80 comments

[ 2.5 ms ] story [ 152 ms ] thread
I own a RX 6600 XT, I bought it for ~$200, best card ever for 1080p, works out of the box on linux

A bit sad about the TDP, from 160W to 190W, but the price remains cheap at $329

I plan to upgrade to a bigger screen so this looks like the perfect upgrade for me

Just curious if you're using it for gaming or other uses? I find it wild that $300 cards are the cheap end now when buying one for that 10 years ago was spending over 50% of my builds price on an upgrade part. Normally I find used eBay mining cards and run with them for four years.

What distro are you using that it worked out of the box? I threw my 580x on an ubuntu fork a few years ago and had a hell of a time getting it working.

> I find it wild that $300 cards are the cheap

The natural consequence of printing a ton of money

> I threw my 580x on an ubuntu fork a few years ago

I’ve found AMD cards to be the most stable under Linux

I use it for both gaming and gamedev

My gaming setup runs on Arch Linux, i had a RX 580 just before that one, i don't remember having issues setting it up

1. (According to my short research) 300$ in 2013 are a bit below 400$ nowadays, so you also kinda gotta take that into account.

2. Low-end GPUs have been/are being attempted to be mostly replaced by APUs (at least on the AMD side) which kinda makes sense.

3. Normal pricing has gone outta the window ever since the NFT craze (or demand if you wanna put it that way) and now the AI craze.

Sorry for off topic.

Wolframalpha makes it easy to compare USD accross years and uses official inflation numbers. "2013 $300" [0] would be worth $386 now (2024).

If you use ddg, that's "!wa 2013 $300" or if you want it in reverse "!wa $300 in 2023"

[0] https://www.wolframalpha.com/input?i=2013+%24300

Didn't know WA actually did that, but not too surprised.

(Also, you writing it out again I kinda forgot that it's 2024, not 2023)

Interestingly, it only seems to work if you force US dollars. When it defaults to Canadian dollars for me it tries to multiply $300 by 2013 inches.
It should be basically any distro. The driver is part of the kernel.
You can also buy a used card. Most mid+ cards from the last 5 years are quite capable. I got a 5700 XT to play some games and it’s I think better than any console you can buy? Cost $170 I think. Can do 1440p in most games still
Went from 480 to 7800 on Arch, both Just Worked (install Mesa and that's about it). On the other hand I tried like 3 different 5600s a few years ago, and it was unstable enough that I got it replaced with a 580 (basically a rebadged 480).
could you just throttle the card down to 160W?
Very easily, AMD has been releasing cards overvolted out of the box for ages, playing it very safe.
GP was not suggesting playing with voltages but simply power limiting the card, which tends to be easier to do and doesn't affect stability but has a performance hit.
I’ve got a regular RX 6600 (non-XT), and it’s been fantastic for the mostly older and/or indie games I tend to play, even at 1440p.

Eager to know how this will compare. I’d like something beefier than what I have now, but what used to buy you a whole upper-mid-range PC barely covers an upper-mid-range GPU.

I use the non-XT with dual WQHD monitors with no issues. But I also don’t play any shooters or other action games.
Could it be a good card to test LLMs at home?

16GBs of ram don't come out often at that price range.

Is the typical limiting factor for local LLM performance the amount of VRAM?
I have MBP M1Max 32GB - it is $20000 now very good for LLM.
I was surprised on how well Mistral's `mixtral` model runs on my work's MBP M1 Max 64GiB. I thought it'd be a total slog but it does work well enough to replace ChatGPT for quite a lot of use-cases.
Does it spin up the fan?
Yes, heavily, I think it's the only time I heard the fans loudly with this machine.
[flagged]
I can't down vote, but I don't understand how your post answers the question, perhaps because I've never owned a Mac. Are you implying that your MBP also has a lot of GPU RAM and therefore, yes, it is good for LLMs?
yes RAM is shared between CPU and GPU, I am able to run Mixtral-8x7B in this machine at decent speed, this is pretty good for testing LLM, yes
Macbook Pros don't cost $20,000 and aren't cards...
oh thanks! I am getting down votted for a typo? reminds me grammar nazies during slashdot days... okay
I didn't downvote you, but someone discussing an article on $329 graphic cards and asking for a card recommendation probably doesn't want a 2000 dollar laptop recommendation.
How's that relevant to my question?
Meanwhile, the AMD Radeon RX 7600M XT GPU exists and is used in both new portable eGPUs: the GPD G1 (which I own) and the ONEXGPU which is being crowdfunded right now. It is a slightly more powerful version of the 7700S. I am sure this naming makes sense and it's only me who misses something.
One of the reasons I never ever consider AMD hardware when I'm looking to buy something is because the naming schemes do not make any sense.

