• Some users have recently had their accounts hijacked. It seems that the now defunct EVGA forums might have compromised your password there and seems many are using the same PW here. We would suggest you UPDATE YOUR PASSWORD and TURN ON 2FA for your account here to further secure it. None of the compromised accounts had 2FA turned on.
    Once you have enabled 2FA, your account will be updated soon to show a badge, letting other members know that you use 2FA to protect your account. This should be beneficial for everyone that uses FSFT.

Legacy Nvidia RTX 3060 12GB returns to retail five years after original launch, priced at $339

MrGuvernment

MrGuvernment

Fully [H]
2FA
Joined
Aug 3, 2004
Messages
24,091
I think it has been rumored NVIDIA said no new GPU's for a year or 2, so lets bring back old cards?

As some have noted, you can buy an AMD 9060 for $10 more and get more performance?

https://www.tomshardware.com/pc-com...lled-a-good-idea-apparently-comes-to-fruition

Legacy Nvidia RTX 3060 12GB returns to retail five years after original launch, priced at $339 — resurrected GPU strategy that Jensen called a 'good idea' apparently comes to fruition​

News
ByJeffrey Kampman published22 hours ago
Ampere rides again for almost the same price as RTX 5060s



Back at CES 2026, Nvidia CEO Jensen Huang said that "it's a good idea" to consider reintroducing older GeForce RTX graphics cards with GPUs from trailing-edge nodes to alleviate pricing and availability concerns around cutting-edge RTX 50-series products. It appears that the idea may now have become reality. After a hiatus that began late last year, RTX 3060 12GB cards (first launched in early 2021) are beginning to reappear as new stock at e-tailers.
We've spotted a Gigabyte RTX 3060 12GB card available direct from new stock at Newegg for $339.99, or just $10 above its MSRP from over five years ago. This Windforce card is a straightforward dual-fan model with no frills, but tellingly, it carries a rare Rev2.0 suffix, directly indicating a revision or new model of some kind. That revision number aside, it's identical (as far as we can tell) to the Rev1.0 card that presumably appeared nearer to the RTX 3060 12GB's launch.
It appears that different board partners are working to quietly bring back the RTX 3060 12GB worldwide, as ComputerBase notes the availability of an Asus Dual card in the European market. We wouldn't be surprised if that product makes its way Stateside, as well.
Gamers may have hoped that the reintroduction of an older product from a foundry and a process node that isn't capacity-constrained (namely Samsung 8nm), as well as an older memory technology, would have resulted in some pricing relief, but that doesn't appear to be the case.
Image 1 of 3


77yhF8ajKTTosserzsARYN.png

(Image credit: Future)

cCbGspLq5yrwLNtkmB8vYN.png

(Image credit: Future)

oUxzMouhsE3X4GJbF62FZN.png

(Image credit: Future)

The RTX 3060 12GB is reappearing for just a few bucks less than the far superior RTX 5060, which offers both stronger baseline performance and unfettered performance with the latest DLSS 4.5 upscaling tech. Heck, even the entry-level RTX 5050 outpaces the RTX 3060 in our most recent tests.
The 3060's extra 4GB of VRAM compared to those products only comes into play with settings and resolutions where its shader horsepower is already inadequate to produce playable frame rates, so it's more of a psychological advantage than a practical one.
So why this product and why now? As AI wafer demand on cutting-edge nodes only grows, it's possible that simple economics dictate that the opportunity cost of producing entry-level Blackwell GPUs for the desktop is simply too great compared to the revenue and margin opportunities of making data center Rubin GPUs for the countless Vera Rubin NVL72 racks that the company is doubtless planning to ship later this year.

