(Updated) AMD Navi GPU to Offer GTX 1080 Class Performance at ~$250 Report Claims

[Valentyn]

2 minutes ago, leadeater said:

Even with the updated drivers though there are plenty of games that gain very little to 15% at best, maybe Nvidia has just hit their scaling limit with CUDA cores like AMD did with stream processors ages ago (current game engine wise).

 

I do actually agree with what Nvidia is doing, everyone else is saying go general compute and scale like hell to brute force the performance but I just don't see that as sustainable. Fixed function extreme fast mixed with general purpose seems to be a much better idea to me. Be smarter and find ways to cut out necessary computation before it gets done.

I think it's a mix of that, and CPU dependant games that are outliers like that on the tests.

Going bigger and bigger just isn't working anymore it seems; specially when looking at the costs associated with those increases.

Shifting to a new core type, and mixing it all up is a fantastic tactic. Especially for offloading work loads; similar to AMD using FP16 in some games.
 

Hopefully AMD will embrace FP16 on their consumer cards also; so that they can take advantage of that in upcoming games that use it.

[laminutederire]

9 minutes ago, Valentyn said:

On the latest gaming reviews, that's moved up to 30-40% with the latest drivers.
https://www.hardocp.com/article/2018/03/20/nvidia_titan_v_video_card_gaming_review

 

Even so you're right, Big Volta is an excellent indication of what to expect on the up coming smaller GPUs. Especially if they're aren't cutting out the Tensor Cores.

It isn't necessarily though. Realistically you won't see the largest nodes on mainstream if they cannot improve their arch. At best they can buy time with a one tier leap by increasing core counts. Allowing them to do this a second time if AMD competes too well (basically if Navi is making a two tier jump, they'll have to do not as well and risk not having something significantly better afterwards. Otherwise they'll span the two tier jump in two dinstinct generations).

It remains that except for tensor cores which use in gaming is only useful with their raytracing API, the rest of the arch us pretty much the same since Maxwell with a few tweaks here and there but the only way it improves really is through higher core count and higher clocks. Which allowed the Pascal jump, and allows the Volta jump as well.

Which is why it's unrealistic to hope too much of Nvidia based on Volta.

[Trixanity]

3 hours ago, leadeater said:

I was assuming Tensor cores were not going to be in the gaming cards but the recent raytracing announcements have confirmed that they will indeed have some amount of Tensor cores.

Nvidia said tensor cores aren't a prerequisite for raytracing but wouldn't explain how or why.

[Valentyn]

1 minute ago, Trixanity said:

Nvidia said tensor cores aren't a prerequisite for raytracing but wouldn't explain how or why.

They've never been a prerequisite; but they significantly speed up the process. 

[Trixanity]

3 minutes ago, Valentyn said:

They've never been a prerequisite; but they significantly speed up the process. 

Nvidia said they could use the tensor cores to train a NN to reconstruct an image using fewer rays. That's it. Other than that, Volta does not use the tensor cores for raytracing whatsoever hence I doubt it'll be a thing for most gaming cards since they'll still be able to do fast raytracing without them. It's a combination of hardware and software that Nvidia won't reveal the specifics of.

[Razor01]

16 minutes ago, Trixanity said:

Nvidia said they could use the tensor cores to train a NN to reconstruct an image using fewer rays. That's it. Other than that, Volta does not use the tensor cores for raytracing whatsoever hence I doubt it'll be a thing for most gaming cards since they'll still be able to do fast raytracing without them. It's a combination of hardware and software that Nvidia won't reveal the specifics of.

 

 

Ah no they won't be able to fast raytracing with them lol.  Without them I'm pretty sure they will be getting 30 FPS or so with today's detail level in games.  They have revealed it and released the libs for it, I was hoping one of those reviewers that bought Titan V was going to be able to test it out, but guess not lol.

