NVIDIA releases CMP lineup and reduces hashing rates on GeForce cards

[porina]

35 minutes ago, leadeater said:

No like SSL, it's the same architecture and uses PKI.

I was thinking of HDCP as an example of a similar system where manufacturers are trusted but can have keys revoked if they should fail to meet their requirements. 

 

Anyway, what nvidia has decided to do is in motion, and we can only see how it plays out.

[PlayerLorenzo]

This will just cause a drift to AMD, AKA higher prices for AMD cards due to Nvidia doing this stuff. It's like putting a restrictor plate on a Lamborghini or Bugatti to stop people from racing with those cars. AKA you're limiting the real power of a GPU just to get more money from the CMP GPUs, even when there is also a silicon shortage that affects the prices too.

[Kisai]

23 hours ago, porina said:

I was thinking of HDCP as an example of a similar system where manufacturers are trusted but can have keys revoked if they should fail to meet their requirements. 

 

Anyway, what nvidia has decided to do is in motion, and we can only see how it plays out.

HDCP is a DRM scheme, and all revoking keys does is result in planned obsolescence. There are also known work-arounds to HDCP that do not involve decrypting the keys or forging them.

 

In respect to how nVidia proposes to impair ETH loads, I've never actually taken a look at ETH source code, but my guess if "memory load" is the operative words, then it's probably why this is on the Ethereum page:

Quote

Why is my hashrate with Nvidia cards on Windows 10 so low?

The new WDDM 2.x driver on Windows 10 uses a different way of addressing the GPU. This is good for a lot of things, but not for ETH mining.

• For Kepler GPUs: I actually don't know. Please let me know what works best for good old Kepler.
• For Maxwell 1 GPUs: Unfortunately the issue is a bit more serious on the GTX750Ti, already causing suboptimal performance on Win7 and Linux. Apparently about 4MH/s can still be reached on Linux, which, depending on ETH price, could still be profitable, considering the relatively low power draw.
• For Maxwell 2 GPUs: There is a way of mining ETH at Win7/8/Linux speeds on Win10, by downgrading the GPU driver to a Win7 one (350.12 recommended) and using a build that was created using CUDA 6.5.
• For Pascal GPUs: You have to use the latest WDDM 2.1 compatible drivers in combination with Windows 10 Anniversary edition in order to get the full potential of your Pascal GPU.

So perhaps by making a different driver that addresses the GPU in a way that is favorable to ETH, is how they plan to make these CMP cards viable. 

 

Also there's notes on the page about needing >4GB of VRAM to operate or the hash rate is poor as well.

 

My impression here is that they might actually put some kind of nerf into CUDA 11 itself (Which is what Ampere cards use) where a 4GB load using straight CUDA puts it into this nerf mode.

 

https://github.com/ethereum-mining/ethminer/issues/1929

Quote

Once the DAG updated to Epoch 288, nvidia 1050ti 4GB models now fail with mem loading errors of the dag.

I'm sure other mining systems have 4GB+ data sets, as well as many neural net AI's, though in the latter, additional libraries are provided by nVidia (eg cuDNN, Tensorflow, etc.) So that can't be the sole detection mechanism.

 

 

[Chipss36]

I am glad to see this...

the miners are sticking it to all us..

nice to seem them, “get some”.

 

 

[jagdtigger]

14 minutes ago, Chipss36 said:

I am glad to see this...

the miners are sticking it to all us..

nice to seem them, “get some”.

 

 

Yup, i hope all the 3xxx cards will get the nerfing tweak.

[Orthusaku]

8 hours ago, jagdtigger said:

Yup, i hope all the 3xxx cards will get the nerfing tweak.

Why not have two seperate Skus?

 

SKU One with the mining nerf and 3/4th the MSRP

SKU Two with no nerf and the standard MSRP.

[Moonzy]

50 minutes ago, Orthusaku said:

Why not have two seperate Skus?

 

SKU One with the mining nerf and 3/4th the MSRP

SKU Two with no nerf and the standard MSRP.

hypothetically, you have GPU A, and GPU B

GPU A: $300, 30MH/s, so 100kH/$

GPU B: $400, 60MH/s, so 150kH/$

 

miners: will finish buying GPU B to the point that GPU B price becomes $600, meaning it's 100kH/s, and possibly still looking for more GPUs.

 

u see the issue with this?

 

and besides, it doesn't benefit nvidia, the mining cards will end up on 2nd hand market, hurting their future sales

they want miners to use a card that gamers cant use in the future, so they can make more money

[jagdtigger]

23 minutes ago, Orthusaku said:

Why not have two seperate Skus?

Becausr its more beneficial to kill mining in the long run.

[Orthusaku]

2 minutes ago, jagdtigger said:

Becausr its more beneficial to kill mining on the long run.

but it is just math

 

  

19 minutes ago, Moonzy said:

hypothetically, you have GPU A, and GPU B

GPU A: $300, 30MH/s, so 100kH/$

GPU B: $400, 60MH/s, so 150kH/$

 

miners: will finish buying GPU B to the point that GPU B price becomes $600, meaning it's 100kH/s, and possibly still looking for more GPUs.

