GeForce RTX 4060 Ti and 4060 Graphics Cards | Hacker News Reader

GeForce RTX 4060 Ti and 4060 Graphics Cards

nvidia.com

47 pointsfomine372 comments

The 4060 (non-ti) has only 8GB of ram, which is less than its predecessor, the 3060, which had 12GB and came out in Feb 2021. I wonder if this is unprecedented?

For anyone not really following the current issues with video cards that have only 8GB of VRAM, there are a bunch of modern games that stutter a lot with only 8GB. I have an 8GB 3060 Ti and regret the purchase since it's been a problem for me in several games so far.

AMD is expected to announce new mid-range models soon, which might provide more options to choose from that have 16GB+.

Just seems like a money grab. The 4060 Ti has half the memory width (128 bit vs the old 256 bit wide) of the 3060 Ti, not much more than half the memory bandwidth. The old 3060 Ti has 448GB/sec and the new 4060 Ti has 288GB/sec.

This card depends on faking textures (DLSS) and faking frames with frame generation to increase performance. Frame generation seems particularly lame since the additional frames are generated without interacting with the game, thus increasing input latency and display latency.

So after 2.5 years you get to pay more for a worse card with less memory bandwidth that depends on cache hits and fakery to show any noticeable performance improvements.

Seems we need a new reversed Moore's style law for the regression taking place in the GPU space.

What we need is more competition. Nvidia has plenty of substantially faster cards they are shipping, they are just minimizing performance delivered and maximizing cost to the consumer.

4070 is comparable in bandwidth to the 3060 Ti and would be a much nicer upgrade for 2.5 years, but instead they increased the price by 50% and named it a 4070 instead of a 4060 ti.

Isn't the current AMD/Nvidia duopoly similar to the Intel/AMD duopoly we had in the CPU race era?

Yes, but AMD GPUs haven't (so far) pulled off any of the triumphs that AMD CPUs have. The ones that spring to mind:

  * Intel tried to reserve 64 bit for servers/Itanium, AMD brought out x86-64 across their line
  * Intel was pushing front side busses or dual busses.  AMD moved the memory controller on chip

As a result Nvidia has a lead in performance (like the 4090), in efficiency (perf/watt), market share, and in software (drivers and CUDA).

I don't think it's fair to say they're winning in drivers. Their driver is an absolute pain to work with. It's to the point that I'd rather have no GPU than an Nvidia one. I'm very much not the common user but their situation is abysmal.

Intel are just starting to enter the discrete GPU space in earnest - I'm hopeful they'll do something innovative, but I'll settle for competition and increased supply at this point.

I am confident that in gaming, AMD and Intel can compete if they can get their act together.

But for AI/ML, we are pretty much stuck with Nvidia.

To me it's more like the 3060 having 16GB is weird. It has the most VRAM in the generation after the 3090 (24GB).

We really need more competition in the GPU space. Nvidia's release strategy seems designed to exploit the mid-tier gamers.

Intel is trying, specifically at the mid-tier. These look good though as long as availability doesn't become a problem.

> as long as availability doesn't become a problem

And you know it will, within minutes of release.

Intel has worse drivers than ati hence the discount

AMD seems to be providing good alternatives. Arguably the 6600XT or the 6700XT are better value than the 4060. And the 6800XT is a good alternative to the 16GB 4060Ti. And AMD's cards are actually purchaseable right now.

6750 XT was on sale a few weeks ago for $330 - $20 rebate and came with a copy of The Last of Us.

Guessing it'll be a while before we see similar deals on Nvidia cards since so many people are determined to buy Nvidia over AMD for reasons I don't understand.

Yes the raytracing performance is better, but not enough to justify what they're charging.

Just using the card for gaming? Absolutely go with AMD. But most of the ML libraries and tooling is built on top of CUDA, so getting things to work with an AMD card can be annoying and have worse performance despite better hardware specs on paper.

