Video Cards

[ddrueding]

Combined with heatsinks that weigh absurd amounts.

 

[sedrosken]

Everyone's cutting costs, no one can make a product that doesn't kill itself to save their life, all anyone wants is your money and they don't care about making a quality product and getting said money legitimately.

 

[Handruin]

Maybe eVGA did make the right call exiting from the GPU industry. If the chips were costing so much along with all the other issues they had, how could they keep going with such thin profit margins without reducing costs.

 

[Mercutio]

Actually, what IS eVGA doing now? They don't make video cards or motherboards. What's left? Cooling setups?

OK I went and looked and... Liquid coolers, PSUs and gaming peripherals. The margins on a gamer mouse might actually be higher than a mid-range GPU and that's crazy.

 

[Handruin]

I was surprised they can make and/or continue a viable business model on those limited items. Maybe they're raising cash to get into other areas over time to supplement the "loss" of GPUs. Difficult to say if the GPUs eventually became a loss-leader after they established a name for themselves given the thin margins.

 

[Mercutio]

4060 came out. It is not particularly better in any way than a 3060Ti until the AI frame generation stuff is put in play. The A770 and 6700XT are both in the same price range. Intel released better Iris/Alchemist drivers just this week and while the 6700XT is 10 - 15% more expensive, it also has 50% more RAM and it's flatly better than nVidia's $300 card.

Oof.

 

[LunarMist]

At some point doesn't nVidia want to move to one technology and the 30 series is getting old and lacking the DLSS? I read on the webs that they are not expected to implement a 50 series until 2025. If $300 is too high and sales volume are poor, then there will be sale prices. I'd really like to add a card like that to a (currently 5700G) system in a µATX case if it fits fully within two slots.

 

[ddrueding]

Some analysis I read showed nVidia selling anything that gamers don't buy to AI startups. The first time they overpriced the gaming market crypto bailed them out, now it is AI. They just don't need to care about the gaming market, especially not the price-sensitive end of the market where the margins are low anyway.

 

[LunarMist]

The Nvidias H100 Hopper is as costly as a small Honda.
I wonder if there is much profit on them.

 

[ddrueding]

Tons. Typically they have similar architecture to the gaming cards, but with more ram and "more reliable" firmware.

 

[Handruin]

Seemed appropriate

 

[LunarMist]

Seemed appropriate

It seems obtuse, but I'm not up to par on video cards humor.
What exactly are those video cards doing for the detainee?

 

[Mercutio]

It seems obtuse, but I'm not up to par on video cards humor.
What exactly are those video cards doing for the detainee?

Everything in the RTX4000 series can be considered downgraded by one designation based on past standards of performance and pricing vs older generations. The 4080 probably should've been a 4070Ti, for example.

 

[LunarMist]

I thought the 4080 was already downgraded to the 4070Ti that I received in Q1.

I'm not sure how your intelligence interrogates the detainees/suspects but video card is a strange graphic.

 

[Handruin]

It seems obtuse, but I'm not up to par on video cards humor.
What exactly are those video cards doing for the detainee?

It's a meme based on the Scooby Do "Let's See Who This Really Is". The joke being one GPU is being an imposter of the other.

 

[LunarMist]

I watched that show a few times in the early 70s, but it made little sense. It wasn't funny and it wasn't scary. The characters seemed like a bunch of dimwits.

I found this. https://knowyourmeme.com/memes/lets-see-who-this-really-is

 

[Mercutio]

Apparently, the XT 7600 and 7700 take nVidia hardware to the shed at their respective price points. This is actually fantastic news. An XT 7700 with performance on par with an RTX 4070 that also has 16GB RAM AND costs $150 less is what we are all looking for for a next Gen update, I think.

 

[LunarMist]

Is it better for the Compute?

 

[Mercutio]

It's hard to say because an awful lot of stuff wants to work with CUDA, which is an nVidia-only deal. It seems to be a generally better graphics card all around though.

 

[LunarMist]

Why is the 'Cuda only on nVidias, is IP exclusivity still in effect or AMD is not interested?

 

[sedrosken]

It will only ever be available on nVidia hardware as it's an nVidia technology and I doubt they're willing to license it out, at least (maybe) until such time as nVidia goes completely mask-off and leaves the PC graphics card market behind for their greener AI pastures. nVidia benefits from leaving everyone else out in the cold as it makes it so a professional only really has one choice.

 

[Mercutio]

The other mainstream machine learning language is OpenCL, which has pretty pathetic support compared to CUDA. There's a dim possibility that Intel and/or AMD could leapfrog nVidia and CUDA, maybe with some new and hyper-specialized compute units or extra instructions, but given where nVidia's bread is being buttered these days, that's not terribly likely.

AMD has been a pretty good egg in a lot of this stuff: Freesync works with any combination of video card and monitor; OpenCL is, uh, open. AMD releases open source graphics drivers itself and apparently its competitor for DLSS, FSR, is also made to work with any combination of display tech.

