AMD Radeon RX 7900 XTX and RX 7900 XT Linux Support and Performance Review
phoronix.com
phoronix.com
Over the last few generations of AMD GPU; I have run Windows, Linux and a MacOS Hackintosh - all with seamless hardware acceleration, relatively good power consumption/thermals and with a great fps per dollar compared with competing products.
Add to it that AMD was instrumental in the creation of Vulkan, which is the graphics API that made gaming on Linux possible (and gaming on M1 MacOS easier). They created FreeSync and FSR which is open to all hardware and honestly pretty decent.
I'm no hardware evangelist but I really respect AMD's consistent drive to improve the industry and it has certainly left an impression on me.
Last time I checked ROCM is light years away from CUDA.
1. World's most powerful supercomputers run on AMD Radeon Instinct cards, which means more funding for better ML stack from AMD.
2. A ton of ROCm libraries developed by AMD has landed recently in Debian, which means there's actual and fast paced development work being done on this stack.
Source: I work on the llvm part of the stack and don't have enterprise hardware. I don't know what the library developers use but expect some of them have similar setups.
I suspect the ROCm situation has not improved since last time I checked, based on their documentation. The installation guide [0] is a bit of a mess. It looks like they are targeting ROCm at existing ROCm users rather than finding ways to convince people to become ROCm users; nothing there is useful introductory material. The opening is basically marketing babble too. Or possibly someone with a history in the GNU projects lineage of terrible recursive names.
It just doesn't look like they've got a leadership figure trying to bring people in to their ecosystem right now. I dunno, maybe the API isn't intended for interested developers.
[0] https://docs.amd.com/bundle/ROCm-Installation-Guide-v5.4/pag... - opens with "ROCm is a brand name for the ROCm open software platform (for software) or the ROCm open platform ecosystem (includes hardware like FPGAs or other CPU architectures)".
So, while the documentation still doesn't make it trivial to get started, the software side definitely seems to have improved in recent years.
ROCm builds are effectively hardware-specific (and officially Linux-only).
With a better inference standard option hopefully it would run on AMD too. But I also hope AMD gets their compute to a better place, preferable using open standards so we don't need "Nvidia code", "AMD code", etc.
I'm a techie and I don't think I have a use for it.
Looks like it's still early days, but clearly it has recently been made to work and there's an effort going on to make it better.
It's good enough (for most user) that you can just install the binary user component driver and run the card without limited to random linux distro or version.
Imaging you can't even do any windows update because nvidia card may have compatibility problem and turn your desktop black. It's the status that nvidia on linux now.
Though even with European electricity prices, the generally higher (real world) power consumption of these doesn't make the (more expensive) 4080 a better deal. Maybe if prices cross over the 1 buck/kWh threshold. Never say never!
The smart move is to wait a couple months and see if AMD can fix the bigger issues with these cards through drivers and only buy one of these if they do.
That's pretty much it.
The assumptions/rumors/leaks and such right now[0] point to underwhelming release drivers. Like they aren't taking advantage of the hardware, or have clutches in place until they can get higher utilization of the hardware to be stable.
By CES, there'll surely be announcements of other RDNA3 cards. By the time these get reviewed, drivers should be closer to where they should be, and informed decisions much easier to make.
Right now, the 20% higher priced 4080 gets you 17% better RT, while raster suffers -8%. Therefore 7900XTX is better value, just not by as much as everybody hoped. This might or might not change with newer drivers.
I do not recommend buying any AMD RDNA3 nor 4000-series NVIDIA right now. Waiting to see what happens once lower-end cards enter the market is my advice.
... matching the 4090 in Cyberpunk, while drawing less power.
They did similar with 5000. So no doubt things will improve with 7000 series.
For power efficiency, considering that Ada is a monolithic TSMC N5 variant (4N) die while Navi31 is a combination of N5 + N6 for the 6 MCDs, maybe it's not so surprising... One thing to note is that historically, GPUs companies are almost always pushing towards the far end of the VF curve for their top-tier cards. For example, while the 4090 has gotten tons of flack for it's power consumption, cable melting, etc, a Korean review site did testing on power limiting and undervolting while gaming [1] and found that you could basically switch to an 80% Power Limit (only 21W of power savings though) and lose <1% performance. Alternatively, with a slight undervolt, you could shave 58W of power consumption while lowering performance by 1.2%, or 95W (from the 347W stock) with 4.6% lower performance.