And yet the Steam hardware survey has Nvidia at 76% compared to AMD at 15%. Most of that isn't because of ML performance.

Yeah, a lot of that is inertia. Most people don't upgrade video cards all that frequently which explains the popularity of the GTX 1650 (2019), GTX 1060 and 1070 (2016), etc. Those cards still work reasonably well for people who don't need bleeding edge performance or ray tracing. I'd expect their market share to improve over the next couple years as some of those cards age out of circulation.

Availability is a huge problem for ati cards. There weren’t that many out there in the last 2 years where nvida vendors at least had a bunch of overpriced options

DLSS, NVENC, CUDA, better drivers(windows) and more. I use a 6700xt, and I enjoy its better support on Linux, but absolutely not how its not suoported by ROCm, while a 3050 is fully supported by Cuda and can run any AI or professional program.

There are many reasons not to get an AMD card.

Is there a power difference?

That's my concern as well. Though if I'm honest I don't game for more than a few hours a week, tops.

According to Steam hardware, the majority of gamers are running their games at 1080p.

The graphics card distribution pretty spread out across a number of generations.

https://store.steampowered.com/hwsurvey/

A lot of gamers chase FPS to go with their high refresh rate monitors, and 1080p is a sweet spot for that right now.

These are the latest generation, not mid-tier. Mid-tier is your 3060 IMO, and will probably become 3080 with the release of these new cards. That mid-tier 3060 gives you 4k gaming or HD gaming at 60+Hz refresh. I think the release strategy seems to mostly work, keeping mid-tier gaming rigs on par or better with latest gen consoles and about a 3-5 year lifespan. PC gaming would not exist if the pricing was exploitive, because the bulk of the market would switch to consoles even though the games themselves are more expensive.

4060 Ti w/16 GB seems like a sweet spot for machine learning hobbyists. Only bummer from what I can see is the 128 bit bus.

> The $399 RTX 4060 Ti will ship on May 24th with just 8GB of VRAM, while a 16GB model is due in July priced at $499.

Is that accurate?

It seems!

I think the more amazing thing is the "only" 165 W power draw for the 4060 Ti. Although, given marketing these days, I am sure that number is lying, but I am still happy to see something reasonable. The 4090 is listed with a TDP of 450W.

I don’t think that’s enough, cloud solutions are still cheaper and faster I believe

Gaming class GPU's provide much better $/FLOPS (ie a 4090 is ~5x better value than a H100)

https://timdettmers.com/2023/01/30/which-gpu-for-deep-learni...

Nvidia forces corporations to buy datacenter GPU's for CUDA licensing reasons, and adds a significant margin on top.

this is not a 4090 and even then your article mentions experiments are better on a cloud solution. Where they only recommend getting a top of the line GPU if you are doing longer than a year work

I unfortunately don’t have $2000 to spend on a GPU. Would be nice though.

*60 TI always seems to be a nice mid-market pick.

AMD's 6800XT should be a pretty good competitor to it; 500 bucks, 16GB vram, and a 256 bit bus. And the 6800XT is already available

Good luck with the software

Why? Buggy drivers? No CUDA?

No CUDA.

ROCm is a thing, though (but not on Windows.)

Don't forget about the Intel Arc A770. 16GB VRAM, 256bit bus for $350. Can use OneAPI for machine learning too.

Unless a better option comes out in the next few months, I might pick one up. It seemed a bit pointless to upgrade to anything below 16 GB, and up until now that meant a 3090 or 4090. And the prices for those are pretty steep for hobbyist use.

The mac mini comes with 32GB

Nowhere near the same computation for floating point.

How big of a factor is the 128 bit bus?

Since there is so much attention on just a few models now AMD could probably make some big inroads by producing some affordable very high memory cards and porting a few models to them... while nvidia continues to limit consumer hardware ram to market differentiate with their high end products.

I just wish they're release a card with mid-range compute but much higher memory to be honest. Would be cool to have an ML focussed series of cards with consumers in mind. But I guess it doesn't make financial sense to them.

