Question 'Ampere'/Next-gen gaming uarch speculation thread

[Ottonomous]

How much is the Samsung 7nm EUV process expected to provide in terms of gains?
How will the RTX components be scaled/developed?
Any major architectural enhancements expected?
Will VRAM be bumped to 16/12/12 for the top three?
Will there be further fragmentation in the lineup? (Keeping turing at cheaper prices, while offering 'beefed up RTX' options at the top?)
Will the top card be capable of >4K60, at least 90?
Would Nvidia ever consider an HBM implementation in the gaming lineup?
Will Nvidia introduce new proprietary technologies again?

Sorry if imprudent/uncalled for, just interested in the forum member's thoughts.

 

[Mopetar]

NVidia ultimately can't stop crytpocurrency mining because there's no singular currency or algorithm for it. If anyone really wanted to they could make a currency where the the proof of work is a frame rendered for a PC game among the top ten most popular titles at a given moment. Nvidia couldn't cripple that without making their cards effectively pointless for gaming.

Maybe that sounds like an utterly stupid basis for a cryptocurrency, but I'm pretty sure it's just a way of describing a distributed Google Stadia service.

Mining has created an alternative use for a product that was previously used for running computer games. Right now that alternative use is being valued much more highly than being able to use a GPU to play games. It's no different than the price of milk skyrocketing if people were willing to pay three times as much for cheese. Solutions that try to rely on people not turning milk into cheese are just doomed to failure.

 

[darkswordsman17]

NVidia ultimately can't stop crytpocurrency mining because there's no singular currency or algorithm for it. If anyone really wanted to they could make a currency where the the proof of work is a frame rendered for a PC game among the top ten most popular titles at a given moment. Nvidia couldn't cripple that without making their cards effectively pointless for gaming.

Maybe that sounds like an utterly stupid basis for a cryptocurrency, but I'm pretty sure it's just a way of describing a distributed Google Stadia service.

Mining has created an alternative use for a product that was previously used for running computer games. Right now that alternative use is being valued much more highly than being able to use a GPU to play games. It's no different than the price of milk skyrocketing if people were willing to pay three times as much for cheese. Solutions that try to rely on people not turning milk into cheese are just doomed to failure.

You really don't understand cryptocurrency if you think that's the case.

 

[Mopetar]

You really don't understand cryptocurrency if you think that's the case.

Proof of work only requires solving a problem that can be quickly verified without having to perform a similar amount of work to arrive at the solution. It doesn't need to be cryptographic in nature. The problem with something like BitCoin is that the work is useless outside of serving as busy work for the purpose of the currency itself. The work for Etherium is similarly pointless, only shifted so that it's memory bandwidth bound instead of compute bound so that it's harder to create an ASIC to mine it.

There's absolutely nothing from stopping someone from developing a currency where the work done has nothing to do with cryptographic hashes. Those are just convenient because they're the sort of problem that requires a lot of work to solve, but very little to verify.

 

[JoeRambo]

NVIDIA Unreal Engine 4 RTX & DLSS Demo Download

NVIDIA RTX Technology Showcase is an interactive demo built from NVIDIA's RTX Unreal Engine 4.26 Branch (NvRTX). This demo shows the benefits that ra

www.techpowerup.com

Some good looking stuff, getting to physical based rendering step by step. And performance with 3090 + DLSS is quite good too. Hard to pass 2x performance gains, for minor quality losses.

Nvidia seems to be working real hard in pushing their technologies to UE4 and Unity engines, to achieve parity it seems AMD will not only have to come up with their implementation of DLSS but also work on getting their stuff in game engines. For UE4 at least, some of this tech is "plugin" based, as easy as it gets.

 

[scineram]

Proof of work only requires solving a problem that can be quickly verified without having to perform a similar amount of work to arrive at the solution. It doesn't need to be cryptographic in nature. The problem with something like BitCoin is that the work is useless outside of serving as busy work for the purpose of the currency itself. The work for Etherium is similarly pointless, only shifted so that it's memory bandwidth bound instead of compute bound so that it's harder to create an ASIC to mine it.