Get Tom's Hardware's best news and in-depth reviews, straight to your inbox.
Contact me with news and offers from other Future brandsReceive email from us on behalf of our trusted partners or sponsorsBy submitting your information you agree to the Terms & Conditions and Privacy Policy and are aged 16 or over.
In the event that Nvidia does prioritize datacenter products over consumer ones, the already elevated prices of entry-level RTX 50-series cards might give way to a situation where one simply can't buy these lower-priced, lower-margin products at any cost. And in that event, the aging RTX 3060 could be one of the few sub-$350 options available to gamers who need to build a new system or upgrade from an aging graphics card.
And frankly, that's a poor outcome for gamers, as the RTX 3060 can't run the latest DLSS 4.5 upscaling model at its full potential due to the lack of proper FP8 acceleration on older Turing and Ampere GPUs. And it can't run DLSS Frame Generation at all, leaving gamers to deal with cross-platform solutions like FSR and XeSS frame generation from other vendors. Those technologies are incredibly useful tools in getting the best experience out of modern games, and if they're only available on a smaller subset of more expensive GPUs, that further stratifies the already stratified modern PC gaming experience.
We've reached out to Gigabyte and Nvidia for comment on the reappearance of the RTX 3060 and will report what we learn if we hear back.
Click to expand...
 
MrGuvernment said:
As some have noted, you can buy an AMD 9060 for $10 more and get more performance?
Click to expand...
for how long would that be the case, newegg do not sell 8GB anymore it seems:
https://www.newegg.com/p/pl?N=600494828 100007709 8000&d=amd+9060

Bestbuy single sku left:
https://www.bestbuy.com/site/search...+Unit+(GPU)~AMD+Radeon+RX+9060+XT&st=amd+9060

microcenter has well, single sku for the 8gb (xfx) left close to the price, and some location with zero.

pcpart show only 2sku left at normal price market wise:
https://pcpartpicker.com/products/video-card/#sort=price&page=1

Could be for a short time that $349.99 9060xt can be fund out there.

Remember all that talk about GPU rapid depreciation that would happen.... 3060 sold at higher (before inflation) msrp 6 years later...
 
1_rick said:
9060 is an 8GB card, though.
Click to expand...
is a 3060 powerful enough to use 12GB of ram? Just thinking back to the days of AMD days, when most games couldn't even use 8G if you got those models, or some of the other NVIDIA cards that were re-released with more RAM, but could never actually use it...
 
MrGuvernment said:
is a 3060 powerful enough to use 12GB of ram? Just thinking back to the days of AMD days, when most games couldn't even use 8G if you got those models, or some of the other NVIDIA cards that were re-released with more RAM, but could never actually use it...
Click to expand...
some use case can happen even if much rarer at that level of card, sometime high resolution texture for example use a lot of memory without using significantly more compute, virtually zero more compute on some game engine, as long as bandwith allow it and vram capacity is big enough very little change in FPS.

9060xt would tend to be the much better choice at a close price, it is 50% ~faster on TPU relative performance chart, a bit like having a 3070 instead of the 3060, exception for niche DLSS case or CUDA affair or local inference for something that love to have 12gb.
 
Is this new silicon? Smell fishy to me. Like the GPU cores have been sitting in a warehouse for years and somehow never made it to e-waste. Conspiracy theorists would say these were salvaged from bitcoin mining rigs.

Uses the same GDDR6 as the RTX 5050, so it's not like this solves any memory shortage problems currently. Even if Nvidia were sitting on a mountain of these, they still gotta source the vram from somewhere. But maybe that's a Nvidia partner problem.
 
MrGuvernment said:
is a 3060 powerful enough to use 12GB of ram? Just thinking back to the days of AMD days, when most games couldn't even use 8G if you got those models, or some of the other NVIDIA cards that were re-released with more RAM, but could never actually use it...
Click to expand...

in AI yes - gaming there's times it can def benefit from over 8GB if not the full 12GB - like 9/10GB usage based on this


View: https://www.youtube.com/watch?v=dx4En-2PzOU
 
Absalom said:
Is this new silicon? Smell fishy to me. Like the GPU cores have been sitting in a warehouse for years and somehow never made it to e-waste. Conspiracy theorists would say these were salvaged from bitcoin mining rigs.

Uses the same GDDR6 as the RTX 5050, so it's not like this solves any memory shortage problems currently. Even if Nvidia were sitting on a mountain of these, they still gotta source the vram from somewhere. But maybe that's a Nvidia partner problem.
Click to expand...

No, it's just cheaper to produce on older nodes. The congestion is on the cutting-edge nodes, so this gives Nvidia the opportunity to launch a card to serve the gaming market they no longer care about with mature driver support, so they don't have to worry so much about that either and can just vibe code the updates.
 
