Volta might not show up anytime soon in the gaming space

[Fluffmeister]

Compression algorithms can get you so far , they need more memory bandwidth just as much as AMD.

The biggest bottleneck inside a GPU comes from accessing what is referred to as global memory (GDDR5 , HBM etc). The issue is not only memory bandwidth but also latency. HBM offers improvements on both of these fronts. Only thing that prevents Nvidia from using HBM is cost and very limited production , trust me as soon as it is viable for them they will switch to HBM for the consumer products regardless of what AMD does.

They will, when they need to and as you say when they think it is viable, which isn't now.

 

[Vya Domus]

I don't think so, gddr6 is around the corner which will put it on same field performance wise as HBM. Even gddr5x is near HBM and its not even been push to its max yet. Issue with HBM besides the low yields atm is the inter-poser and mounting the chips to it. It does add possible failure point for more wasted chip's. I see no reason for them to use it if GDDR6 can match up pretty close to HBM at a much cheaper price.

I have seen this argument a million times , GDDR6 is not really that much faster than current GDDR5X , 14 Gbps vs 11 Gbps.

GDDR6 also does not help with latency pretty much at all.

HBM is where it's at in terms of high performance GPUs.

To give an example Nvidia uses HBM2 for the V100 core which is incidentally the fastest GPU on the planet. They'll never switch to GDDR6 because it's simply not as good.

GDDR6 is a stopgap in order to save some time while HBM becomes more viable. But make no mistake : it does not match it , it just doesn't.

 

[arbiter]

HBM is where it's at in terms of high performance GPUs.

To give an example Nvidia uses HBM2 for the V100 core which is incidentally the fastest GPU on the planet.

the market they are using HBM on is a completely different market then consumer gpu's. They can work around the latency for consumer gpu's no problem, they have been doing it for how many years now?

 

[xkm1948]

HBM may be awesome for certain workloads, but it may very well remain a niche product until it is CHEAPER to use than GDDR.

Quick example, when using GPU+GDDR, there can be quite a lot of combinations based upon the VRAM width(128bit, 192bit, 256bit and etc) simply by enabling and disabling group of VRAMs. When using HBM available combination is restricted, as changing bus width will mean a brand new design of interposer. Given the current high cost of HBM, it is only natural to use HBM for only the top of line halo products.

HBM will mature, however GDDR will probably be around for a very long time, serving low~mid end graphics cards.

On another hand, HBM would make perfect sense for consoles. They don't need so many different designs. It is quite likely the wide adoption of HBM will happen in the console system first.

 

[GorbazTheDragon]

I have seen this argument a million times , GDDR6 is not really that much faster than current GDDR5X , 14 Gbps vs 11 Gbps.

GDDR6 also does not help with latency pretty much at all.

HBM is where it's at in terms of high performance GPUs

You're hammering down the wrong doors... GP102 has the same, if not better memory bandwidth on a 384-bit bus. Sure, HBM should have latency advantages, but the SP throughput of GP102 and Vega is similar, but the 1080Ti blows the 64 out of the water in performance. So clearly AMD is doing something wrong somewhere. They might have better memory latency, but they don't have anything to show for it in terms of performance. Vega is worse in every regard it should outperform Pascal in (on basis of having HBM).

AMD is fixing a problem that doesn't exist... or at least is non-existent in relative terms to the fact that their arch is so much less efficient (practical/theoretical throughout, not power efficiency) than what NVidia has put on the table.

 

[Vya Domus]

They can work around the latency for consumer gpu's no problem, they have been doing it for how many years now?

It's catching up on them , like I said memory is by far the biggest bottleneck in a GPU. They can't really work around it. This isn't a CPU where they can just throw more cache/improve the cache hit ratio and be done with it.

Look how much memory technology has evolved in the context of GPUs and how much for CPUs. Yeah... they need every ounce of performance they can get.

the market they are using HBM on is a completely different market then consumer gpu's.

The market is different , the benefits are the same for both.

You're hammering down the wrong doors... GP102 has the same, if not better memory bandwidth on a 384-bit bus. Sure, HBM should have latency advantages, but the SP throughput of GP102 and Vega is similar, but the 1080Ti blows the 64 out of the water in performance. So clearly AMD is doing something wrong somewhere. They might have better memory latency, but they don't have anything to show for it in terms of performance. Vega is worse in every regard it should outperform Pascal in (on basis of having HBM).

You're just scratching the surface of something that is far more complicated than you make it look.

AMD is fixing a problem that doesn't exist.

Oh it sure as hell exists , regardless of what you think.

 

[xkm1948]

It's catching up on them , like I said memory is by far the biggest bottleneck in a GPU. They can't really work around it.

Care to back it up with some good old facts?

I feel like GPU architecture itself is the biggest bottleneck. As most performance improvement occurs are due to change of core structure, not using a new type of VRAM

 

[GorbazTheDragon]

I feel like @xkm1948 has the same idea as I do...

The fact is that NV does more with less silicon, less bandwidth, and less power. AMDs chips wouldnt be memory starved if they actually made the arcitecture and the drivers better.