[Trixanity]

1 minute ago, Razor01 said:

 

 

Ah no they won't be able to fast raytracing with them lol.  Without them I'm pretty sure they will be getting 30 FPS or so with today's detail level in games.  They have revealed it and released the libs for it, I was hoping one of those reviewers that bought Titan V was going to be able to test it out, but guess not lol.

I have no idea what you're saying  ¯\_(ツ)_/¯

[laminutederire]

34 minutes ago, Trixanity said:

Nvidia said they could use the tensor cores to train a NN to reconstruct an image using fewer rays. That's it. Other than that, Volta does not use the tensor cores for raytracing whatsoever hence I doubt it'll be a thing for most gaming cards since they'll still be able to do fast raytracing without them. It's a combination of hardware and software that Nvidia won't reveal the specifics of.

If I understood it right, they send a few rays, get a super noisy image, and use a neural network (probably derived from a CNN) to denoise more efficiently the image to get something smooth.

The raytracing part does not seem to be particularly optimized in Volta (as in, the architecture does not try to make the ray tracing part faster), but they leverage tensor cores for the nn evaluation. That's why it's consistently slower without them as it's pure ML oriented.

[JediFragger]

3 hours ago, Ryujin2003 said:

I can only imagine what that's going to mean for the new entry price of PC gaming on green side.

Wouldn't surprise me if the lower-end cards are just re-branded Pascal's and the upper range ones the new architecture.

[Trixanity]

5 minutes ago, laminutederire said:

If I understood it right, they send a few rays, get a super noisy image, and use a neural network (probably derived from a CNN) to denoise more efficiently the image to get something smooth.

The raytracing part does not seem to be particularly optimized in Volta (as in, the architecture does not try to make the ray tracing part faster), but they leverage tensor cores for the nn evaluation. That's why it's consistently slower without them as it's pure ML oriented.

This is pretty much what we know of it:

Quote

...not many technical details are being disclosed, making it difficult to piece together what appears to be a multi-layered technology. NVIDIA could only confirm that some indeterminate functionality in Volta does accelerate ray tracing, and that RTX is a mix of both hardware – NVIDIA also described Volta in a separate blogpost as having a "ray tracing engine" – along with various bits implemented in software running on a GPU's CUDA cores. Meanwhile NVIDIA also mentioned that they have the ability to leverage Volta's tensor cores in an indirect manner, accelerating ray tracing by doing AI denoising, where a neural network could be trained to reconstruct an image using fewer rays, a technology the company has shown off in the past at GTC. RTX itself was described as productizing certain hardware and software algorithms, but is distinct from DXR, the overlying general API.

What I can glean from that is Nvidia has changed things in the backend to make raytracing faster while having optimized software running on the CUDA cores. Meanwhile they can leverage tensor cores to accelerate certain things but that the heavy lifting is done elsewhere.

[Razor01]

42 minutes ago, Trixanity said:

I have no idea what you're saying  ¯\_(ツ)_/¯

Think about the demos we have seen with Dx's Ray tracing or Vulkan.  No animated objects,  The reason for this is because ever time animated polys are in the scene the ray tracing is coming from the the vertex locations, animated polys are going to force a full scene rerender   The more animated polys ya have the more horse power its going to take.

 

The denoising is needed because the more rays we have in a scene the heavier its going to be heavy on the GPU.  So for now we have to use as few rays as possible.  While using few rays the scene gets noisy, because there are going to pixels that aren't rendered, that is where the tensor cores come in to predict what the values of those missing pixels are.  This is the heaviest part of raytracing the amount of rays being used.

[Razor01]

3 hours ago, laminutederire said:

It isn't necessarily though. Realistically you won't see the largest nodes on mainstream if they cannot improve their arch. At best they can buy time with a one tier leap by increasing core counts. Allowing them to do this a second time if AMD competes too well (basically if Navi is making a two tier jump, they'll have to do not as well and risk not having something significantly better afterwards. Otherwise they'll span the two tier jump in two dinstinct generations).