Though I do find the binning interesting...I guess they bin on compute and then restrict the bus width of the memory controller based on that. Would be cool if they could also bin on memory controller/bus width, then just offer some average compute along with that. Means people go from being able to run smaller models faster (3090/4090) to being able to run larger models (where they couldn't before) just a little slower.

Gimme a 4060-4070 with like 24GB or vram, hell even 40-48GB.

What amount of ram and GPU are people considering adequate for playing around with all the local inference possibilities over the next year?

As much as your budget allows for.

Agreed. But your choices are now the 4060 to 16gb for $500 or a way more expensive 3090/4090 with 24gb and the power draw to go with it.

I started out with a 1080ti that I had lying around. Within days upgraded to an RTX3060. Within weeks of that upgraded to an RTX 4090.

In retrospect I wish I had of known that "buy once, cry once" might have been the way to go here.

That's just me.

3090s go for under $800 used, and you can set the power limit to whatever you want.

I found that 290W keeps the fans from spinning up audibly under heavy CUDA workloads with ~2% lower performance.

Out of curiosity, are there any real reasons to buy these over used RTX 3070's or used RX 6700XT's or other stuff like that?

Only if you care about power consumption.

3070 is I think 220~250 watts, while the 4060 sits at 160~200 watts.

DLSS3 makes these better than any previous gen GPUs including 3080s for games with DLSS3 support.

My 3070 has 8gb vram. I’d rather have 16gb of vram of the 4060 Ti for a few hundred less than I paid for mine.

https://www.youtube.com/watch?v=Y2b0MWGwK_U

Well, looks like there is literally 0 reason.

4060 performs worse than a 3060Ti in some benchmarks.

Do they even want to sell these when they could use their silicon budget for much higher margin data center chips to an unsatisfiable AI demand?

Based on their "Performance"graph on that link, Autodesk Arnold & Call of Duty are faster on the regular 4060 than the 4060Ti

They are being compared to two different cards. The TI is to the 3060 Ti, and the non Ti version to the non ti 3060.

What’s the story on the blocks they’ve put in to prevent bitcoin mining. Does that also make them bad for other GPU based workloads?

it's not profitable to mime Bitcoin on graphics cards any more, you need a dedicated ASIC. Ethereum was profitable, but they have moved ot proof of stake now, so there's not nearly as much demand for gpu mining

Is there a performance chart comparing last gen (30xx series) to current gen? Specifically current and last gen mid-tier cards ?

Games are cool I guess, but how well will they perform at llama.cpp and stable diffusion? That's really all I care about.

In that case the 8 GB card is probably right out.

Llama.cpp runs mostly on CPU. I say "mostly" because it's possible to compile it with cublas to use GPU for some of the computations.

My 2060 Ti 12GB is usably fast for SD, so I’m guessing these will work great.

Actually it's only a 2060 12GB, not even a Ti.

How is posting this link a legit use of hacker news? I get that Hacker news is often used for important product releases/updates, but when it comes to Nvidia and AMD, this feels like product placement. Nothing in this is particularly newsworthy hardware, and we're talking about the coke vs pepsi of the pc gaming adapter world. If it was someone reviewing it, that would be different, but this is right to the vendor. I expect more from the mods.

HN is for talking about stuff that hackers find interesting. Hackers are interested in cheap GPUs to level up their machine learning abilities.

Yes I think that cheaper Nvidia board with 16GB VRAM is newsworthy (though memory bandwidth is bad)

People are very interested in generative AI which can be VRAM hungry. LLMs, Graphics, etc. One of these is a mid range SKU with 16GB of VRAM. The next notch up doesn't even have a 16GB VRAM SKU.

As far as the Coke vs. Pepsi comparison, Nvidia's CUDA is like not even that right now. It's a curbstomp in developer experience. You can get in on that action for cheaper now.

4060 Ti is big news for generative ai and llm hobbyists on a budget.

Reply on news.ycombinator.com