNA 2 comes from what they learned with Zen 3 and their new "Infinity Cache", which is a high-performance, last-level data cache they say "dramatically" reduces latency and power consumption while delivering higher performance than previous designs. You can see some of the benchmarks they showed in the image below:
As always, it's worth waiting on independent benchmarks for the full picture as both AMD and NVIDIA like to cherry-pick what makes them look good of course.
Here's the key highlight specifications:
|RX 6900 XT||RX 6800 XT||RX 6800|
|Process||TSMC 7nm||TSMC 7nm||TSMC 7nm|
|Game clock (MHz)||2,015||2,015||1,815|
|Boost clock (MHz)||2,250||2,250||2,105|
|Infinity Cache (MB)||128||128||128|
|Memory||16GB GDDR6||16GB GDDR6||16GB GDDR6|
You shouldn't need to go buying a new case either, as AMD say they had easy upgrades in mind as they built these new GPUs for "standard chassis" with a length of 267mm and 2x8 standard 8-pin power connectors, and designed to operate with existing enthusiast-class 650W-750W power supplies.
There was a big portion of the event dedicated to DirectX which doesn’t mean much for us, but what we’ve been able to learn from the benchmarks shown is that they’re powerful cards and they appear to fight even NVIDIA’s latest high end consumer GPUs like the GeForce 3080. So not only are AMD leaping over Intel with the Ryzen 5000, they’re also now shutting NVIDIA out in the cold too. Incredible to see how far AMD has surged in the last few years. This is what NVIDIA and Intel have needed, some strong competition.
How will their Linux support be? You're probably looking at around the likes of Ubuntu 21.04 next April (or comparable distro updates) to see reasonable out-of-the-box support, thanks to newer Mesa drivers and an updated Linux Kernel but we will know a lot more once they actually release and can be tested.
As for what’s next? AMD confirmed that RDNA3 is well into the design stage, with a release expected before the end of 2022 for GPUs powered by RDNA3.
You can view the full event video in our YouTube embed below:
Additionally if you missed it, AMD also recently announced (October 27) that they will be acquiring chip designer Xilinx.
Last edited by Shmerl on 28 October 2020 at 4:23 pm UTC
Quoting: ShmerlToo much talk about DX12. They said nothing about Vulkan so far...
While this is bothering me as well, they usually talk about DX "feature levels", because "DX 12 Ultimate" goes off better than "Vulkan 1.2.34103 release 3".
Last edited by Liam Dawe on 28 October 2020 at 4:30 pm UTC
Quoting: GuestIf the driver breaks with a kernel update then one would assume that it's obvious that it is the kernel which breaks its own interface and not vice versa. The kernel doesn't provide any stable interface. Read about the latest breakage: https://lwn.net/Articles/827596/
We are definitely talking about different things then. AFAIK, Nvidia proprietary driver doesn't implement DRM api, so you are installing a piece of software that doesn't follow Linux design guidelines.
Regarding the problem you mentioned, the issue Nvidia has with each new release of the kernel is that some parts of the driver may get broken because the guts of the kernel may change from time to time, which means that if your driver is not mainlined you will have to take care of following and patching your driver in parallel to the kernel releases (there isn't any other way around). But still, mainlining your kernel driver doesn't mean in anyway that you cannot still provide your dkms driver for downloading and installing in older kernels (of course, the company in charge of the module will have to take care of adapting their code for older kernels if necessary). You can see an example of this with AMDGPU-PRO, where they include AMDGPU DRM as dkms for the kernel supported by Ubuntu LTS and Rhel versions.
Last edited by x_wing on 28 October 2020 at 4:42 pm UTC
Quoting: KROMAwesome presentation. But I think I just saw my wallet running out of the door screaming.
So did mine. The 6800 (non XT) is more than enough for my needs.
Quoting: GuestThis is what an amd fanboy would say, not a conscious customer who evaluates a product properly. Nothing you mentioned is actually an evaluation of any aspect,
This is based on the actual testing of 50+ hardware pieces, more than 70 configurations, it is months of testing (and in fact, sometime, years of debugging).
All those tests in that page were driven by developers or under developer supervision, they are not automated. I was involved in all of them in several stages.
Quoting: lunisAs a game developer you develop your software for a specific driver/hardware combo.
We don't develop for a specific driver/hardware combo. You're talking like a web developer from Internet Explorer 6 era.
The specific bug I talked about is a bug where Nvidia driver wrongly announces the hardware supports a feature that is not supported. That's unrelated to how the game is developed or not. A game written properly against the OpenGL standard would query the availability of the feature and would enable code the hardware does not support because the driver made a false statement. So we have to do some guesses to attempt to detect the faulty driver/hardware combinations to not trust what the driver says.
And outside of game development, I have seen Nvidia cards requiring to plug a screen on VGA port to stop the Nvidia driver to complain when using the HDMI port (while not being able to display something on that VGA port anyway). I have seen Nvidia cards displaying the early BIOS screen on one port and the operating system on another, requiring to unplug/replug the screen during the boot process to get continuous display. This has nothing to do with game development. Some cards just disconnect themselves from the PCIe bus with the proprietary driver, not with the open source driver. This has nothing to do with game development.
I had myself to write Xorg.conf things to make Nvidia stuff working on actual hardware, things people using AMD or Intel don't do since year 2008 or so. What a blast from the past!
The thing is that Nvidia products are like dragster: they can shave 1 second over their concurrent time on a straight line, you don't know what happens if you have to take turns, they are not meant to make yourself able to take your children to school or complete the Dakar Rally, and it may just explode in your face before starting the race.
[ 195.564010] NVRM: GPU at 0000:01:00.0 has fallen off the bus.
Last edited by illwieckz on 28 October 2020 at 4:59 pm UTC
See more from me