 

[LunarMist]

So if I select the Open CL box it does nothing for the RTX 4070Ti compared to just selecting the video card?

 

[Mercutio]

So if I select the Open CL box it does nothing for the RTX 4070Ti compared to just selecting the video card?

OpenCL is probably somewhat slower than using CUDA to do CUDA stuff.

 

[sedrosken]

No 'probably', definitely. OpenCL is treated as a back-path option only to be used if you absolutely must in most software if it supports it at all. If you're using any sort of professional compute solution, you're using CUDA or you're basically running in software. The only software I saw make decent use of it was Folding@Home. (I have a bad enough electric bill as it is, I don't need to dump more heat into my office to deal with.)

 

[LunarMist]

Selecting the Open CL box does nothing with the physical video card in my desktop systems. Now I understand the 'Cuda function.
The auto setting works best in the ultralight laptop with the puny XeRemis video system (no physical video card).

 

[Mercutio]

Apparently, second-hand RTX 3080TIs are now a very reasonable $400 - $500. If you don't mind the power draw, that's about as reasonable as anything I can see from nVidia. The 4070TI only just barely better for $300+ more, which says nVidia is still insane, even after another round of price cuts.

 

[LunarMist]

I received the basic RTX 4060 and tested it in my old 5950x system. Then I restored the 3060 Ti that was just sitting around since it was in my main system when they both rebuilt in Q1 (the 4070Ti was later). The 4060 is about 10% lower on the benchmarker points compared to 3060 Ti, but the heat difference is quite palpable! The 3060 Ti is an inch longer and construction includes a heavier HSF although both are NMT 2 slots in width. I think the low-res user (21-24MP) will be more than satisfied to get the 4060. Meanwhile I will have the backup system with the 3060 Ti instead of the unpowered Quattro Pasquale. According to the DXO the expectation for PL 7 is a 2060, but I think the Quattro would still have worked.

 

[LunarMist]

The conversion time for each D500 file using DeepPrime XD on the RTX 4060 worked out to be 5 seconds.
That's just a bit off of my calculations of 4 seconds and far better than the >1 minute needed for the internal radeon of the 5700G.

 

[Mercutio]

According to PugetBench, it seems that Resolve Studio works best on a 24GB Radeon 7900XTX, which is kind of interesting because its closest price competitor is the 4070 Ti, which is considered by most metrics a mid-tier card.

This generation of GPUs is insane. I want an Arc for hardware codec support, a Radeon for Resolve Studio and at least $800 worth of nVidia hardware for basically anything else.

 

[LunarMist]

I find it difficult to believe that you cannot get adequate results on 4K video using $800 video cards.
In any case you must work all production costs into the business model.

 

[Mercutio]

I'm sure I CAN. The obnoxious issue is that the $800 cards don't do something the entry level Intel cards can.

 

[Mercutio]

nVidia is about to announce new RTX Super hardware at CES. AMD may have new cards coming as well. The 4070Ti Super has 16GB RAM and a not-crippled 256-bit memory bus for $800 MSRP and there are price cuts expected for the current 4080, 4070Ti and 4070 and probably on the AMD side as well.

I still feel like pricing on these cards is in point and laugh territory.

 

[CougTek]

The new pricing for NVidia GPU have been leaked:

First up, MLID reports that, per one of his NVIDIA sources, the RTX 4080 SUPER price is set at $999, the same MSRP as the RX 7900 XTX. The source also alleged the RTX 4070 Ti SUPER and the RTX 4070 SUPER prices to be $799 and $599 respectively, matching the launch MSRPs of the RTX 4070 Ti and the RTX 4070.

Interestingly, the source suggested that NVIDIA is also dropping the “official MSPR” of the RTX 4070 to $549 while discontinuing the vanilla RTX 4070 Ti and RTX 4080 following the launch of the SUPER variants.

 

[LunarMist]

Is there a new manufacturing process or something that will improve performnce per watt, or are the Super duper versions just build with more RAM and Cuda cores? Does this all mean that there are no 50 series GPUs coming in 2024?

 

[Mercutio]

There is no new manufacturing process. I think we're just doing more RAM and wider memory buses for this round.

 

[LunarMist]

Then there is little value in upgrading from another 40 series? I suppose it makes more sense for someone that has been waiting with older video cards. The "Super" prices are lower for more cores and you are complaining about the cost?

 

[Mercutio]

I'm just not willing to spend $500+ on crippled hardware. If Intel and pretty much every contemporary mobile SoC can handle 10 bit 422 HEVC in hardware, why the hell can't nVidia?

 

[LunarMist]

Probably the number of users needing that function are less than 4.55%.
Don't you have space in those massive servers for several video cards, maybe three brands at the same time?

 

[Mercutio]

Probably the number of users needing that function are less than 4.55%.
Don't you have space in those massive servers for several video cards, maybe three brands at the same time?

I'm sure it is, but there are plenty of BS features on GPUs that only a minority will use. Doesn't mean there's an excuse for not supporting it if every decent mobile device can do it without issue or complaint.