There's absolutely nothing from stopping someone from developing a currency where the work done has nothing to do with cryptographic hashes. Those are just convenient because they're the sort of problem that requires a lot of work to solve, but very little to verify.

Someone needs to invent Primecoin.

 

[IntelUser2000]

You really don't understand cryptocurrency if you think that's the case.

Describing why he doesn't understand cryptocurrency would help. Otherwise you are saying nothing other than "I'm better than you".

I mean, there's Proof of Space that uses empty storage. And Proof of Network or whatever it's called that uses routers and being connected to other routers. Proof of Frames/Proof of Performance that uses performance in an actual game is viable too.

Of course transitioning majority of mining to first two would help with GPU demand.

Actually in the short term EIP-1559 would help. But as long as lockdowns exist, shortages will not stop.

 

[darkswordsman17]

Describing why he doesn't understand cryptocurrency would help. Otherwise you are saying nothing other than "I'm better than you".

I mean, there's Proof of Space that uses empty storage. And Proof of Network or whatever it's called that uses routers and being connected to other routers. Proof of Frames/Proof of Performance that uses performance in an actual game is viable too.

Of course transitioning majority of mining to first two would help with GPU demand.

Actually in the short term EIP-1559 would help. But as long as lockdowns exist, shortages will not stop.

No matter what, it will involve cryptographic algorithms, which inherently is not graphics rendering type of processing (GPU compute capability is why GPUs are popular for crypto work). GPUs are being bought up because they're still pretty good at doing that type of processing.

But the entire point of any of it is that it involves blockchain work, so it will invariably require other processing. There is no value in any of this stuff without the crypto blockchains.

Plus, Nvidia isn't interested in stopping crypto mining, they're trying to push the people wanting the compute capability for it to pony up for more expensive Nvidia stuff, so that Nvidia gets that extra money instead of retailers and resellers.

Those things are essentially giving value for locking in hardware capability, but again is all tied to the crypto. Storage and network makes sense, as would crypto processing capability. Not sure that graphics rendering does.

So, sure someone could (and I'm sure someone will if they haven't already) tie it to some sort of graphics rendering (arguably many NFTs are already that), but that is largely meaningless. What its tied to essentially does not matter, it is the blockchain keys (tied to crypto) that the value of any and all of this stuff is actually derived. Linking it to graphics rendering will still require cryptographic algorithm work. Nvidia could still flag anything doing such.

 

[Mopetar]

Cartographic algorithms aren't inherently required for proof of work, they're just simple to use, quick to verify, and often make it easy to adjust the difficulty of the problems being solved to show the proof of work. Some currencies, like Gridcoin are based around algorithms that can be used for scientific analysis where the work being done by the miners is useful because the problem it solves is potentially useful. The main issue is that it isn't always trivial or quick to verify solutions to some of these types of problems, whereas you can quickly tell if a particular input into a hash function produces the expected output.

Suppose you wanted to have a currency based around rendering frames as part of some Stadia-like service. There's an incentive to degrade the quality a little (or even a lot) in order to render more frames. Humans can subjectively tell if this has happened after a point, but verifying that the person doing the rendering isn't cheating the system would be a far more complex problem. The same goes for other things like a protein-folding problem. Some of those practically require a proof of stake to be used instead or in conjunction with that work because it's far easier to cheat than it is to catch someone at it, particularly if it's all decentralized.

Obviously anything involving transactions, etc. is still going to make use of cryptography, but it doesn't need to serve as the basis for determining who gets to mine a block. However, even with a blockchain there's a difference between using cryptographic algorithms to add to the blockchain and solving problems based on guessing the correct input to a cryptographic function to arrive at some output. One could just as easily create a blockchain that has no currency behind it at all. Of course that doesn't create a lot of incentive for anyone to participate in recording to it.