I can't wait to buy this new 5-year old GPU to pair with a brand-new, newly-launched 4-year old 5800x3D and buy a brand-new AM4 motherboard based on 16-year old technology. The future of the PC hardware enthusiast is bright, my friends!
 
Absalom said:
Is this new silicon? Smell fishy to me. Like the GPU cores have been sitting in a warehouse for years and somehow never made it to e-waste. Conspiracy theorists would say these were salvaged from bitcoin mining rigs.

Uses the same GDDR6 as the RTX 5050, so it's not like this solves any memory shortage problems currently. Even if Nvidia were sitting on a mountain of these, they still gotta source the vram from somewhere. But maybe that's a Nvidia partner problem.
Click to expand...
If they were on tsmc like RTX 5050 new silicon would make little sense that true.

But that old Samsung 8nm that was restarted and left in part unused after peak produciton of switch 2, they could be "new", made during switch 2 soc dull time in recent time for this.

I think there was long rumors about this relaunch because they were not already made 3060 (which would have been easy to keep secret until last minutes) but new one made:
https://www.hankyung.com/article/2026030866841 (https://www.techpowerup.com/347174/samsung-to-resurrect-nvidias-geforce-rtx-3060-using-8-nm-node), This has been confirmed by the Korean media outlet Hankyung, which reports that Samsung is restarting its 8 nm node production to meet NVIDIA's needs.

Nvidia still usually ship the ram with the gpu so it would be there problem, but it is not the exact same, 3060 8-12GB used 1875mhz 15 Gbps memory gddr6 while 5050 used more expensive-hard to do 2500Mhz 20 Gbps effective memory according to tpu, those old that were not close to be top of the line GDDR6 back in 2020 modules must have near perfect yield now.
 
Last edited:
HockeyJon said:
I can't wait to buy this new 5-year old GPU to pair with a brand-new, newly-launched 4-year old 5800x3D and buy a brand-new AM4 motherboard based on 16-year old technology. The future of the PC hardware enthusiast is bright, my friends!
Click to expand...

Big :(

We are really living the dream right now. And by dream, I mean nightmare.
 
MrGuvernment said:
is a 3060 powerful enough to use 12GB of ram? Just thinking back to the days of AMD days, when most games couldn't even use 8G if you got those models, or some of the other NVIDIA cards that were re-released with more RAM, but could never actually use it...
Click to expand...
In many modern titles 8GB of VRAM is insufficient even at minimum settings at 1080p.

Whenever it has to start waiting for resources to swap in from system RAM via PCIe pretty much any GPU is going to take a real performance hit.
 
MrGuvernment said:
I think it has been rumored NVIDIA said no new GPU's for a year or 2, so lets bring back old cards?

As some have noted, you can buy an AMD 9060 for $10 more and get more performance?

https://www.tomshardware.com/pc-com...lled-a-good-idea-apparently-comes-to-fruition
Click to expand...

Just waiting for people to try to figure out how to get four of these into a single system to run ~70B Ollama models (which generally require 48GB) all in VRAM 🙄

You'd need something with a lot of PCIe lanes, but there are lots of decommed older server boards out there...

The VRAM bandwidth on these would likely limit the performance some, but it would still be way faster than running on the CPU and system RAM, which means people will likely try as it would wind up being one of the most cost effective ways to run larger models locally, presuming they don't overspend on the CPU/Motherboard platform.

That unfortunately seems to be the way everything goes these days. The AI people generally have a limitless thirst for capacity, and they are usually willing to pay for it. Four 3060's may not perform as well as two 4090's, but it is a hell of a lot cheaper.

In fact, it wouldn't surprise me if this is really who these GPU's are intended for rather than the consumer market.
 
Last edited:
Zarathustra[H] said:
Just waiting for people to try to figure out how to get four of these into a single system to run ~70B Ollama models (which generally require 48GB) all in VRAM 🙄
Click to expand...
I poked around on eBay looking at used gear and came to the conclusion that getting 4+ good slots was going to cost a 3-4 grand for the platform, so that doesn't include any GPUs. Just board, proc(s), ram, case, PSU and some storage. There are multiple ways to get there, but however you slice it you're looking at ThreadRipper, EPYC or Xeon for 4+ decent PCI-e vid card slots, and the cheapest route for an AI machine is something that uses registered ECC DDR4. DDR4 RDIMMs are the only type of ram you can still get reasonably cheap used unless you want to go really old and start buying DDR3 or something. I kinda want to build an inferencing box, but I'm really not sure it's worth building vs. just buying a Strix Halo, DGX Spark or maybe a Mac. I'd be aiming for 128GB vram, probably using AMD or Intel cards (NV prices are nuts), but building that is more expensive than even an M5 Max MacBook Pro with 128GB. I kinda like the DGX Spark just because it's a wierd new thing to play with, Strix Halo because it's the cheap 128GB option, upcoming Gorgon Halo for 192GB for about the same price as a 128GB DGX Spark, and then a top end Mac laptop would be an awesome laptop for many years. By the time I make up my mind Gorgon Halo will probably be out and I'll just end up getting one of those.

If you want 48+GB vram in a consumer grade box using new hardware I think it'd be a lot cheaper to just use 2x AMD Radeon AI Pro 9700s or Intel ARC B70 Pro cards than 4 3060s. That'll get you 64GB vram, and a dual slot ordinary desktop board is much, much cheaper than something that has enough slots and lanes to run 4 cards at a reasonable speed. AI isn't mining. Slot speed counts.

I only see one real use for this 3060 relaunch at $340 - pro/creative/etc. users who want NV for software support (CUDA, etc.) and want >8GB but aren't willing to pay up for a 16GB 5060Ti or 12GB 5070. That actually makes sense with some apps - some pro/creative apps are far more vram limited than GPU limited.

I really don't see the point to a 3060 for gaming. Between a much faster 8GB 5060 or 8GB 9060XT and a 12GB 3060 I'd take the 8GB option for gaming. Yeah 8GB gets in the way and limits settings, but not all that much when your GPU is a 3060 and there are so many 8GB cards out there game devs will be figuring out how to make them work in the latest AAA game for years. Then if you really want that 12GB a $310 Intel ARC B580 will generally smoke a 3060 in games. I've got one in one of my rigs, and XeSS is well supported in AAA titles. I doubt the reason has anything to do with ARC cards. Pretty sure it's all about the iGPUs.
 
Not the first time NVIDIA has done this.
First with the GT 730 during the first GPU/crypto apocalypse in 2017-2018.

Then again with the GTX 1050 Ti around 2020-2022 with the second GPU/crypto apocalypse and first supply chain issues this decade.
History, and the market NVIDIA itself represents, repeats itself.

11.jpg
 
Zarathustra[H] said:
That unfortunately seems to be the way everything goes these days. The AI people generally have a limitless thirst for capacity, and they are usually willing to pay for it. Four 3060's may not perform as well as two 4090's, but it is a hell of a lot cheaper.
Click to expand...

As far as I know, the "sweet spot" for """inexpensive""" multi GPU concurrency is 3090s, since they can share memory more easily. I feel that with 3060's the cost of the motherboard(s) and networking solutions that you would need to actually slot that many GPUs in and have them try to (awkwardly) share memory would kind of outweigh the savings. Plus all of the upkeep, noise, and power supplies. Then again I'm still kind of a layman, relatively speaking. There is probably someone that could pull off actually scaling with a massive amount of these, but I feel like they could make money working anywhere else that could actually afford better GPUs instead, if they could do that.

zandor said:
I poked around on eBay looking at used gear and came to the conclusion that getting 4+ good slots was going to cost a 3-4 grand for the platform, so that doesn't include any GPUs. Just board, proc(s), ram, case, PSU and some storage. There are multiple ways to get there, but however you slice it you're looking at ThreadRipper, EPYC or Xeon for 4+ decent PCI-e vid card slots, and the cheapest route for an AI machine is something that uses registered ECC DDR4. DDR4 RDIMMs are the only type of ram you can still get reasonably cheap used unless you want to go really old and start buying DDR3 or something. I kinda want to build an inferencing box, but I'm really not sure it's worth building vs. just buying a Strix Halo, DGX Spark or maybe a Mac. I'd be aiming for 128GB vram, probably using AMD or Intel cards (NV prices are nuts), but building that is more expensive than even an M5 Max MacBook Pro with 128GB. I kinda like the DGX Spark just because it's a wierd new thing to play with, Strix Halo because it's the cheap 128GB option, upcoming Gorgon Halo for 192GB for about the same price as a 128GB DGX Spark, and then a top end Mac laptop would be an awesome laptop for many years. By the time I make up my mind Gorgon Halo will probably be out and I'll just end up getting one of those.