As to the smart move, IMO from a price-perf perspective, all of these new cards are a bit out-to-lunch. For gaming, high-end last generation cards (like a 6800XT at $600) will get you to 4K60 on almost all AAA titles and are a fraction of the price of the Radeon 7000 or RTX 4000 series. Also, while I'm a big fan of open source drivers, those doing ML are basically forced to buy Nvidia atm.
[1] https://quasarzone-com.translate.goog/bbs/qc_plan/views/3068...
[0] https://www.phoronix.com/forums/forum/phoronix/latest-phoron...
100 W just the card or the whole system? In Linux? Because I've seen ~20W windows numbers for just the card for AMD. Mind, less powerful cards.
NVidia does do better for multimonitor desktops tho.
I'm currently browsing on one monitor and having a twitch stream with GPU accelerated decoding on another, and nvidia-smi reports 49 W for my 3080. Nvidia 525 drivers & Fluxbox as window manager.
Using an AMD 5950X CPU, currently with powersave governor, my PSU reports a total power usage of 120 W.
Without competition Nvidia will continue to drive prices sky high.
And let's hope Intel manages to stay in the game for the same reasons.
It's hard to get my head around a top end video card costing AUD$3,000.
Even AUD$1,500 seems a ridiculous amount to spend on a graphics card.
On paper yea, but in practice it isn't really though. AMD 7900 series are cheaper than the super expensive Nvidia 4090/4080, but only because they're weaker in several categories, so you're not getting any real benefits for the same money. Less price but also less performance.
If you're doing DL/ML work, then Nvidia is still the only choice for most consumers/businesses.
Also, industry insiders are reporting that the AMD 7900 reference cards at MSRP are in very limited supply, with most going to the US, and that once they sell out, which they will, the ABI board partners will bring a $100 price bump to bring them closer to Nvidia.
Gamers aren't getting the salvation they were hoping for this time from AMD. It's not really a competition anymore, AMD is simply joining Nvidia in the club of price-gouging consumers too.
You're getting identical rasterization performance at a lower cost.
Now, that cost is still obscenely expensive and the 7900 XT (non X) is also terribly priced for the performance. So I'm not defending that.
But I also don't use raytracing personally because in the current generations it just costs too much. I don't want to degrade and noise my image with DLSS/XESS/FSR/whatever just to get raytracing. So raytracing performance is low on my priorities.
Not that lower if you go beyond the MSRP which most people won't see in stock. Industry insiders are reporting ABI prices for the Radeon 7900 XTX on the shelf to be $100 higher than MSRP, while Nvidia is reportedly lowering the 4080 price by 100$ to meet them in that same $1,100 price-point.
So in the end 7900XTX and 4080 are "the same performance for the same price" if you restrict yourself only to rasterization performance as your only metric and ignore the current bugs the AMD cards have and the extra ML/DL features Nvidia cards offer.
They're still both $1,100 GPUs, far beyond what most gamers are willing to spend on a not-the-highest-end GPU, which makes me think they'll still not sell, as the whales are still going for the $1,600 4090 which is selling like hot cakes while the value chasing gamers are waiting for something much cheaper or buying cards one or two generations older at steeper discounts.
For my life circumstances, I'd tend to agree.
But if I were fabulously wealthy, or if the 4090's were for some reason the ideal cards for my machine learning work, it might be reasonable.
Yeah OK
I trust the word of the developers.
My experience has been great so far, with emulators for recent consoles.
* system crashes if I turn on my monitors at the wrong time when waking from sleep. One trigger seems to be a monitor powering on while I'm logging in. Both screens go black and I'm forced to do a hard reboot.
* flickering and parts of the screen going black, which seems to be related to video playback but occurs at other time (for example, while typing this comment)
I didn't have either problem with the NVIDIA card that I used previously.
Yeah, that was ATI, pre-AMD.
This might be your problem - the kernel that ships with Ubuntu is often very old.
My 5700XT ran like crap under Ubuntu - switching to a distro[0] with the latest kernel fixed all my problems.
You can't be mad at AMD because Canonical ship an old kernel!
[0] Any distro with a recent kernel will do. I picked Gentoo, but that's irrelevant.
Ubuntu's stack is just prehistoric.
If you never hear from me again, know that I don't regret bricking my machine and I'm sure 2023 will be the year of Linux on the desktop.
