ai Looking for some GPU options.... my R9 290 is getting old

[TheNuzziNuzz]

12 minutes ago, Duckster said:

I'm sorry.... I don't know what to tell you but it seems like you will have to wait until it drops, or if it's out of stock wait until they can restock. This is a pretty idiotic idea, but you can buy a prebuilt with a 3070 and take the gpu and sell the rest....

yeah that's not worth it. And it's okay I'm not blaming you, it was silly to stay up til 4am programming haha!

[TheNuzziNuzz]

1 hour ago, Duckster said:

I'm sorry.... I don't know what to tell you but it seems like you will have to wait until it drops, or if it's out of stock wait until they can restock. This is a pretty idiotic idea, but you can buy a prebuilt with a 3070 and take the gpu and sell the rest....

NOOOO!!

 

I added a card to my cart, got into checkout, and it said "No stores available with card" then I tried again a few times and it actually let me select a store for pickup, after that I tried submitting my order and then it said "Item unaavable at store" so I tried with every best buy within 200 miles and the same thing happened. 

 

I wouldn't have cared if I just missed it and it was sold out, but I actually got to the payment info for a 6800, but I decided the extra ctost wasn't worth the risk of not having CUDA cores...which I need for Artificial Intelligence, so this purchase would be kinda wasted on me.

[LadyFelina]

Same same same... got no sleep and been up since 6 am trying to get a damn card. BB gave me the runaround for a good hour and now says sold out... I desperately want my son to build his own PC but can't find ANY video card in the $350-$500 range to purchase! I'm actually contemplating buying a pre-built and taking it apart for him to rebuild... but that's just insanity  

[Harry Voyager]

14 hours ago, TheNuzziNuzz said:

Yes, but also they don't have CUDA cores...which is what I need for tensorflow

 

It sucks because I've always leaned AMD, i have over 10 AMD cards in my house and no nVidia...but now that I'm doing so much work that can benefit from CUDA, ray tracing, and nVidia's just plain better software support....

Do you know if your analysis is limited by integer operations or by floating point operations? 

 

If you are integer limited you will get more mileage out of the the 2080 and 1080 series. The 3080 significantly improved the floating point performance, but does it by repurposing the integer cores. There was an "ask me anything" nVidia did on the architecture changes. 

 

This wasn't the article I was trying to find, but seems to cover the changes: NVIDIA's RTX 3000 cards make counting teraflops pointless | Engadget
https://www.engadget.com/nvidia-rtx-3090-3080-3070-cuda-core-int32-fp32-210059544.html

[TheNuzziNuzz]

21 minutes ago, LadyFelina said:

Same same same... got no sleep and been up since 6 am trying to get a damn card. BB gave me the runaround for a good hour and now says sold out... I desperately want my son to build his own PC but can't find ANY video card in the $350-$500 range to purchase! I'm actually contemplating buying a pre-built and taking it apart for him to rebuild... but that's just insanity  

OR you re-sell the system with a lesser-gpu you find. A lot of people are doing that. Could always go for an older card too. Fury's are selling for like 100, they have like 1/2 - 1/3 of the performance of a $400 3060ti, but you could actually get one. 

[TheNuzziNuzz]

17 minutes ago, Harry Voyager said:

Do you know if your analysis is limited by integer operations or by floating point operations? 

 

If you are integer limited you will get more mileage out of the the 2080 and 1080 series. The 3080 significantly improved the floating point performance, but does it by repurposing the integer cores. There was an "ask me anything" nVidia did on the architecture changes. 

 

This wasn't the article I was trying to find, but seems to cover the changes: NVIDIA's RTX 3000 cards make counting teraflops pointless | Engadget
https://www.engadget.com/nvidia-rtx-3090-3080-3070-cuda-core-int32-fp32-210059544.html

I honestly don't know, I'm running pre-trained SSD Mobilenet to detect faces in hundreds of gigabytes of video files. I have an R9 290 but there's no GPU acceleration with AMD, it takes an absolute eternity.

