NVIDIA GeForce 3070/3080/3080 Ti (Ampere): RTX Has No Perf Hit & x80 Ti Card 50% Faster in 4K! (Update 6 ~ Specs / Overview / Details)

[ShinRamen]

25 minutes ago, Yanashydo said:

Feeling the same way. I have the money, to buy 4,000 EUR PC, but to build it now, would be an insanity. 3Q/4Q should come out new GPUs, new Ryzen CPUs. I originally wanted to build the PC for the Cyberpunk 2077 release, but if I would have to wait a month or two for the new components, I would do that. 

I wanted to build a new rig about 1.5 years ago, then life came in the way and after that I decided to wait for Ryzen 3000. Suddenly Nvidia 3000 and Big Navi get closer and closer. It'd be great to have new Ryzen 4000 and Ampere out at about the same time and get a new top-end rig, but we'll see - top end models will most likely be released 1-2 months after the release of the rest.

I'm sure Nvidia wants their new gen out as soon as possible, but AMD might be first - there've been some rumors that Nvidia will paper-launch the new Titan just to keep the gpu-performance-crown, but mainstream cards might take until the very end of this year.

[igormp]

9 hours ago, Bombastinator said:

Though this new rtx tensor core stuff may be interesting for machine learning people.

 

9 hours ago, GDRRiley said:

I swear nivida will screw over its users again if they don't give these cards enough vram. the 9xx didn't have enough just like the 400/500 series.

This very much! Even though the tensor cores have been a blessing for anyone working with FP16, the lack of vram on regular cards makes really hard to do anything state-of-the-art. Having something like 16gb on their top GP102 consumer card would be awesome, although I doubt this will happen

[TOMPPIX]

On 10/30/2019 at 10:18 PM, Flying Sausages said:

Bitcoin miners will destroy the msrp really quick. I hate those mf bitcoin miners ruining msrp and supply.

Nah, NO competition is destroying the price.

[GDRRiley]

15 minutes ago, igormp said:

This very much! Even though the tensor cores have been a blessing for anyone working with FP16, the lack of vram on regular cards makes really hard to do anything state-of-the-art. Having something like 16gb on their top GP102 consumer card would be awesome, although I doubt this will happen

funny thing is, right now a Radeon VII in Reslove 16 which loves GPU memory and compute power. It equals 2080ti (both have very close raw FP32 and FP16 numbers) but that bandwidth for GPU memory.

[Axeonelite]

I need them to chill with these new graphics cards. I wallet can't keep up! I had a 1070 when they first came out and now I have a razer blade that i JUST got with a 2070mq. 

[Bombastinator]

54 minutes ago, igormp said:

 

This very much! Even though the tensor cores have been a blessing for anyone working with FP16, the lack of vram on regular cards makes really hard to do anything state-of-the-art. Having something like 16gb on their top GP102 consumer card would be awesome, although I doubt this will happen

It sounds like what is needed then is a tensor core apu or iGPU. Something that uses system ram which is variable.  Iirc there are other types of graphics solutions that use ram from elsewhere as well.  With one of those the ram wouldn’t be as fast, but one could assign as much as the whole machine could hold.  There are motherboards that will take 256gb. 

[RejZoR]

If there will really be no performance hit from RTX, I might actually go with it. One of rare times I might grab RTX card just because of a single game (Cyberpunk 2077). And curiosity. Also Marty McFly's RTGI shader.

[StDragon]

Will 

15 minutes ago, RejZoR said:

If there will really be no performance hit from RTX, I might actually go with it.

Yeah, that is interesting. I game at 2560x1080 (UltraWide) anyways, and the 1060 Super has been working perfectly for me. Runs Doom (2016 and Eternal) butter smooth. I really hope there's an equivalent 3060 as I would seriously be in the market for one of those if only to play Cyberpunk.

[GDRRiley]

23 minutes ago, Bombastinator said:

It sounds like what is needed then is a tensor core apu or iGPU. Something that uses system ram which is variable.  Iirc there are other types of graphics solutions that use ram from elsewhere as well.  With one of those the ram wouldn’t be as fast, but one could assign as much as the whole machine could hold.  There are motherboards that will take 256gb. 

if you need lots of ram it is usually needs it fast too. nothing says you couldn't make a GPU with 4 or 8 sodim ram slots.

point still stand Nvida needs to stop trying to cripple and limit cards with VRAM.

[Bombastinator]

19 minutes ago, RejZoR said:

If there will really be no performance hit from RTX, I might actually go with it. One of rare times I might grab RTX card just because of a single game (Cyberpunk 2077). And curiosity. Also Marty McFly's RTGI shader.

I’m not seeing “no performance hit” I’m seeing “less performance hit”. Might be none or near none.  Usefully low.  One of those things we likely can’t find out about till it arrives.

[Bombastinator]

5 minutes ago, GDRRiley said:

if you need lots of ram it is usually needs it fast too. nothing says you couldn't make a GPU with 4 or 8 sodim ram slots.

point still stand Nvida needs to stop trying to cripple and limit cards with VRAM.