If you want 48+GB vram in a consumer grade box using new hardware I think it'd be a lot cheaper to just use 2x AMD Radeon AI Pro 9700s or Intel ARC B70 Pro cards than 4 3060s. That'll get you 64GB vram, and a dual slot ordinary desktop board is much, much cheaper than something that has enough slots and lanes to run 4 cards at a reasonable speed. AI isn't mining. Slot speed counts.

I only see one real use for this 3060 relaunch at $340 - pro/creative/etc. users who want NV for software support (CUDA, etc.) and want >8GB but aren't willing to pay up for a 16GB 5060Ti or 12GB 5070. That actually makes sense with some apps - some pro/creative apps are far more vram limited than GPU limited.

I really don't see the point to a 3060 for gaming. Between a much faster 8GB 5060 or 8GB 9060XT and a 12GB 3060 I'd take the 8GB option for gaming. Yeah 8GB gets in the way and limits settings, but not all that much when your GPU is a 3060 and there are so many 8GB cards out there game devs will be figuring out how to make them work in the latest AAA game for years. Then if you really want that 12GB a $310 Intel ARC B580 will generally smoke a 3060 in games. I've got one in one of my rigs, and XeSS is well supported in AAA titles. I doubt the reason has anything to do with ARC cards. Pretty sure it's all about the iGPUs.
Click to expand...

Well not quite, you can technically find older PCIE 3.0 Intel HEDT platforms for not that much (obviously some caveats, like the handicapped memory bandwidth due to the 3.0 slot). There are also PCIE 4.0 TR's with 3 x16 slots and 1x8 slot (plenty for this) if you really look hard enough (I think I parted it out to like 700-800ish), but the amount you would spend on them still comes close to the cost of 4 3060's and you're left with a pretty weird and shitty system. Overall agreed. I have the DGX Spark that I got for 3k, and it's been a fun little machine. Not the fastest at spitting out, but it can run a 122B model at ~50 tok/s so I'm fine with it.
 
StoleMyOwnCar said:
Well not quite, you can technically find older PCIE 3.0 Intel HEDT platforms for not that much (obviously some caveats, like the handicapped memory bandwidth due to the 3.0 slot).
Click to expand...
A GPU won't saturate the pcie bus at pcie 3.0 16x like M2 storage will. Having said that, the difference between pcie 3.0 and pcie 4.0 in relation M2 bandwidth isn't large enough to really notice a difference in daily use - and any M2 storage based on pcie 5.0 is just stupidly expensive and out of reach for most.

X299 has the capacity (depending on CPU used) for 48 pcie lanes direct to the CPU, and an additional 24 lanes via the chipset. My memory bandwidth running DDR4 PC3600 CL16 in quad channel is actually faster in terms of reads and about equal in terms of writes than most DDR5 PC6000 AMD consumer grade dual channel systems running Intel's MLC checker
 
Last edited:
Mazzspeed said:
A GPU won't saturate the pcie bus at pcie 3.0 16x like M2 storage will. Having said that, the difference between pcie 3.0 and pcie 4.0 in relation M2 bandwidth isn't large enough to really notice a difference in daily use - and any M2 storage based on pcie 5.0 is just stupidly expensive and out of reach for most.
Click to expand...


It really depends on what you are doing with it.

For games, No. You'll get little bursts here and there, particularly at load time, but even then the PCIe bus is not the limiting factor. Usually storage is. The only exception is if you are running with insufficient VRAM and thus swapping to system RAM, and even with a fully maxed out PCIe bus that tends to slow things down to a crawl anyway, but I guess at least in theory, if you are just very slightly above your VRAM limit, and only need to swap a little to system RAM, a fast PCIe bus could in theory help mitigate that, but in practice I have never seen it work that way.

Also, for mining, the PCIe bus is mostly irrelevant. Those damn miners often run (ior at least used to) several GPU's at 1x PCIe each just to max out how many they could fit in a system.