It remains that except for tensor cores which use in gaming is only useful with their raytracing API, the rest of the arch us pretty much the same since Maxwell with a few tweaks here and there but the only way it improves really is through higher core count and higher clocks. Which allowed the Pascal jump, and allows the Volta jump as well.

Which is why it's unrealistic to hope too much of Nvidia based on Volta.

the tensor cores are decoupled from the rest of the GPU just like the DP units.  I don't think they are decoupled from the DP units though, because if there are problems with those DP units or tensor cores, it would make sense to turn them all off, not just one set since they they pretty much need each other when doing certain work loads.  Its a possibility that if the Volta architecture comes down to next gen gaming cards, they will have tensor cores to some degree.  It might be very little like 1/32 amounts.

[Trixanity]

8 minutes ago, Razor01 said:

Think about the demos we have seen with Dx's Ray tracing or Vulkan.  No animated objects,  The reason for this is because ever time animated polys are in the scene the ray tracing is coming from the the vertex locations, animated polys are going to force a full scene rerender   The more animated polys ya have the more horse power its going to take.

 

The denoising is needed because the more rays we have in a scene the heavier its going to be on the GPU.  So for now we have to use as few rays as possible.  While using few rays the scene gets noisy, because there are going to pixels that aren't rendered, that is where the tensor cores come in to predict what the values of those missing pixels are.  This is the heaviest part of raytracing the amount of rays being used.

I haven't watched any of the demos. Just read the articles. Similar to how I consume Playboy magazine ( ͡° ͜ʖ ͡°)

 

So you're saying any raytracing needs ML techniques to make it feasible regardless of how beefy the hardware and how good the software implementation is (in relation to Volta or any other new arch)? If that's the case any chip would have to have some sort of neural processor onboard. That wouldn't bode well for the future of Nvidia's gaming cards or anything less than high end. They'd have to use the die space for that instead of adding CUDA cores or other things. Not to mention AMD cards - I haven't heard any plans to add dedicated neural processing so they'd have to rely on the NCUs to do that.

[Razor01]

1 minute ago, Trixanity said:

I haven't watched any of the demos. Just read the articles. Similar to how I consume Playboy magazine ( ͡° ͜ʖ ͡°)

 

So you're saying any raytracing needs ML techniques to make it feasible regardless of how beefy the hardware and how good the software implementation is (in relation to Volta or any other new arch)? If that's the case any chip would have to have some sort of neural processor onboard. That wouldn't bode well for the future of Nvidia's gaming cards or anything less than high end. They'd have to use the die space for that instead of adding CUDA cores or other things. Not to mention AMD cards - I haven't heard any plans to add dedicated neural processing so they'd have to rely on the NCUs to do that.

 

The way current graphics cards are setup, they cut down the DP units, I am expecting the same from the tensor cores.  If we are looking at 1/32 the cut downs, its not a much of an issue with die space. 

[Trixanity]

1 minute ago, Razor01 said:

 

The way current graphics cards are setup, they cut down the DP units, I am expecting the same from the tensor cores.  If we are looking at 1/32 the cut downs, its not a much of an issue with die space. 

But what of performance then?

[Razor01]

You don't need full performance right, the demo that we did see with RTX is like 10 years away from a real gaming experience.  That used 4 Titan V's lol and ungodly levels of polys lol and reflective surfaces.  Today's games won't be doing all that.  So one card with cut down tensors, I would say around 60 fps, but something runnable.  That is why without those tensor cores I don't its going be a very good gaming experience.  If this was 10 years ago and 30fp was a good target for new games, yeah, but with today's monitor technologies kinda need 60 fps.....

[Kamjam21xx]

If thats true... imagine what the $500 cards would do in quad crossfire.

[laminutederire]

1 hour ago, Trixanity said:

This is pretty much what we know of it:

What I can glean from that is Nvidia has changed things in the backend to make raytracing faster while having optimized software running on the CUDA cores. Meanwhile they can leverage tensor cores to accelerate certain things but that the heavy lifting is done elsewhere.