 

Does the performance gain with GPUs on AI only apply to training? I see people getting like over 6x performance with an nVidia card when it comes to training.

[wyliecoyote827]

the new cards are pretty great, but as everybody has told you by now they are impossible to find. the nice thing is old cards are going down in price. i saw an rtx 2080 dual for like 600 canadain on facebook. 2060 supers are down to 400, and this is all BEFORE the 3060ti launch. if you wait a week or so, you can probably get a 2070 super or better for almost nothing used.

 

you could always try to fight for a non ti 3060 card when they come out, but thats probably a while and no guarantee you will get it.

[TheNuzziNuzz]

27 minutes ago, wyliecoyote827 said:

the new cards are pretty great, but as everybody has told you by now they are impossible to find. the nice thing is old cards are going down in price. i saw an rtx 2080 dual for like 600 canadain on facebook. 2060 supers are down to 400, and this is all BEFORE the 3060ti launch. if you wait a week or so, you can probably get a 2070 super or better for almost nothing used.

 

you could always try to fight for a non ti 3060 card when they come out, but thats probably a while and no guarantee you will get it.

If I could find a 2070 for under $400, id jump right on that. Maybe i'd even go for a 2080. I don't want the best of the best, I just need something that has decent CUDA cores and can run my 3 monitors. 3060ti's are listed used for like $800 right now, but hopefully because of their low retail the 20 cards drop low.

[Duckster]

"When you check out the Q&A of 3060 ti..." 

 

This is actually real

Edited December 2, 2020 by Duckster
I'm just having fun at this point.... get 2070 used, as you said.

[LadyFelina]

1 hour ago, TheNuzziNuzz said:

OR you re-sell the system with a lesser-gpu you find. A lot of people are doing that. Could always go for an older card too. Fury's are selling for like 100, they have like 1/2 - 1/3 of the performance of a $400 3060ti, but you could actually get one. 

I wouldn't mind doing this... where can I find them at this price?

Except I'm guessing I wouldn't be able to upgrade later? Or could I get the same mobo/etc. that would work with a 3060 and use it today with a Fury? 

[Action_Johnson]

Do you really need it right now? 

 

If you're using software that actively takes advantage of Tensor and CUDA, AMD should be completely off the table. 

[igormp]

On 12/2/2020 at 2:34 PM, Harry Voyager said:

Do you know if your analysis is limited by integer operations or by floating point operations? 

 

If you are integer limited you will get more mileage out of the the 2080 and 1080 series. The 3080 significantly improved the floating point performance, but does it by repurposing the integer cores. There was an "ask me anything" nVidia did on the architecture changes. 

 

This wasn't the article I was trying to find, but seems to cover the changes: NVIDIA's RTX 3000 cards make counting teraflops pointless | Engadget
https://www.engadget.com/nvidia-rtx-3090-3080-3070-cuda-core-int32-fp32-210059544.html

ML is mostly FP operations (unless they're using some network that has gone through quantization), so Ampere does bring tons of benefits.

 

Even if the workload had a mix of INT+FP (hence working just like Turing), there would still be a perf increase due to the new arch with higher IPC anyway, not to mention the newer tensor cores that have some nice features.

 

On 12/2/2020 at 2:54 PM, TheNuzziNuzz said:

Does the performance gain with GPUs on AI only apply to training? I see people getting like over 6x performance with an nVidia card when it comes to training.

Also applies to inference, but that's not really relevant since inference is mostly done on end devices. Training is more relevant for performance numbers because that's where devs/researchers spend most of their time on.

[Mark Kaine]

On 12/1/2020 at 9:58 PM, TheNuzziNuzz said:

Its not that easy though, as you can't find these GPUs for their retail if at all.

 

I'm seeing them for over 1k, which so isn't worth it to me.

yeah true, which is why right now upgrading a pc might not be the best idea, especially before the holidays. 