AI/LLM loads however are different. If you are running a multi-GPU setup and don't have the benefit of something like NV-link to tie them together, the PCIe bus can become a huge limiting factor, especially during the prompt decode stage when the multiple GPU's in a system need to coordinate with eachother a lot. (it is less relevant during the output generation stage which is mostly VRAM bandwidth bottlenecked)
 
StoleMyOwnCar said:
As far as I know, the "sweet spot" for """inexpensive""" multi GPU concurrency is 3090s, since they can share memory more easily. I feel that with 3060's the cost of the motherboard(s) and networking solutions that you would need to actually slot that many GPUs in and have them try to (awkwardly) share memory would kind of outweigh the savings. Plus all of the upkeep, noise, and power supplies. Then again I'm still kind of a layman, relatively speaking. There is probably someone that could pull off actually scaling with a massive amount of these, but I feel like they could make money working anywhere else that could actually afford better GPUs instead, if they could do that.
Click to expand...

That is probably true, but AI right now has such an insatiable appetite that as soon as it gets around that something is a "sweet spot" for low cost local models, whatever that "something" is tends to radically increase in price overnight.

3090's are going for as much as $1,400 each right now on eBay and every time I look they seem to get more expensive.

I sold my 4090 locally last year. If I had held on to it I could have sold it for quite a bit more than I sold it for then today.

At $300 a pop, you could throw together four 12GB 3060's and make it work for $1,200 instead of $2,800 for two 3090's. It won't be as fast as dual 3090's but it will certainly be way faster than running entirely on CPU, or running on GPU's with less VRAM than fits the model and swapping to system RAM, so it would be a pretty decent budget option.

Yes, whoever does this would need a system with a lot of PCIe lanes, but a lot of home labbers have spare old server boards kicking around they could stick them in. I have no plans of doing this, but I have a couple of spare Supermicro server boards I am no longer using in servers that have quite a lot of PCIe lanes. My X9DRI-F - for example - has 80 total Gen3 lanes. If I wanted to set up an Ollama machine it would cost me nothing apart from the GPU's.

This is what gamers are competing against.

A popular local target model these days is Llama 3.3 70B with 4 bit quantization, which requires about 42GB of RAM. A little bit more if you factor in the working prompt cache and conversation history, so they just about fit perfectly in 48GB of VRAM, which is why that has become a "rule of thumb" target level for local home lab LLM's.

Four 3060's would probably get up to 8 tokens per second. Typical adult human reading speed (silent, in head) is about equivalent to 5 tokens/s. Skimming speed goes up to about 8 tokens/s. So - while not ideal - it is certainly usable for many things. Really depends on what you are doing with it.

By comparison dual 3090's would likely output at about 22 tokens per second. Certainly a lot nicer, especially for complex prompts, but it would come at more than twice the price.

I honestly hope more GPU's wind up in the hands of gamers, but I'm trying to have realistic expectations of what I think will actually happen.
 
You can get a 9060 XT with 16GB at Microcenter in Dallas right now for 390. Yes, that is more money, but well worth the modern architecture.
 
Zarathustra[H] said:
Just waiting for people to try to figure out how to get four of these into a single system to run ~70B Ollama models (which generally require 48GB) all in VRAM 🙄

You'd need something with a lot of PCIe lanes, but there are lots of decommed older server boards out there...

The VRAM bandwidth on these would likely limit the performance some, but it would still be way faster than running on the CPU and system RAM, which means people will likely try as it would wind up being one of the most cost effective ways to run larger models locally, presuming they don't overspend on the CPU/Motherboard platform.

That unfortunately seems to be the way everything goes these days. The AI people generally have a limitless thirst for capacity, and they are usually willing to pay for it. Four 3060's may not perform as well as two 4090's, but it is a hell of a lot cheaper.

In fact, it wouldn't surprise me if this is really who these GPU's are intended for rather than the consumer market.
Click to expand...
Ya, I was curious about them as LLM GPUs but seems they are sort of far down the list in terms of tokens speed vs a 4060 even...

I was looking at Xeon Gen 3 boards, LGA 4196 for PCIe 4, but those start around $700 CAD +.....

I already have some systems, like an HP Z6 G4 with a Xeon 2nd gen gold and 192GB of ram, so I keep thinking add in 2 GPU's and run some local models...but then I see the price and go ya, don't feel like spending $1k CAD + to play with LLMs :D

I know I could get a model like Ollama that can share system ram to offload too...but I also hate things being slow.
 