Oh yeah maybe they reworked a bit the culling to allow it to be more versatile in order to perform better in intersection. That would probably mean a rework of the cache structure a bit. From what I gathered in my computer graphics class, the hard part about ray tracing on GPU is actually the intersection part. With CPU rendering, you can build fancy spatial trees or hierarchies to accelerate intersection which make up for a big part of computation, since its somewhat expensive, and repeated a lot (for each ray). Issue with GPU being that you work with smaller cache per core. Rasterization techniques can use caches for pools of cores efficiently since you have coherency most of the time of the data used (same textures,same objects etc.). You can have this as well with tweaks for ray tracing. Issue comes when you want to perform path tracing, as following rays reflected onto surfaces can go in very different directions if you do it the CPU way. You then have to tweak memory hierarchy to allow for this eventuality.

As @Razor01 said, if you need movement it makes things harder because the data structure  of static scenes falls apart as is. You need to think about how to efficiently stream changes and allocate resources to avoid cache misses and so on.

1 hour ago, Razor01 said:

the tensor cores are decoupled from the rest of the GPU just like the DP units.  I don't think they are decoupled from the DP units though, because if there are problems with those DP units or tensor cores, it would make sense to turn them all off, not just one set since they they pretty much need each other when doing certain work loads.  Its a possibility that if the Volta architecture comes down to next gen gaming cards, they will have tensor cores to some degree.  It might be very little like 1/32 amounts.

I think they will have a great deal of them, but Nvidia will surely make it modular in a way that allows preventing the use of those through regular APIs to deter scientists to use them for research. (Not that many would necessarily use those anyway).

 

1 hour ago, Trixanity said:

 

So you're saying any raytracing needs ML techniques to make it feasible regardless of how beefy the hardware 

It does not necessarily needs it, but a good neural network implementation can make renders muuuuuuch faster as long as the denoising scales well enough. Disney uses that to make their rendering faster for instance. But in their case they shoot for quality, so their design takes a few minutes per image, which is okay considering the quality of the results would take hours more to get through rendering only. Issues then become robustness on scenes, motion etc.

[AlTech]

 

Quote

Gaming, wasn't it obvious how much trouble AMD was in with Maxwell.  I saw it with the 750ti, that card even though low end and first generation of Maxwell, showed so much potential.  I'm sure at that point AMD knew they were in deep shit.

 I'm sure that's why the R7 260X competed well against the GTX 750Ti and I'm sure that's why the R9 270X is now on par with the GTX 770 and GTX 680 despite the R9 280X being the primary competition against the GTX 770 at the time.  An

 

And I'm sure that's why the R9 390 competed well against the GTX 970. And I'm sure that's why the R9 380 and 380X competed well against the GTX 960.

 

The biggest threat to AMD other than Nvidia's illegal and/or unethical business practices with the GPP is Nvidia's customers and the mindshare of consumers.

Quote

Pascal video where he mentions it won't be much more over Maxwell because of process change,

Pascal is Maxwell on 16nm with some tweaks. It's not a totally new revamped architecture and it's so heavily targeted at Direct X 11 that Nvidia's Direct X 12 performance sucks because of the lack of Direct X 12 dedicated hardware.

Quote

People with any base knowledge of how chips are designed and manufactured will cringe at the things he has stated.  There was no way he could support Vega, even the most diehard fans know its not very good at gaming when it comes to everything it has shown vs the competition.

 

After that he couldn't say anything about Vega because of how wrong he was about his "forecasting", its not even forecasting its throwing a dart blind folded with a red blind fold on.   So GPU side he had to tone it down a bit and go the other way. 

Ummm excuse me?

 

Nvidia felt threatened by Vega 56 to the extent that they greenlit a GTX 1070Ti to take it head on and limited factory clockspeed not to cannibalize GTX 1080 sales.

Quote

 

Big differences in errors.  One is lack of knowledge and passing it off as he has some, which is ignorance and damages his audience, vs a person that has the knowledge and does a good job getting that across but not telling us where and how they got the tests to being with.