 

1-2 months ago it looked a lot better, but alas... 

 

I would recommend to wait until jan/feb to make the decision tbh 

[Guest]

A used 1660 Super off ebay is great bang for buck.

[Downkey]

On 12/1/2020 at 3:51 PM, TheNuzziNuzz said:

So as we know, the market is pretty weird right now with everything going on, but I'm in need of an upgrade so I want to look at my options.

 

Price range is something between $300 and $1000 USD. I know that's a weird range and ill explain why

 

I've been running an R9 290 for a few years, I game occasionally, but primarily I do workstation tasks / video rendering / programming. Recently i've been getting in AI for one of my businesses, and I realized that Tensorflow only works with Nvidia GPUs. I also have started getting into 3d design, and the softwares just don't seem to take advantage of my AMD GPU.

 

My CPU is i5 8600k (I get like 4.5ghz), 32GB ram. And a few terabytes of SSD storage. This is my power supply: https://pcpartpicker.com/product/kKFPxr/raidmax-power-supply-rx700gh I know it looks sketchy but its actually quite awesome. 

 

Given my outdated Graphics platform, I don't even know where to start. Almost anything even on the low $200 end of new GPUs now seems to beat the 290 I have now. I'm a big fan of bang for the buck, but I'm not sure where that even exists today. Looking for some education and guidance.

 

From my research it seems nVidia has much better software support. I'm also totally open to used GPUs, back when I built computers I always used used GPUs and 95% of the time they worked amazing, like the 290 I have now that I got for $100 2 years ago.

 

Final note! I have 3 HDMI monitors (thinking about upgrading them too) connected via 1 HDMI, 1 DVI, 1 Displayport on my GPU with active adapters. I've also heard about new display port hubs?

 

Thank you guys!

A used 1080ti will do you just fine, as it's around 300 - 450 used

[TheNuzziNuzz]

3 hours ago, Downkey said:

A used 1080ti will do you just fine, as it's around 300 - 450 used

 

4 hours ago, Shabba said:

A used 1660 Super off ebay is great bang for buck.

 

3 hours ago, Downkey said:

A used 1080ti will do you just fine, as it's around 300 - 450 used

 

4 hours ago, Mark Kaine said:

yeah true, which is why right now upgrading a pc might not be the best idea, especially before the holidays. 

 

1-2 months ago it looked a lot better, but alas... 

 

I would recommend to wait until jan/feb to make the decision tbh 

 

5 hours ago, igormp said:

ML is mostly FP operations (unless they're using some network that has gone through quantization), so Ampere does bring tons of benefits.

 

Even if the workload had a mix of INT+FP (hence working just like Turing), there would still be a perf increase due to the new arch with higher IPC anyway, not to mention the newer tensor cores that have some nice features.

 

Also applies to inference, but that's not really relevant since inference is mostly done on end devices. Training is more relevant for performance numbers because that's where devs/researchers spend most of their time on.

 

On 12/2/2020 at 4:29 PM, Action_Johnson said:

Do you really need it right now? 

 

If you're using software that actively takes advantage of Tensor and CUDA, AMD should be completely off the table. 

 

On 12/2/2020 at 3:42 PM, LadyFelina said:

I wouldn't mind doing this... where can I find them at this price?

Except I'm guessing I wouldn't be able to upgrade later? Or could I get the same mobo/etc. that would work with a 3060 and use it today with a Fury? 

 

On 12/2/2020 at 3:39 PM, Duckster said:

"When you check out the Q&A of 3060 ti..." 

 

This is actually real

I found 2 1080s for $400 total locally. They work. WHAT

[Downkey]

18 hours ago, TheNuzziNuzz said:

 

 

 

 

 

 

 

I found 2 1080s for $400 total locally. They work. WHAT

Just get 1! Sli is useless

[igormp]

4 hours ago, Downkey said:

Just get 1! Sli is useless

They mentioned working with ML, so SLI isn't used and multiple GPUs are really useful.