Zarathustra[H] said:
For games, No. You'll get little bursts here and there, particularly at load time, but even then the PCIe bus is not the limiting factor. Usually storage is. The only exception is if you are running with insufficient VRAM and thus swapping to system RAM, and even with a fully maxed out PCIe bus that tends to slow things down to a crawl anyway, but I guess at least in theory, if you are just very slightly above your VRAM limit, and only need to swap a little to system RAM, a fast PCIe bus could in theory help mitigate that, but in practice I have never seen it work that way.
Click to expand...
This is something I actually monitor from time to time, and I've never come close to saturating the pcie bus playing games, even running 4k textures.

Zarathustra[H] said:
AI/LLM loads however are different. If you are running a multi-GPU setup and don't have the benefit of something like NV-link to tie them together, the PCIe bus can become a huge limiting factor, especially during the prompt decode stage when the multiple GPU's in a system need to coordinate with eachother a lot. (it is less relevant during the output generation stage which is mostly VRAM bandwidth bottlenecked)
Click to expand...
Agreed.
 
Balkroth said:
I'd probably pick up an Arc B580 over the 3060 personally if looking at 12GB cards.
Click to expand...
Yup. Haven't looked at current pricing for them, but I got one for my daughter's PC for Christmas last year for around $250 and it has been great for her. Techpowerup shows it being about 17% faster on average as well, so even for about the same price I would prolly go with the B580 still.

Also, I'm not sure the point in bringing this back still when the RTX 5050 exists for considerably cheaper than $340 ($290 right now on Amazon) and is also about 14% faster according to Techpowerup's database. I would rather have a bit more performance (esp. using transformer model DLSS and access to MFG) and deal with the 8GB VRAM while being cheaper. Hell there are 5060s on Amazon right now for the same $340 that are 44% faster as well. I don't see that extra 4GB VRAM making the difference in most games at 1080p.
 
Not really, but kind of related, I took boxes of old ewaste to the local dump yesterday, there were easily a dozen video cards the newest was a Gtx460, and I saw that 8800gtx that gave me good memories too. About 15 pounds of old hard drives, which total capacity doesn't even come close to the SSD I currently have in my computer, and about 20 sticks of RAM which felt kind of bad to just get rid of... but it was PC100 and PC133 stuff, even though it predates the earliest of DDR modules, it still kind of hurt to get rid of it given the current climate. But me wanting to clear clutter is much more important to me than hanging onto old junk that someone somewhere might find value in.
 
T4rd said:
Yup. Haven't looked at current pricing for them, but I got one for my daughter's PC for Christmas last year for around $250 and it has been great for her. Techpowerup shows it being about 17% faster on average as well, so even for about the same price I would prolly go with the B580 still.

Also, I'm not sure the point in bringing this back still when the RTX 5050 exists for considerably cheaper than $340 ($290 right now on Amazon) and is also about 14% faster according to Techpowerup's database. I would rather have a bit more performance (esp. using transformer model DLSS and access to MFG) and deal with the 8GB VRAM while being cheaper. Hell there are 5060s on Amazon right now for the same $340 that are 44% faster as well. I don't see that extra 4GB VRAM making the difference in most games at 1080p.
Click to expand...


3060 8nm old tech chip easier to make probably
 
MrGuvernment said:
is a 3060 powerful enough to use 12GB of ram? Just thinking back to the days of AMD days, when most games couldn't even use 8G if you got those models, or some of the other NVIDIA cards that were re-released with more RAM, but could never actually use it...
Click to expand...
It is and it isn’t…. At 1080p you aren’t likely to need the memory, but I can’t say with any certainty that will still be the case 6 months from now.
 
I would almost purchase the Intel B580 before the 3060 12GB.
1783308929080.png
1783308981007.png
 
Intel is currently ready to sell 272mm of fully enabled TSMC 5N chips with 12GB of 19 Gbps memory (450 GB/s bandwith) for a bit cheaper than the upcoming revival 276mm of 93% enabled samsung 8nm with 15 Gbps memory (360 GB/s)

I think for many, they should buy the B580 before the 3060 even if they were the same price, would just make sure their favorite games do run well before doing it... for how long that would be possible, (sold by them) NewEgg has only 2 sku, BestBuy 0, the 3060 12GB does not make much sense versus the 5050/5060/B580 and others right now, but that could be a sign... that they think it will in 2-3 months.
 