 

[jpiniero]

NVIDIA GeForce RTX 3080 Ti's GA102-225 GPU pictured and specifications leaked - VideoCardz.com

NVIDIA GeForce RTX 3080 Ti GPU pictured and tested in ETH Mining The RTX 3080 Ti is definitely coming. After numerous delays, NVIDIA has finally settled to launch its new enthusiast graphics card in May, multiple source claim. As we approach the launch of GeForce RTX 3080 Ti, the usual cycle of...

videocardz.com

Videocardz has details on the 3080 Ti - this QS (without the hash limiter) does 112 MH albeit with overclocked memory. Aparently the hash limiter was only added with the release version.

 

[Aikouka]

NVIDIA GeForce RTX 3080 Ti's GA102-225 GPU pictured and specifications leaked - VideoCardz.com

NVIDIA GeForce RTX 3080 Ti GPU pictured and tested in ETH Mining The RTX 3080 Ti is definitely coming. After numerous delays, NVIDIA has finally settled to launch its new enthusiast graphics card in May, multiple source claim. As we approach the launch of GeForce RTX 3080 Ti, the usual cycle of...

videocardz.com

Videocardz has details on the 3080 Ti - this QS (without the hash limiter) does 112 MH albeit with overclocked memory. Aparently the hash limiter was only added with the release version.

The rumored prices are $1000-1100, and it just seems like too much for right now. The problem is that the 3090 just isn’t fast enough to warrant the > +100% premium. (Not that I need to tell anyone here that.) The 3080 Ti ought to be close to the 3090 in performance, so would ~10-15% performance increase be worth about a 60% increase in price? It still seems like the 3080 is the sweet spot for higher-end gaming.

Although, I’m liking the 3070 for my HTPC since it has a decently lower TDP.

 

[MrTeal]

The rumored prices are $1000-1100, and it just seems like too much for right now. The problem is that the 3090 just isn’t fast enough to warrant the > +100% premium. (Not that I need to tell anyone here that.) The 3080 Ti ought to be close to the 3090 in performance, so would ~10-15% performance increase be worth about a 60% increase in price? It still seems like the 3080 is the sweet spot for higher-end gaming.

Although, I’m liking the 3070 for my HTPC since it has a decently lower TDP.

If the actual price you paid was close to $1100, it wouldn't be terrible by any stretch. It's not not great compared to 3080 FE pricing for the performance uplift, but it's great compared to what 3080s are actually selling for.

 

[psolord]

NVIDIA GeForce RTX 3080 Ti's GA102-225 GPU pictured and specifications leaked - VideoCardz.com

NVIDIA GeForce RTX 3080 Ti GPU pictured and tested in ETH Mining The RTX 3080 Ti is definitely coming. After numerous delays, NVIDIA has finally settled to launch its new enthusiast graphics card in May, multiple source claim. As we approach the launch of GeForce RTX 3080 Ti, the usual cycle of...

videocardz.com

Videocardz has details on the 3080 Ti - this QS (without the hash limiter) does 112 MH albeit with overclocked memory. Aparently the hash limiter was only added with the release version.

Do you know what the sad thing is?

GPU engineers have been straggling for decades, to make their designs smarter, more efficient, improve and enlarge caches, make the texturing units better, enhance the cores so they can do more and more, make the rops faster and do more per clock and here we are, having reached 2021, where for some people, the interesting part of a gpu, is how much MH a card can do.

This is beyond ridiculous and an insult to anything IT stands for.

I honestly don't know why the governments of the world are just standing there and don't ban the useless thing.

STOP BEING MONEY WHORES PEOPLE!

 

[Ajay]

NVIDIA GeForce RTX 3080 Ti's GA102-225 GPU pictured and specifications leaked - VideoCardz.com

NVIDIA GeForce RTX 3080 Ti GPU pictured and tested in ETH Mining The RTX 3080 Ti is definitely coming. After numerous delays, NVIDIA has finally settled to launch its new enthusiast graphics card in May, multiple source claim. As we approach the launch of GeForce RTX 3080 Ti, the usual cycle of...

videocardz.com

Videocardz has details on the 3080 Ti - this QS (without the hash limiter) does 112 MH albeit with overclocked memory. Apparently the hash limiter was only added with the release version.