I had a similar issue with my previous GPU, an RX560, where my system would sometimes freeze either while turning the monitors on or other times when it was locked with both monitors off.
On the 560 I had no choice but to connect 1x DP + 1x HDMI but on my newer 6700xt I connected both monitors via DP and the whole class of issues just went away.
Unfortunately, a lot of things don't get shaken out before release because of limited access to hardware. Within around 6 months things are generally a lot better. I wound up switching to a 3080 as I wanted to play with RTX and CUDA, thinking of jumping to an RX 7900 XTX in March/April if prices are still competitive. It won't be as good as 4080 for RTX, but similar to 3080 performance, which for me has been usable.
I suppose it shows the investment split between platforms, where Nvidia win on Windows and AMDs fully OSS strategy working out on Linux.
I guess fixes for AMD based platforms like the Steam Deck help these newer cards too
My only gripe is that, despite these being better value than Nvidia, GPU prices are still creeping up. The technology future I was sold was one where technology becomes cheaper over time, not more expensive.
And it is suddenly a hot cake now because nvidia introduced dlss3???
It's so confusing from my perspective.
In my opinion. Frame generation only looks good when use with texture without too much detail, it looks shitty when use with complex(realistic) textures.
The RRP of the XTX is $1000, while the 4080 is $1200. The delta is at least $200, but more once tax is included. Sales tax is 25% here in Europe, so the delta is even higher.
People need to get over "NVIDIA = AI" hasn't been true for at least a year.
Contrary to popular belief, high-end cards actually are priced appropriately. It's the low to midrange cards that are ludicrously overpriced; their price-performance ratios make them terrible investments.
The RTX3050 (250) is 2 times more expensive than the 1070 (125) and is 10% faster than that card.
The RTX3060 (400) is 3 times more expensive than the 1070 (125) and is 40% faster than that card.
The RTX3080 (750) is 3 times more expensive than the 3050 (250) and is 300% faster than that card.
The RTX4090 (1600) is 2 times more expensive than the 3080 (750) and is 200% faster than that card.
The high-end GPUs are currently vastly cheaper per unit of performance than the midrange cards and have been for the last 2 years (the 3080 was twice the price of the 3070, but had twice the performance). And that completely destroys the value proposition (and PCMR cache) of PC gaming over just buying a console (IIRC both the newest ones have the functional equivalent of an RTX2060).
And so it shouldn't be a surprise that people continue to bitch about "yet another overpriced card". What the builder market needs is a 3080 competitor at the 500-dollar price point to forcibly correct the pricing of the cheaper cards down to the 200 dollars the secondary market reveals they're actually worth; once that happens, the pricing structure will likely begin to return to normal. (It's also weird that nobody talks about this.)
Of course, there's always the possibility that the RTX 5000 series doubles in speed (and price) again; after all, not only have gamers proven they'll pay for 3080s at massive markups, but companies training ML models will pay top dollar to cut the time that takes in half. Intel seems to have bet on it; taking advantage of in-house fabrication + that market gap is what Arc was made to do, but it only survives provided current price/performance ratios stay the way they are for a little while longer.
Just for the sake of clarity I understand you to mean "200% as fast as," (as in twice as fast) and not "200% faster than."
I agree with your premise that cheaper cards are even worse value, but taking this example of the best value of this generation, it's not cheaper per unit of performance. It's exactly the same as the last generation. GPUs (and indeed most technology) has been getting cheaper per unit of performance over the years. To me, this marks a radical departure from my expectations.
I like what I am seeing in regards to the new AMD cards vs. my current RTX 3080 for gaming performance. Especially since Nvidia has raised prices to a point I am no longer even begrudgingly willing to pay. But, as I am currently playing around a bit with stable diffusion I am more inclined to stay with CUDA for the time being. However this is a very recent change in interests, so I am not sure whether my priorities will change long term. I am also not planning an upgrade any time soon, so I will wait for new developments on all sides.
At work we have 20+ workstations running Nvidia on Linux, and they are such a pain in the ass. The idiots at my place don't even consider hardware-software compatibility issues before making the purchases, makes me mad.
Nobara which is a patched version of Fedora published by Glorious Eggroll includes it: https://nobaraproject.org however no secure boot if you think its important.
For streaming, would guess that the new 7000 series encoder is at least comparable to the RTX 3000 series.