Last edited:
Ive been playing Crimson Desert lately and it shows how much VRAM is being used while playing and at 2560x1440 on a RTX 5070 it usually sits at around 4.5GB. I also sometime play the game on another TV using Geforce Now Ultimate at 4K resolution and it shows using about 5.5GB of VRAM. You guys are talking about games being unplayable at 1080p....is this maybe because of poor optimization and not because 8 or 12GB is legitimately unusable?
 
westrock2000 said:
Ive been playing Crimson Desert lately and it shows how much VRAM is being used while playing and at 2560x1440 on a RTX 5070 it usually sits at around 4.5GB. I also sometime play the game on another TV using Geforce Now Ultimate at 4K resolution and it shows using about 5.5GB of VRAM. You guys are talking about games being unplayable at 1080p....is this maybe because of poor optimization and not because 8 or 12GB is legitimately unusable?
Click to expand...
I haven't run into anything that wouldn't run on an 8GB card, and I don't expect any significant changes there for years. There are just far too many 8GB cards out there. At less than 8GB you have to go all the way back to the RTX 2060, RX 5600XT and GTX 1660Ti/Super to find a "60" card with less than 8GB. So early 2020 at the latest.

I do have a couple that are a gimped on settings in 8GB even at 1080p. Indiana Jones and the Great Circle and Doom the Dark Ages can't run max settings in 8GB no matter how fast your card is. In Indy you're stuff with a minimum sized texture pool. DTDA is a little better but you still can't max it out. It doesn't make a lot of difference in Doom. The game looks like hell anyway. >8GB matters more with Indy. It still can't quite max out the texture pool with 12GB, and while Doom is an ugly hellscape covered in dirt and rock Indy tries to be pretty and realistic. Indy is also a slow paced game so you don't need as much FPS. Same goes for path tracing between those two. For Indy it's a hell yes if your rig can do it. The difference is blatantly obvious and path tracing looks much better. DTDA nah turn it off not worth the FPS hit. I think I'd rather play Indy on a B580 (I have one) than an 8GB 5060Ti, but for Doom it's the opposite. IIRC they don't support 4k on 8GB cards even with a pile of DLSS, or at least Indy doesn't. Borderlands 4 also requires 8GB, but I haven't read up on that one so I'm not sure how it behaves in 8GB.

I think what we're going to see is more games where 8GB is the minimum, 8GB can't run max settings, and 8GB + 4k isn't supported. All three of the games I mentioned use dynamic lighting. Hardware ray tracing for DTDA & Indy, and Unreal Engine 5 software Lumen for BL4. That's a big driver for the 8GB minimum. Game devs like it because they don't have to bake lighting. It makes game development much faster since they're not waiting for a server farm to bake lighting, test it, make changes, then do it again. It also shrinks game downloads - ID estimated DTDA would be 50-100GB larger if they hadn't required RT global illumination.
 
westrock2000 said:
.is this maybe because of poor optimization and not because 8 or 12GB is legitimately unusable?
Click to expand...

It all depends on the game, the settings requirements etc. like RT, just global RT illumination, dlss, frame rate doubling, all of those techs take ram to run. More ram lets you use higher quality textures.
 
I run an RTX 4070S with 12GB of vram, and at 2560x1440 with settings maxed out in many games and RT as well as PB RT enabled in those that support it, along with a touch of DLSS and FG at times (not all the time) - I never even get close to running out of vram.
 
I calculated in Canadian dollars it would be $450 CAD. I'm seeing over $500 CAD right now. Any chance of seeing this go down over time or worse?
 
wandplus said:
I calculated in Canadian dollars it would be $450 CAD. I'm seeing over $500 CAD right now. Any chance of seeing this go down over time or worse?
Click to expand...

Everything costs more to bring it into Canada. Smaller market, regulatory differences like French language laws, taxes, the fact that they can and Canadians just take it, etc. Doing a straight conversion won't reveal the sell price like you'd think it would. We can thank our brilliant government for most of it though. Elbows up!
 
You must log in or register to reply here.
Back
Top