 

[Vya Domus]

Care to back it up with some good old facts?

I feel like GPU architecture itself is the biggest bottleneck. As most performance improvement occurs are due to change of core structure, not using a new type of VRAM

You can pretty much look at every OpenCl/CUDA documentation where they talk about the limitations of the memory hierarchy system inside GPUs where the biggest limiting factor is the global memory. You might say that only applies is in the case of compute but it doesn't , even when doing graphics you run into the same problems , hardware limitations don't go away.

Of course the architecture itself matters a lot and can contain it's own bottlenecks , I am speaking in a general context of an already existing architecture where faster/lower latency memory definitely helps.

 

[Fizban]

Toys for gamers.


Glad they see how their customer base loves toys, not graphics cards that cost hundreds of dollars, but toys...... cause we are all teens living in our parents basements, playing with toys.


What a dbag.

If you'reusing it to play games, it basically is a toy though...

 

[rtwjunkie]

Translation: no competition from AMD

That was my exact mental translation as I read what he said. "We have no one truly pushing us, so we will just keep what we have out there."

So unless Mr. Huang had a crystal ball, or was privy to privileged information,

That's actually a lot more common than you might think among business rivals. Corporate spying is probably as big as spying between countries. I don't believe there is much that either of them is able to keep totally secret from the other.

 

[GorbazTheDragon]

You can pretty much look at every OpenCl/CUDA documentation where they talk about the limitations of the memory hierarchy system inside GPUs where the biggest limiting factor is the global memory. You might say that only applies is in the case of compute but it doesn't , even when doing graphics you run into the same problems , hardware limitations don't go away.

Of course the architecture itself matters a lot and can contain it's own bottlenecks , I am speaking in a general context of an already existing architecture where faster/lower latency memory definitely helps.

This is so immensely application dependent... There are lots of tasks that have large numbers of muOps per data block, and others that have a very small number of simple Ops and scroll through large, often far apart (but interdependent) pieces of data.

With compute you cant generalise, quite simply because there is such a broad variety of tasks to be performed.

For example, I've studied the implementation of GPGPU for data reduction and correlation on radio telescope arrays, and there the primary limiting factor is that pre-pascal GPUs are not able to perform low (<4-bit) precision tasks at all efficiently due to the nature of the stream processors being made for 32-bit ops. The other limitation is not in memory latency but size, GPUs can't fit the datasets on-board as is needed for the data reduction steps. While on the other hand in the correlation steps neither bandwidth, size, or latency are problematic as the data is addressed entirely serially and there is no interdependency between time blocks.

 

[Vya Domus]

This is so immensely application dependent... There are lots of tasks that have large numbers of muOps per data block, and others that have a very small number of simple Ops and scroll through large, often far apart (but interdependent) pieces of data.

With compute you cant generalise, quite simply because there is such a broad variety of tasks to be performed.

For example, I've studied the implementation of GPGPU for data reduction and correlation on radio telescope arrays, and there the primary limiting factor is that pre-pascal GPUs are not able to perform low (<4-bit) precision tasks at all efficiently due to the nature of the stream processors being made for 32-bit ops. The other limitation is not in memory latency but size, GPUs can't fit the datasets on-board as is needed for the data reduction steps. While on the other hand in the correlation steps neither bandwidth, size, or latency are problematic as the data is addressed entirely serially and there is no interdependency between time blocks.

Just because you found some niche application that requires the use of a data format that isn't even natively supported by said hardware doesn't say much. That would be a bottleneck on pretty much every modern GPU/CPU.

Your example is also the worst possible use case of a GPU , since their functionality relies on independent access of memory in order to make use of the data parallelisation they are designed to do. Accessing data in a serial fashion is ridiculously wasteful on a GPU , I am sorry but your example is simply out of the scope of the typical use case scenario.

That was not my point however. Whether it is compute or graphics you use the same memory hierarchy and same execution resources that are bounded by the exact same hardware limitations. If you don't believe me read said documentation , there are a lot of mechanisms in place to deal with the issue of latency. Registers/shared memory operate at/or close to the actual clock speed , where as accessing global memory can take up to several hundreds of clock cycles. It is a huge bottleneck that is difficult to deal with on something like a GPU that needs to be fed data constantly in huge volumes , graphics or compute ! On CPUs this is easier to deal with since they don't produce anywhere near the throughput that a GPU does.

 

[TheoneandonlyMrK]

https://www.techspot.com/news/70584-nvidia-volta-gaming-gpus-not-foreseeable-future.html

Looks like Jensen himself has said the following :

"Volta for gaming, we haven't announced anything. And all I can say is that our pipeline is filled with some exciting new toys for the gamers, and we have some really exciting new technology to offer them in the pipeline. But for the holiday season for the foreseeable future, I think Pascal is just unbeatable. It's just the best thing out there. And everybody who's looking forward to playing Call of Duty or Destiny 2, if they don't already have one, should run out and get themselves a Pascal"

He also confirmed one of my suspicions , that Volta is very expensive to make in it's current iteration :

"The price of Volta is driven by the fact that, of course, the manufacturing cost is quite extraordinary,” said Huang. "These are expensive things to go and design."