For example: Did you know some Ryzen 7000 CPUs, introduced with the Zen 4 architecture, are Zen 2 and Zen 3 and also Zen 3+?[1] Did you know Ryzen 3000 spans Zen, Zen+, and Zen 2?[2]

And here's the kicker: Ryzen 5000 is the 4th generation of Ryzen CPUs, based on Zen 3 which is the 4th generation of Zen, except the Ryzen APUs which are based on Zen 2 which is the 3rd generation of Zen.

Oh by the way, Zen 4 is the 6th generation of Zen.

Yeah. I very sincerely ain't got time for this fucking nonsense.

[1]: https://en.wikipedia.org/wiki/Ryzen#Mobile_6

[2]: https://en.wikipedia.org/wiki/Ryzen#Ryzen_3000

Is Intel or Nvidia even better about this?

> Yeah. I very sincerely ain't got time for this fucking nonsense.

aye, is also most of what you said. if following a naming schema is too hard for you mon ami then stop buying hardware independently and pick up whatever Alienware is releasing this quarter.

>Is Intel or Nvidia even better about this?

Surprisingly: Yes.

They (usually) don't label inferior/older tech as new tech, for one. If Intel says something is 12th gen or 13th gen or 14th gen, it generally is. Likewise Nvidia, if they say something is 3rd gen or 4th gen, it generally is. This already greatly simplifies things; numbers actually mean something.

Then they don't have more suffixes and prefixes than I can be bothered to remember. Intel? The first number (eg: i7) denotes class within generation. First one or two digits in the second number denote generation (eg: 12 in 12700K). The third number denotes tier within the class (eg: 700 in 12700K). The suffix, if any, denotes overclocking (K), supremely-special (S), no graphics (F), powerful mobile (H), desktop repurposed for mobile (HX), power-saving (P), and ultra power-saving (U).

For Nvidia, it's either GTX or RTX prefix for no raytracing or raytracing enabled respectively. Then the first one or two digits in the number denotes generation, and the following digits denote tier within the generation. Super or Ti are appended for higher tier models of the same number.

Meanwhile, what the sincere fuck is AMD?

> If Intel says something is 12th gen or 13th gen or 14th gen, it generally is.

10th and 11th generations included both 14 nm and 10 nm CPUs, now 14th gen is in some weird state when it coexists with Core Ultra on Intel 4 (7 nm) chips.

> The first number (eg: i7) denotes class within generation. First one or two digits in the second number denote generation (eg: 12 in 12700K). The third number denotes tier within the class (eg: 700 in 12700K).

So what is i7-1065G7 then? Or Core Ultra 5 135U?

> Then they don't have more suffixes and prefixes than I can be bothered to remember.

You forgot at least T, XE, N, G. That's only the ones I remember, there are definitely more.

>If Intel says something is 12th gen or 13th gen or 14th gen, it generally is.

14th gen is 13th gen with a minor clockspeed boost on desktop.

Don't forget NVIDIA trying to sell the now named 4070 as a different 4080.
When you get down to the lower marketed brands, like Pentium and Celeron, and sometimes i3, things get really messy.

If it's got a suffix, it's a core core (having to say core core is of course, an Intel naming problem too), and if it's got a prefix, it's an atom core. And of course, the numbers are offset from the other brands for no particular reason.

Intel is definitely on the same level. 10-th gen Intel Core had two different architectures for laptop chips - both 14 nm Comet Lake and 10 nm Ice Lake. 11-th gen Intel Core was 14 nm on desktop, but 10 nm on laptops. Now Intel 14-th gen is 10 nm Raptor Lake Refresh on desktops and laptops, but also 7 nm (Intel 4) Meteor Lake on a specific laptops, as Core Ultra.
> Is Intel or Nvidia even better about this?

I remember when Intel brought out some new mobile processor and didn't mention it's entire name ONCE. https://www.youtube.com/watch?v=aFHBgb9SY1Y&t=17s (This is just the most memerable part. The entire video is an interesting watch)

Then they also recently (december) brought out slides about "Core Truths" in which they actually explained AMDs naming scheme decently well, while completely ignoring their own mobile naming scheme (they removed those slides very fast, but here is a video covering it https://www.youtube.com/watch?v=xUT4d5IVY0A)

Not to say AMDs naming scheme is completely faultless (e.g. they could have at least move the architecture number from 3rd to 2nd index), but I'd rather have an explained consistent naming scheme, than one that changes on how they are feeling that day (https://www.intel.com/content/www/us/en/processors/processor...).