 

u see the issue with this?

 

and besides, it doesn't benefit nvidia, the mining cards will end up on 2nd hand market, hurting their future sales

they want miners to use a card that gamers cane use in the future, so they can make more money

 

I really wasn't thinking about what benefits nvidia, I was more concerned about not having cards getting nerfed each time a new workload gains traction. I'm still upset RTX 3000 series doesn't have SR-IOV, and VM monitoring is still used.

[Moonzy]

4 minutes ago, Orthusaku said:

I was more concerned about not having cards getting nerfed each time a new workload gains traction

Then you should be against it, not accepting compromises

Because having a company dictate what we can and can't do with our hardware isn't the future I want

[Blademaster91]

9 hours ago, jagdtigger said:

Yup, i hope all the 3xxx cards will get the nerfing tweak.

Yeah i've seen rumors of the 3070 and 3080 getting discontinued for a new version with the mining nerf in the vBIOS,but it doesn't make much sense because nvidia is likely making record profits on selling 3070's and 3080's directly to miners.

Unless Nvidia is going to replace the 3070 and 3080 with a super version or a card that has more vram, which would be nice because the 3080 should have more than 10GB imo.

[Rym]

Logically, Nvidia would make more money by discontinuing current 30 series cards and implementing hard nerfed 3080s and 3070s that cannot mine efficiently at all.

 

There is a heavy stock limit atm, which is partly because of shortages, then because of covid, then because of artificial price gauging. Regardless, it's extremely simple to determine that hard nerfing 3080s and 3070s in mining is a good thing for profit.

 

Failed SKUs that would never exist as either 3080 or 3070 would now be able to exist as CMP, functional especially for mining and fulfilling it's intended purpose. This increases profit by default.

 

RTX video cards can now be sold to gamers due to a higher availability, prices will decrease because gamers won't pay 200 to 300% over MSRP, but the normal prices of these GPUs plus the addition of failed SKUs under the form of CMP should in theory provide an increased profit overall when compared to selling all your stock overpriced to miners.

 

 

Of course, the most profitable way possible is to simply sell both CMPs and functional SKUs to miners for 200%+ above MSRP, while claiming shortages. This depends on competition and other factors however and if Nvidia considers whether their GPUs should from now on be focused solely on mining for profit which is volatile at best or on gamers which have kept them in business for decades.

[WkdPaul]

1 hour ago, Rym said:

Logically, Nvidia would make more money by discontinuing current 30 series cards and implementing hard nerfed 3080s and 3070s that cannot mine efficiently at all.

no, they wouldn't

 

nvidia makes money regardless who buys the GPU, miners or gamers.

 

Also, they've only nerfed the Dagger-Hashimotto algorithm for ETH ... that's a VERY narrow nerf that will be bypassed at one point or another. Meaning, miners will still continue to buy GPUs (nerfed or not), and CMP won't be sold since they don't have any second life as a regular GPU, and as such, no resale value.

 

BTW, even if the nerf can't be bypassed and miiners decide to mine ETH on the Dagger-Hashimotto algorithm @ 50%, they'll still have something with a great resale value at the end compared to a CMP.

[eece_ret]

Of note on the CMP cards being a "throw away card" 

https://ethereum.org/en/eth2/staking/

When ETH goes Proof of Stake, mining essentially is game over for Eth.  Doesn't this diminish the shelf life of these CMP products even more?

 

Due to profit motive, this issue will return over and over.  NVIDIA sells chips, they will follow demand, right now, that demand is with Crypto.

 

The only way to change this situation would be for a device that can cover various hash algorithms at a better MegaHash/Watt rate and was produced on a noncompeting process/lithography.

 

RISC-V could be an interesting option in this space due to its Royalty Free IP offering.   

 

 

 

[Moonzy]

7 minutes ago, eece_ret said:

The only way to change this situation would be for a device that can cover various hash algorithms at a better MegaHash/Watt rate and was produced on a noncompeting process/lithography

they arent gonna spend millions/billions in RnD for this

and there are ASICs already

[Kisai]

On 2/24/2021 at 4:24 PM, eece_ret said:

 

 

The only way to change this situation would be for a device that can cover various hash algorithms at a better MegaHash/Watt rate and was produced on a noncompeting process/lithography.

 

That is what a GPU is. It’s a very parallel math processor.

 

The alternative is to build FPGA-based miners, and unfortunately FPGA’s are substantially more expensive and not parallel in this regard. There isn’t a factory putting out cheap FPGA’s at all. So regardless if you could buy one or one thousand of them, you would program the ASIC for the algorithm needed and they would still become useless when GPU’s run circles around it.