There is a mention about Volta GPUs costing 1000$ to produce. I'd assume they are talking about V100 but I am not exactly sure what this price actually means , if it's a per-die cost for example.

Looks like they are doing the smart thing and might skip on this one.

I wonder what those "new toys" are though.

As I expected and divulged 2018 for Christmas and thats optimistic in the meantime Pascal refresh called volta for no real reason.

 

[R0H1T]

the market they are using HBM on is a completely different market then consumer gpu's. They can work around the latency for consumer gpu's no problem, they have been doing it for how many years now?

I feel this latency thing is being blown way out of proportion, do we have actual numbers as to how much slower (or faster) GDDR5x or GDDR6 is as compared to HBM, or how much would a normal GPU bottleneck due to latency?

 

[EarthDog]

Memory bandwidth is memory bandwidth , the more the better. The benefit of using HBM is very real.

indeed... at 4k, where less than 1% of steam users play.

Yes, bring it on now....its real.

Clearly power consumption isnt terribly relevant for nvidia now. Its leading that race already.

 

[Vya Domus]

I feel this latency thing is being blown way out of proportion, do we have actual numbers as to how much slower (or faster) GDDR5x or GDDR6 is as compared to HBM, or how much would a normal GPU bottleneck due to latency?

instruction throughput =instructions in execution / instruction latency , basic design law.

Come on people we are not reinventing the GPU. This is how these things work , memory latency and bandwidth are critical for achieving more performance.

indeed... at 4k, where less than 1% of steam users play.

So...how is this relevant ? We are talking about high end GPUs , you don't buy a 1080 and up to play at 720p.

 

[EarthDog]

You don't buy a 1080 for 4k either... you buy a 1080ti. You also don't buy vega 64 for 4k either...too slow even with gobs of bandwidth.

The point is nobody (less than 1%) uses it where it benefits them, 4k. So, yes, the benefit is real for those few people. It's too bad 99% can't utilize the benefits... my point. Ahead of its time at its own detriment.

Hbm isnt remotely needed now. Gddrx5 does a more than acceptable job, even at 4k. 3 years from now, different story. There will also be new gens of cards out by then as well.

 

[RejZoR]

If people only knew the wonders of 120Hz and more, they'd realize you can stick something this fast even on a 1080p screen... You know, when 60fps just isn't enough.

 

[dcf-joe]

If people only knew the wonders of 120Hz and more, they'd realize you can stick something this fast even on a 1080p screen... You know, when 60fps just isn't enough.

+1 to high-refresh gaming.

 

[GorbazTheDragon]

Latency means jack all for serialised data streams, a huge amount of compute tasks are extremely predictable, and a huge amount that require a lot of random data access fall well outside of the boundaries of what HBM is currently providing. It's a question of cost benefit, not just simple performance improvements.

 

[Solaris17]

Translation: no competition from AMD with a rubbish Vega so they're milking Pascal for all its got and are gonna give us "toys" to play with in the meantime.

And it's us who lose out. I'm sure if Vega had leapfrogged the 1080 Ti in framerate, power, heat and noise, that suddenly a Volta card would be close to release and none of these lame excuses.

Honestly I don't think this is the case. If you look back to the mid to early 2000s and even up until fermi both AMD and Nvidia have shortened there release cycles tremendously. I think the market is whats out of control and we are just used to it now. I am more then happy to hold on to my $700 GPU for more then 6 months before its depreciated. I think both nvidia and AMD just got ahead of themselves. They should slow release cycles back down to what it was. It would be better for consumers as well. better device and driver maturity. These arent cell phones no need to rush.

 

[the54thvoid]

.... I am more then happy to hold on to my $700 GPU for more then 6 months before its depreciated..... no need to rush.

Yeah, I agree, I discovered holidaying in Canada and it aint cheap. I can't afford a new top end GPU each year anymore AND holidays in Canada... And the bears win every time.

 

[phanbuey]

Not surprised... if Vega was more of a threat they may have pushed volta, but as it sits, the 1080ti is kind of on its own.

 

[qubit]

Honestly I don't think this is the case. If you look back to the mid to early 2000s and even up until fermi both AMD and Nvidia have shortened there release cycles tremendously. I think the market is whats out of control and we are just used to it now. I am more then happy to hold on to my $700 GPU for more then 6 months before its depreciated. I think both nvidia and AMD just got ahead of themselves. They should slow release cycles back down to what it was. It would be better for consumers as well. better device and driver maturity. These arent cell phones no need to rush.

hmmm... slower release cycles. It's that old dilemma between progress and getting a return on investment from your current product and I don't think that there's a single right answer there. If the release cycles are longer, then I think the performance jumps should be bigger.

It still makes sense to me that if Vega had been better then NVIDIA would have responded with the next gen GPU, whether that's good for customers or not.