Even if the finally product come with a signed driver with hash limiting, how long before these older pre-release drivers are 'accidentally' released so that mining can continue on at full rate. The chain is already broken.

 

[Dribble]

Even if the finally product come with a signed driver with hash limiting, how long before these older pre-release drivers are 'accidentally' released so that mining can continue on at full rate. The chain is already broken.

They don't need to be "accidentally" released, they are already out there. I would hope the release cards won't work with them however.

 

[biostud]

They don't need to be "accidentally" released, they are already out there. I would hope the release cards won't work with them however.

The 3080ti has another device ID, so they won't work with the older drivers.

 

[Krteq]

You can still simply modify .inf file with correct DevIDs etc., main binary/blob (and everything else) is the same as for other GA102 based cards

 

[Ajay]

You can still simply modify .inf file with correct DevIDs etc., main binary/blob (and everything else) is the same as for other GA102 based cards

Supposedly there is a new hardware (on the ASIC ?) that locks down the hash rate (well, for ethereum).

 

[Heartbreaker]

You can still simply modify .inf file with correct DevIDs etc., main binary/blob (and everything else) is the same as for other GA102 based cards

That remains to be seen, but I seriously doubt it.

 

[jpiniero]

That remains to be seen, but I seriously doubt it.

It's going to get cracked, it's just a matter of how long it takes.

 

[Mopetar]

Well if someone at Nvidia doesn't screw up and releases another driver that doesn't have limiter code in it we just might find out this time.

Considering that the 3090 is going for around $3k and this will have the same potential hash rate, it's market value to a miner is already $1.5k even with the limited hash rate, so it's already going to be botted and bought up by any miners who can get one at MSRP.

Presumably someone has already figured out how to crack it just from experimenting on a 3060. If they were successful, there's no reason for them to tip their hand to Nvidia or other miners right away because it puts them in a far better position to get the cards much cheaper than they otherwise might be able to.

 

[Leeea]

Well if someone at Nvidia doesn't screw up and releases another driver that doesn't have limiter code in it we just might find out this time.

Considering that the 3090 is going for around $3k and this will have the same potential hash rate, it's market value to a miner is already $1.5k even with the limited hash rate, so it's already going to be botted and bought up by any miners who can get one at MSRP.

Presumably someone has already figured out how to crack it just from experimenting on a 3060. If they were successful, there's no reason for them to tip their hand to Nvidia or other miners right away because it puts them in a far better position to get the cards much cheaper than they otherwise might be able to.

Thing is, there is a great deal of $$$ motivation to "accidently" release the driver. If that is the cheapest solution, every employee at nvidia will be targeted.


But I feel you are right, odds are the someone already hacked the 3060, and kept it in house waiting for the next round. Competitive advantage and all that.

 

[JoeRambo]

https://arxiv.org/pdf/2104.08378v1.pdf

Some good info about sparse ops for Tensor cores. While focus is A100 and academia ML, there is an quite some detail about sparsity matrix generation overall.

As DLSS 2.0 is basically "generic" neural algorithm, that takes as inputs: low resolution image + motion vectors + previuos frame high resolution image and outputs new high resolution frame, there seems to DL amateur like me a lot of potential for using sparse operations. At the cost of additional image quality, but potentially speeding up performance.

 

[JTsyo]

nVidia is reportedly stockpiling 3060s to flood the market in August to counter the 6600.

 

[gdansk]

I don't buy it. What company would decide to increase their inventory when they could immediately sell them instead...

 

[Hans Gruber]

I don't buy it. What company would decide to increase their inventory when they could immediately sell them instead...

I think he is implying that nvidia would flood the market with 3060 cards to counter the AMD 6600 when it is released in August.