Folding@Home Performance Is Looking Good On The GeForce RTX 2080 Ti

[audir8]

AMD could have a DLSS alternative pretty quickly since it's DL-based. It just means AMD would also need to dedicate GPU real estate for DL inference, and they've already done so to a limited extent in Ryzen. Though the benefits from DLSS would really need to hit it out of the park, which might not be true for all titles. 2080 vs 1080 comparisons would show any IPC gains Nvidia has really made, since the clocks and memory BW are more equal on those cards, though the 2080 Ti benchmarks do show more cores and more memory BW means significantly more performance. AMD has already had an alternative for the mesh-shading that hasn't seen wide adoption likely because it's Vega only, so the vendor-specific stuff really needs to go mainstream at lower price points before really having an affect on adoption in the market.

If DLSS can move downstream and offer 30+% better AA in a $200 card, that would be a big shift towards a combination of DL and linear algebra in a GPU.

[coder]

AMD could have a DLSS alternative pretty quickly since it's DL-based.

And certainly patented.

It just means AMD would also need to dedicate GPU real estate for DL inference, and they've already done so to a limited extent in Ryzen.

Wut? They did?

AMD has already had an alternative for the mesh-shading that hasn't seen wide adoption likely because it's Vega only, so the vendor-specific stuff really needs to go mainstream at lower price points before really having an affect on adoption in the market.

And they never enabled it. Assuming you're referring to primitive shaders, that is.

If DLSS can move downstream and offer 30+% better AA in a $200 card, that would be a big shift towards a combination of DL and linear algebra in a GPU.

Lol wut?

GPUs are already competent at linear algebra, because... 3D graphics. Conversely, Tensor cores are pretty bad for general linear algebra, because much of their benefit comes at the expense of precision.

[audir8]

And certainly patented.


Wut? They did?



And they never enabled it. Assuming you're referring to primitive shaders, that is.


Lol wut?

GPUs are already competent at linear algebra, because... 3D graphics. Conversely, Tensor cores are pretty bad for general linear algebra, because much of their benefit comes at the expense of precision.

I don't think patents can exclusively stop AMD or Intel from making a different API that does AA through DL until it's standardized.

Ryzen uses NNs for branch prediction/pre-fetch (though I think Intel does this too): https://www.anandtech.com/show/10907...hz-boost-steps https://www.amd.com/en/technologies/sense-mi

I do mean primitive shaders, and AMD might have have listened to game devs saying no more than Nvidia did. having 70+% market share means you can do things like this, but that doesn't guarantee adoption if the cards cost $800+.

Nvidia dedicating resources to Turing tensor cores means exactly what I said, if you read their white paper on page 10, they say these are the new cores used for DLSS.
https://www.nvidia.com/content/dam/e...Whitepaper.pdf

The tensor core parts of a full core seem significant too ~20% of a SM core? Picture of 4-core block on page 18.
RT cores also seem to be around ~20% of a SM core. So seeing how far ray tracing goes in this generation will be interesting to see too.

These are Nvidia's two bets, DL inference and ray tracing, separate from the general purpose linear algebra of GPUs until now.

[coder]

I don't think patents can exclusively stop AMD or Intel from making a different API that does AA through DL until it's standardized.

Agreed. I'm just pointing out an obstacle they'll face, when they even reach that point.

Ryzen uses NNs for branch prediction/pre-fetch

Oh, right. But it's extremely hard-wired and not at all exposed for general usage. It's not comparable to Tensor Cores in any way. Not just different animals, but like different phyla.

These are Nvidia's two bets, DL inference and ray tracing

True. Exciting times!

I actually think the RTX series is getting a bad rap. The only real gripe people have is on price. But if you dig into the gaming benchmarks out there, a RTX 2080 FE is about 4% faster than a GTX 1080 Ti. So, I'd pay like $50 more for that + its new features. The biggest worry I'd have is that its ray tracing might not be fast enough to actually use, as support for it matures.

If the pricing structure were a bit closer to their previous generation, I think there'd be much more enthusiasm about this launch. The other self-inflicted wound is their month-long buildup, which seemed designed to drive demand for their Pascal inventory.

[audir8]

coder Total guessing on my part, but I actually think AMD's NN implementation seems like online learning, as in, each time a processor boots and starts new threads, it's re-training the network. There aren't too many examples of this in the wild, if it really is online learning. It's hard-wired and everything, but still a NN replacing a part of the processing pipeline like AA in graphics. Call it however similar or dissimilar as you want.

I think DLSS could have a more immediate impact along with any other DL training/inference applications for the tensor cores, and ray tracing will have a larger longer-term impact once it matures. Really both could become integral parts of GPUs.

[coder]

coder Total guessing on my part, but I actually think AMD's NN implementation seems like online learning, as in, each time a processor boots and starts new threads, it's re-training the network.

Yes, that's the point. But it must be implemented in a very limited, power-efficient way that's customized to the specific problem of branch prediction. There's no way it has any convolution layers.

It's hard-wired and everything, but still a NN replacing a part of the processing pipeline like AA in graphics. Call it however similar or dissimilar as you want.

No, it's not just a matter of perspective. Nvidia's Tensor cores are quite general, by comparison, and really not tied to AI. Moreover, you were trying to make a practical point that Ryzen's NN branch prediction somehow positions them to offer a competing DLSS solution. This has no basis in reality.

Why get all defensive? You clearly strayed out of your depth. I advise simply dropping the point.

[audir8]

Yes, that's the point. But it must be implemented in a very limited, power-efficient way that's customized to the specific problem of branch prediction. There's no way it has any convolution layers.


No, it's not just a matter of perspective. Nvidia's Tensor cores are quite general, by comparison, and really not tied to AI. Moreover, you were trying to make a practical point that Ryzen's NN branch prediction somehow positions them to offer a competing DLSS solution. This has no basis in reality.

Why get all defensive? You clearly strayed out of your depth. I advise simply dropping the point.

I'm not sure my guessing was any worse than your guessing. I will give you it's more hardwired than tensor cores clearly, and probably doesn't take up much die space at all, but like I said, online learning means essential bits of a tensor core are all there.

Also, mentioning layers, and precision the way you did in the last two posts, shows you're out of your depth. There, I let it go.

[coder]

I'm not sure my guessing was any worse than your guessing.

Really? So, you're also building deep learning into commercial products?

I will give you it's more hardwired than tensor cores clearly, and probably doesn't take up much die space at all, but like I said, online learning means essential bits of a tensor core are all there.

No, it doesn't. A tensor core is so-called because it computes tensor products. That's a useful, generic programmatic abstraction in deep learning, but not an essential element. If you're embedding a hard-wired network that needs to be extremely power and area-efficient, then you wouldn't pay the cost of the extra data movement and control logic for something generic, like a tensor core. You'd hardwire all the multipliers and adders in a fixed topology. That's what hardwiring means.

Also, mentioning layers, and precision the way you did in the last two posts, shows you're out of your depth. There, I let it go.

WTF? If you want to take issue with something I write, go on and make your case. Snipes like these are just so much trash talk.

For serious linear algebra, you need GPUs with decent fp64 provisioning, since they don't implement IEEE 754-conformant floating point. Specifically, they lack denormals. Given that fp32 is lacking, you can imagine fp16 is virtually worthless, as a generic linear algebra platform. Tensor cores are really about signal processing, which typically has much more modest precision demands.