(Also, as a general tip: don't think too much about the naming and treat them like almost anyone does on TVs/Monitors, aka just as product identifiers)

From Ryzen 7000 forward, the third number is the core generation, and this weird mixing is only true outside desktop too (see the 5600X3D).

That's also because AMD keeps die shrinking older generations of processors (i.e. the Steam Deck), but not as far down as Zen 4 because they're still capacity limited there. So it makes sense to keep 6-7nm designs alive for longer.

Intel is not any better with their increasingly long number names. I guess Apple so far has followed a logical numbering scheme but they only have a few SKUs compared to hundreds from AMD and Intel
You don't buy a naming scheme, you buy a particular CPU.
This is my next upgrade from an Nvidia 1660.
Looks like the power draw could be almost double your 1660. Better check to see if you need an upgraded power supply.
Chat, what is your preferred AMD GPU vendor? I have had mixed results. My biggest gripe is acoustics - squealing, screeching, coil whine, etc. I've had mixed results with PowerColor, terrible results with Sapphire. I have an ASRock Taichi card (7900XTX) that is so large the pcb/frame began to bend and that introduced a noise into the fans. The support bracketry that came with the card was missing all hardware, so I ran without it for a while. Perhaps that is on me. I supported it with a snus can and it has slowly cooked itself back to being straight. I always prefer the aesthetic of the reference cards, but believe the aftermarket manufacturers do a better job with cooling.
I've personally had a good experience with XFX cards. I haven't had that many (not a hardcore gamer), but they've all worked perfectly well over multiple years, and never developed any kind on noise. The youngest is a 5600XT I bought used in 2020.

They have never been top of the line models, but one or two steps down. I've always used "silent" cases, since I hate noise with a passion, but I'm not sure how much that actually changes things.

I’ve also had good luck with XFX cards - even used ones. But this is wildly anecdotal of course
You might have a power supply issue if you get coil while this consistently. Very happy with my Sapphire Nitro 7900 XTX.
I have a Sapphire RX 5700 and it's basically fine, although I don't ask much of it.
I have had Sapphire cards for the last 10+ years and have always liked them enough to go back. Maybe I've been lucky, but my cards have been cool and quiet.
On my third Sapphire, just upgraded to a 7800xt. Excellent hardware QC, never had any issues beyond AMDs occasionally terrible drivers.
Every sapphire (3) I have had has failed. I luckily got one rmaed but that starting freezing my computer after a year.

Switched to XFX Merc 319 6950xt. I hope this serves me better.

If you can then buy "founders edition" cards.

I have an RX 6700 XT from AMD themselves and it's great.

The reason I went down that route was that the AIB partners all beef the cards coolers up and thus the dimensions. I build in small SFF cases for travel reasons and usually end up undervolting and downclocking my cards so I really don't need the monster coolers the AIBs are so fond of.

Unfortunately they generally only seem to be available in limited quantities and for a limited time which is a real shame.

My 6900XT FE has _terrible_ coil whine. I can troubleshoot some types of issues with my models by ear.

It's so bad I thought my card was defective the first time I heard it. It's usually fine if I cap framerate to a reasonable number though.

Maybe look at undervolting, you can do it from the AMD software pretty easily and for me it solved a bit of a coil whine problem under boost scenarios but YMMW it depends on sample to sample I think.
I can tell how close stable diffusion is to finishing based on the current noise pattern it is almost useful
In your case, it seems that investing in some case-specific support bracket for $10 would have prevented all issues, so I'm unsure why you propose vendor switching as the solution.

After experiencing coil whining with one game and loud fans after extended perios of full load on the system (CPU+GPU), I switched to an aftermarket all-in one liquid cooling system, like Alphacool Eiswolf. I just think that the standard 3 fans blowing air onto the GPU kill most useful airflow in every PC case and with every vendor that uses this style.

I've had issues on many cards, this was just one example.

Every GPU I have ever had, particularly AMD, have been noisy cards. I have hyperacusis and can hear a butterfly fart a block away, so silence is paramount to me.

I don't even allow deep sleep states on my Intel processors because I can hear it. Modern motherboards are beginning to include "acoustic adaptations" to try and fix, but they rarely work in my experience.

What helped me with the fan noise for extended periods of CPU + GPU load was moving the GPU one PCIE port over. Initially, the GPU was very close to the CPU's heatsink (millimeters away).