 

Right now it looks to me like you're going on a crusade against AdoredTV for not being an Nvidia fanboy. Personally, I think fanboying is stupid imho.

 

Business should be rewarded for making good products by us buying their good products.

[Mira Yurizaki]

5 minutes ago, AluminiumTech said:

Pascal is Maxwell on 16nm with some tweaks. It's not a totally new revamped architecture and it's so heavily targeted at Direct X 11 that Nvidia's Direct X 12 performance sucks because of the lack of Direct X 12 dedicated hardware.

What's "DirectX 12 dedicated hardware"?

[Razor01]

10 minutes ago, AluminiumTech said:

 

 I'm sure that's why the R7 260X competed well against the GTX 750Ti and I'm sure that's why the R9 270X is now on par with the GTX 770 and GTX 680 despite the R9 280X being the primary competition against the GTX 770 at the time.  An

 

And I'm sure that's why the R9 390 competed well against the GTX 970. And I'm sure that's why the R9 380 and 380X competed well against the GTX 960.

 

Ah they didn't compete well that is why we saw the huge market share decline from AMD starting with the Maxwell line of cards!

10 minutes ago, AluminiumTech said:

 

The biggest threat to AMD other than Nvidia's illegal and/or unethical business practices with the GPP is Nvidia's customers and the mindshare of consumers.

Pascal is Maxwell on 16nm with some tweaks. It's not a totally new revamped architecture and it's so heavily targeted at Direct X 11 that Nvidia's Direct X 12 performance sucks because of the lack of Direct X 12 dedicated hardware.

Ummm excuse me?

 

You are excused lol, because what lack of DX12 dedicated hardware?  Up till recently with Vega, nV was ahead of AMD with DX12 their tier levels.  And if you are talking about Async compute, you need to talk to a programmer about that.  Oh yeah I am a graphics programmer lol.  Well that is just incorrect then.  You don't seem to understand what the needs of being DX12 compliment are.  Concurrent processing of graphics and compute kernels is not a requirement of DX12.  Even though Pascal can do that too though.  Its not a good as GCN but then again it didn't need to be since it had base performance even higher.  Maxwell was capable of doing that as well, but only at the first partitioning of its units, after words unless the programmer took care of how they wrote shaders it would fall apart since the chip had to be flushed of all data before re partitioning was done.

 

Now do you want me to talk more about this?  I can at a lot more depth.

10 minutes ago, AluminiumTech said:

Nvidia felt threatened by Vega 56 to the extent that they greenlit a GTX 1070Ti to take it head on and limited factory clockspeed not to cannibalize GTX 1080 sales.

Right now it looks to me like you're going on a crusade against AdoredTV for not being an Nvidia fanboy. Personally, I don't find fanboying a hugely attractive trait and fanboying is stupid imho.

 

 

Personally I don't find ignorance highly unattractive, actually I find it outright distasteful.  Look above, if you want to talk about DX12 compliance, but don't know what it is then there is a problem.  Just like AdoredTV, he doesn't know what his talking about, his is ignorant of the facts, everything from yields, to EE, to programming.  You name it, he has made mistakes in ever single one of his videos.

[Razor01]

2 minutes ago, M.Yurizaki said:

What's "DirectX 12 dedicated hardware"?

 

He is talking about async compute units lol, wow, more shit from AMD marketing department.

[Mira Yurizaki]

4 minutes ago, Razor01 said:

He is talking about async compute units lol, wow, more shit from AMD marketing department.

I'm assuming as such, but I'd still like to hear what that is from someone who asserts this is a thing.

[AlTech]

1 minute ago, Razor01 said:

Ah they didn't compete well that is why we saw the huge market share decline from AMD starting with the Maxwell line of cards!

They did compete. It's just that Nvidia has been always in the mind of consumers whereas even if AMD had a better product they wouldn't be dominating Nvidia in sales.

1 minute ago, Razor01 said:

because what lack of DX12 dedicated hardware? 