[Strykker]

I think this article nails the why of the CMP series Nvidia cards.

 https://wccftech.com/nvidia-cmp-series-anti-consumerists-intentions-looks-to-be-to-kill-second-hand-resale-market/

[Kisai]

1 hour ago, Strykker said:

I think this article nails the why of the CMP series Nvidia cards.

 https://wccftech.com/nvidia-cmp-series-anti-consumerists-intentions-looks-to-be-to-kill-second-hand-resale-market/

"nVidia wants to kill the second hand market" ... I doubt that's the case. That may be the consequence for those who buy the cards for mining, but I'm sure those mining cards could be used for cheap ML applications that don't necessarily require the entire GPU for inference, let alone training. Just I don't see Google buying these cards on the second hand market.

 

Since nobody has one to actually try it we won't know, and it's likely people doing ML stuff at home want GPU's that work for everything, not CMP cards that have reduced utility.

[igormp]

1 hour ago, Kisai said:

but I'm sure those mining cards could be used for cheap ML applications that don't necessarily require the entire GPU for inference, let alone training.

They have less vram than their regular consumer counterpart tho. But yes, if they're cheap enough, buy 2~4 of those would be nice.

 

1 hour ago, Kisai said:

and it's likely people doing ML stuff at home want GPU's that work for everything, not CMP cards that have reduced utility.

I could keep going with my 2060 Super or even a 1050ti I have laying around just for display, and use those CMPs for actual computation though, that's a non-issue.

[Kisai]

1 minute ago, igormp said:

They have less vram than their regular consumer counterpart tho. But yes, if they're cheap enough, buy 2~4 of those would be nice.

 

I could keep going with my 2060 Super or even a 1050ti I have laying around just for display, and use those CMPs for actual computation though, that's a non-issue.

I guess we will have to wait and see if the CMP cards end up on ebay at $50 a piece when this mining boom crashes.

[Mark Kaine]

You know, "and reduces hashing rates on geforce cards" kinda implies they're doing this retroactively... which has several implications - but I'm not sure it's actually meant like that... 

[Kisai]

2 hours ago, Mark Kaine said:

You know, "and reduces hashing rates on geforce cards" kinda implies they're doing this retroactively... which has several implications - but I'm not sure it's actually meant like that... 

They could always push a vBios update and cripple all the Geforce cards hashing rates, and for other cryptocoins. I doubt it, because miners will just not update the drivers, and thus all the algorithms get stuck at the GPGPU API used (eg CUDA 11) 

 

Honestly, if nVidia wanted to be stealthy about it, they would incrementally update the vbios over several driver updates, and then pull what DirectTV did to kill the pirate card market.

https://slashdot.org/story/01/01/25/1343218/directvs-secret-war-on-hackers

 

In a sense, update the vbios on the gpu's and not allow the drivers to install on older/unsigned/mis-matched vbios versions. People can hack the drivers, but to hack both the drivers and the vbios is a lot more of a pain in the behind. The bios has been encrypted since Turing.

 

For all we know nvidia could retroactively cripple all windows and linux drivers from mining by having the drivers for Geforce and Quadro drivers require 8x/16x links to initialize, and not initialize on 1x links, and reversing that for CMP.

 

 

[Moonzy]

@leadeatercould the old dedicated mining cards do F@H?

 

edit: short googling says yes, but the PPD seems lower...?

[Craftyawesome]

54 minutes ago, Kisai said:

They could always push a vBios update and cripple all the Geforce cards hashing rates, and for other cryptocoins. I doubt it, because miners will just not update the drivers, and thus all the algorithms get stuck at the GPGPU API used (eg CUDA 11) 

 

Honestly, if nVidia wanted to be stealthy about it, they would incrementally update the vbios over several driver updates, and then pull what DirectTV did to kill the pirate card market.

https://slashdot.org/story/01/01/25/1343218/directvs-secret-war-on-hackers

 

In a sense, update the vbios on the gpu's and not allow the drivers to install on older/unsigned/mis-matched vbios versions. People can hack the drivers, but to hack both the drivers and the vbios is a lot more of a pain in the behind. The bios has been encrypted since Turing.

 

For all we know nvidia could retroactively cripple all windows and linux drivers from mining by having the drivers for Geforce and Quadro drivers require 8x/16x links to initialize, and not initialize on 1x links, and reversing that for CMP.

 

 

Does nvidia update the vBIOS with their driver installer? Considering how many partner cards there are, that sounds like a logistical nightmare.

 

I'm not sure why incrementally updating would be an option here. My understanding is that directtv's strategy only works because the updates are mandatory to be used.

 

They can make it a pain for any miner that gets a card with an updated vBIOS, but I don't think they can stop one with an spi programmer and downloads of an older vBIOS. Unless the card already has fuses to burn to prevent downgrades or something.

[leadeater]

1 hour ago, Moonzy said:

@leadeatercould the old dedicated mining cards do F@H?

 

edit: short googling says yes, but the PPD seems lower...?

You mean the old/original mining GPUs or just old mining GPUs in general which includes these CMP cards? Technically any can so long as CUDA drivers are implemented correctly for the card.