I know this isn't always possible with newer CPUs which tend to be stingy with PCIE lanes, but in my case, I have a boatload of them, so I could do it with no issue.

On my last build I completely bypassed the issue. I minimized noise and improved thermals by deshrouding the GPU (Zotac Amp 3080) and using 120mm Noctua fans for cooling, as shown in this video: https://www.youtube.com/watch?v=QUaZVpN51Po
For anyone looking to do this, there are much better 120mm fans available nowadays, like Alphacool Stealth Metal.
Interesting. Apparently the Alphacool got a great review from Igor's Lab, but der8auer was not able to show substantial difference compared to Noctua: https://www.youtube.com/watch?v=8J7jkRbxLYk

This is based off of me doing a 1 minute search for a comparison, so I could be missing something completely. I haven't really researched things in depth since my last build 3 years ago.

Sapphire for AMD.

EVGA used to be go-to vendor for Nvidia but since they're gone I'd settle for Gainward and its derivatives (Palit, Galax, KFA2)

My past two AMD cards have been Sapphire. First, an RX 5700 XT Nitro+ and second an RX 6900 XT Nitro+ SE.

The first was used on a machine on my desk while the latter is in a living room gaming machine. It’s harder to judge the 6900 because of the increased distance from it while running but the 5700 had no coil whine.

The EVGA 3080 Ti FTW3 in the machine at my desk on the other hand can be quite audible at times, particularly during some menu screens in games. It’s a great card otherwise.

Sapphire is the safest bet. ASRock, MSI and Powercolor also tend to make good cards.

And yeah, with the monster-sized high-end GPUs, you'll need to take some precautions no matter the brand.

Basically 12ish years after the Radeon HD series with the GCN architecture we return to the 7XXX numbers. :-) https://en.wikipedia.org/wiki/List_of_AMD_graphics_processin...
*3rd time.

Not sure what became of my Radeon 7200 but I have a sweet Radeon HD 7770 (GHz Edition!!!) that still works.

My ASUS Radeon HD 5850 Cypress Pro is still going strong! Never any issue ever in the years I've had it.
The thing that’ll kill it is games moving to DX12/Vulkan only which that card doesn’t support.
There are rumors AMD won't do high end cards anymore. So 7900xtx becomes the last one to compete with NVidia on that front.

Honestly, after owning 7900xtx for one year, I find it difficult to saturate it with tasks. I only play non-graphics heavy games like rimworld or aoe2de. Sometimes I play Hogwards Legacy with my partner, and the performance there is not the highest imaginable. But I read online that is because the game is not well optimized. Perhaps, you don't even need more powerfull graphics, since games are oriented at mid-market anyways.

Game can run like crap even on RTX 4090 if you turn up settings all the way up
>There are rumors AMD won't do high end cards anymore.

No there aren't. There are rumors AMD isn't doing a high end card for the next generation. That isn't the same thing.

In any case, graphics cards are improving faster than the ability to make use of the power, especially with upscaling technology. Now that "midrange" cards are able to hit 144hz at 1440p with maximum graphics, the market for high-end is probably going to dry up a bit, at least until full path tracing is viable, or 4k monitors get cheaper.

What would make sense for AMD is to spend less time on fancy hardware architectures (dual-chip GPUs) and more time focusing on getting AI working well, improving their software stack, and not having issues like the "power bug" that supposedly held back the 7000 series GPUs from reaching their intended performance goals. I expect that's what they're doing.

The rumor is that RDNA 4 is a mulligan gen for AMD while they get chiplet based GPUs ready for RDNA 5 in 2025. While Nvidia is going fully monolithic for Blackwell and won’t be using chiplets until at least the 6000 series. So there’s a chance AMD could actually pull out in front of Big Green in two years.
I’d like to know how it fairs with 8K or 4K. I have an 8K 60hz Dell monitor which my Nvidia 4090 runs games on beautifully.

But the real dream is the tv which plays 4K at 120hz.

If you at least want to future proof your graphics card, surely you’d want to know about 4K support at the very least.

Damn that's 8k at 32 inches? Do you use that for 4-5-8k video editing or so do other use case? 4000$ monitor
The Dell is mostly just for productivity as there is a tonne of real estate. I have ADHD and I get totally distracted when I don’t have all the windows on the same screen, then my chain of thought slips and I can’t remember what I am doing. I still get this on the 8K screen, but it’s more manageable.

Of course, I cannot go to native resolution because the windows are so small I can only read anything - 200% is good for me. I found it 2nd hand off eBay for a good bit less than what you said.