Async Compute and similar stuff to it.

 

Maxwell and Pascal were designed for DX11.

1 minute ago, Razor01 said:

Up till recently with Vega, nV was ahead of AMD with DX12 their levels.

It's not like Nvidia's card are especially great at DX12. Nvidia is doing better than AMD because Nvidia has the money to spend on a good driver team.

 

Their drivers make great performance come out of mediocre and moderate hardware.

1 minute ago, Razor01 said:

You don't seem to understand what the needs of being DX12 compliment are.

*Compliant.

 

I do. Async Compute is a nice optional extra but is hugely beneficial for gaming when implemented correctly. Unfortunately we don't see many implementations because Nvidia isn't good at Async Compute.

1 minute ago, Razor01 said:

  Concurrent processing of graphics and compute kernels is not a requirement of DX12.  Even though Pascal can do that too though.  Its not a good as GCN but then again it didn't need to be since it had base performance even higher. 

What do you mean base performance?
 

DX11 isn't base performance.

1 minute ago, Razor01 said:

Personally I don't find ignorance highly attractive, actually I find it outright distasteful.  Look above, if you want to talk about DX12 compliance, but don't know what it is then there is a problem.  Just like AdoredTV, he doesn't know what his talking about, his is ignorant of the facts, everything from yields, to EE, to programming.  You name it, he has made mistakes in ever single one of his videos.

Linus makes mistakes in every single video he makes. Are you gonna call him out for it?

[Razor01]

22 minutes ago, AluminiumTech said:

They did compete. It's just that Nvidia has been always in the mind of consumers whereas even if AMD had a better product they wouldn't be dominating Nvidia in sales.

Bullshit!  Please look up market share numbers and product release cycles!  You will see there are strong trends in favor of AMD, when they had good products, products that competed well with nV or beat them they have always gained market share.  Don't make blanket reasoning like this without factual backup, specially when its easy to find!

22 minutes ago, AluminiumTech said:

Maxwell and Pascal were designed for DX11.

It's not like Nvidia's card are especially great at DX12. Nvidia is doing better than AMD because Nvidia has the money to spend on a good driver team.

 

Their drivers make great performance come out of mediocre and moderate hardware.

*Compliant.

 

I do. Async Compute is a nice optional extra but is hugely beneficial for gaming when implemented correctly. Unfortunately we don't see many implementations because Nvidia isn't good at Async Compute.

What do you mean base performance?

 

Yeah and you are wrong lol.  What else I can say.  Load up a DX12 game and profile it and you will see concurrent execution of kernals on Pascal and GCN they will function the exact same way, when it concerns async compute!

 

Again, that is something that can be shown easily.  Maxwell couldn't do async compute after the first partitioning.  That is why programmers had to optimize differently or leave that out.  This has nothing to do with dedicated hardware per se at least not with async compute units.  It was the ability to partition the SM's on a per cycle need.

 

Base performance = throughput, nV is so far ahead of AMD its not funny.  That is why nV with 1/3 to 1/2 less the compute performance can still keep up with AMD cards in gaming.

22 minutes ago, AluminiumTech said:

Linus makes mistakes in every single video he makes. Are you gonna call him out for it?

Making a mistake is one thing being ignorant is completely something else. Look up the definition.  Ignorant is when a person thinks he is right about something even though he/she is wrong, look above about async compute and dedicated hardware, those don't equal each other.

 

Async compute isn't even hugely beneficial,  if getting 5 to 10% extra performance where AMD cards were underutilizing their CU units is a good thing, its better to increase your base utilization then you don't need async compute to give you the performance back.  Async compute doesn't give you extra performance above what the GPU is capable of. It gives you back resources that were otherwise not usable before.  Of course there will be times undersutilization always takes place, because code can't be broken up all the time perfectly for optimal throughput, (critical path code, race conditions, etc become problems).  This is where concurrency helps out, to fill those otherwise unusable gaps in.

 

There is no free lunch with async compute.