Depends on what the cards are made for though.  They’re not actually made for machine learning, they’re made for doing gaming video.  They aren’t crippled for that.  My understanding is that attaching memory to video cards is really really complicated and has hard limits based on  trace length in the card itself.  Might be possible to go with some sort of cache thing.  Use what fast memory can be added but allow the card to use that memory as cache for system memory which is slower but larger.  Maybe that’s not viable.  I don’t know how machine learning works. 

[GDRRiley]

Just now, Bombastinator said:

 They’re not actually made for machine learning, they’re made for doing gaming video. 

cool so 1440p right now can eat up 6gb+ with modern games given a 3080ti should be doing 4K+ 16gb+ is what I want to see.'

many get used for compute workloads like Video editing or 3d modeling

[Void Master]

This means that we will see the RTX 3050 and the 3050 ti ?

Didn't NVIDIA say that they will include ray tracing across their whole graphics card lineup ?

What about the "RT" 3030 .

Man that would be a cool GPU ...

 

[Bombastinator]

35 minutes ago, GDRRiley said:

cool so 1440p right now can eat up 6gb+ with modern games given a 3080ti should be doing 4K+ 16gb+ is what I want to see.'

many get used for compute workloads like Video editing or 3d modeling

There is gddr, ddr, series versions 1, 2, 3, 4, 5, and 6 ( the latest gddr)  memory series’s appear first as gddr, then generally ddr a generation later.

If you were willing to accept ddr5 (which isn’t even out yet) instead of gddr6 as the definition of “fast” a sodimm solution might be feasible.  Right now the fastest sodims are ddr4

[TetraSky]

Cool. Will still be waaaay out of my budget. Even the 2060 super cost over $500 (canada)... I miss the day I could get a high end GPU for $350~$400... Now I can't even get a low end one for that price.

[igormp]

2 hours ago, GDRRiley said:

funny thing is, right now a Radeon VII in Reslove 16 which loves GPU memory and compute power. It equals 2080ti (both have very close raw FP32 and FP16 numbers) but that bandwidth for GPU memory.

Too bad AMD can't into machine learning at all. Even my 2060S performs better than a Radeon VII with ROCm in both FP32 and FP16 on tensorflow benchmarks.

 

2 hours ago, Bombastinator said:

It sounds like what is needed then is a tensor core apu or iGPU. Something that uses system ram which is variable.  Iirc there are other types of graphics solutions that use ram from elsewhere as well.  With one of those the ram wouldn’t be as fast, but one could assign as much as the whole machine could hold.  There are motherboards that will take 256gb. 

It'd be no better than using your SIMD instructions on the CPU. That's what's usually done when vram gets really constraining on your workload, CPUs can go past 1tb without much problems.

[The_Strict_Nein]

New video from Nvidia - What has Jensen been cooking:

 

https://www.youtube.com/watch?v=So7TNRhIYJ8

 

[Bombastinator]

5 minutes ago, The_Strict_Nein said:

New video from Nvidia - What has Jensen been cooking:

 

https://www.youtube.com/watch?v=So7TNRhIYJ8

 

Some sort of 4u industrial thing.  An appliance maybe?  Not gonna fit in a consumer pc.

[RejZoR]

2 hours ago, StDragon said:

Will 

Yeah, that is interesting. I game at 2560x1080 (UltraWide) anyways, and the 1060 Super has been working perfectly for me. Runs Doom (2016 and Eternal) butter smooth. I really hope there's an equivalent 3060 as I would seriously be in the market for one of those if only to play Cyberpunk.

I mean, RTX 2060 was already sort of usable. Battlefield 5 had some initial performance issues, but when they fixed it, RTX was quite usable even on this lowest end RTX card at the time.

[C2dan88]

15 minutes ago, Bombastinator said:

Some sort of 4u industrial thing.  An appliance maybe?  Not gonna fit in a consumer pc.

New DGX module.

[Brooksie359]

20 minutes ago, RejZoR said:

I mean, RTX 2060 was already sort of usable. Battlefield 5 had some initial performance issues, but when they fixed it, RTX was quite usable even on this lowest end RTX card at the time.

I imagine the 3060 to destroy the 2060 in terms of raytracing performance and even be quite a bit ahead in normal performance. I do think if someone is planning on playing the new cyberpunk game then it is probably not a bad idea to upgrade to the next gen gpus if ampere is as good as the rumors say. 

[Bombastinator]

19 minutes ago, C2dan88 said:

New DGX module.

https://scicomp.aalto.fi/triton/usage/dgx.html
 

a DGX-1 apparently contains 8 v-100s

[Bombastinator]

39 minutes ago, RejZoR said:

I mean, RTX 2060 was already sort of usable. Battlefield 5 had some initial performance issues, but when they fixed it, RTX was quite usable even on this lowest end RTX card at the time.

I have heard quite the opposite pretty much everywhere.  May go to the definition of how “usable” is being defined.

[Hilltrot]

1 hour ago, Bombastinator said:

I have heard quite the opposite pretty much everywhere.  May go to the definition of how “usable” is being defined.

I get 60+ 1080p RTX on the 2060.

 

Is usable for you 300+ FPS at 4k?

[williamcll]

Here's the motherboard of the new DGX, seems like Jensen learned a thing or two from Linus: