NVIDIA 460.84 Linux Driver Released With GeForce RTX 3080 Ti Support

[phoronix]

Phoronix: NVIDIA 460.84 Linux Driver Released With GeForce RTX 3080 Ti Support

NVIDIA has now published the 460.84 Linux driver as the latest in their long-lived 460 driver series...

https://www.phoronix.com/scan.php?page=news_item&px=NVIDIA-460.84-Linux-Driver

[zexelon]

Yay! Finally more software we can download for cards we cant buy!

Maybe they should start charging for the software and giving the cards away for free? And you have to pay an extra 1,000 for drivers with mining built in... they should make sure its a completely different software stack so people could not patch enable it :P

[Myownfriend]

miners don't like RDNA2 cards because the infinity cache (128mb) is to small for the Ethereum working thread file what is 4GB...

That makes no sense. The infinity cache is a cache. It doesn't impose any limit on a data set in external memory. Plus you're acting like "compute" just means crypto-mining. It isn't. One of the CDNA's selling points is its FP64 performance which isn't useful for cryptomining.

because if you see the mm² of the chip the 3D part of the chip is so big that you can put the nearly the double amount of shaders on the chip if you remove the 3D graphic part.

What do you mean by "the 3D part"? The majority of any GPU is it's execution units (what you called shaders) and caches which are used for vertex and pixel shading. Sure there's some fixed function blocks like the geometry engine, texture units, and ROPs but they don't take up anywhere near as much space. You wouldn't fit anywhere near twice as many in the same amount of space

Vega64= 4096 shaders 500mm² 14nm
CDNA1= 8192 shaders 750mm² 7nm

What are those numbers supposed to show? That you can fit twice as many shaders of a different architecture in 50% more area with a smaller manufacturing process?

[Paradigm Shifter]

I know a "local" computer shop which since the farce of the 3000 series launch keeps a few cards which they will only sell in store, if you ask. However, they will charge their online price for them. I was going to duck out during lunch today and see if they had any in, but saw the price announcement last night - nearly 220,000JPY for a card that (at current exchange rates with USD) shouldn't be going for more than 135,000. I was prepared to overpay to around 165,000, but upwards of 200? Not a chance.

Looks like I have to hope my GTX1070 stays with me for another year. Funny thing is, that 1070 is already my longest-commissioned card - it's been going running daily for four years now. The only card to beat that was my Voodoo 5 5500, which saw daily service for about three years, then intermittent (twice a week or so, in an old box for playing old Glide games on Win 98) for a further two until it developed a weird issue where sometimes I'd get no video out after I installed drivers...

[Myownfriend]

of course it makes sense the 128mb cache is to small to have any usefull hit rate. other cards without infinity cache and higher memory bandwidth.

Hit-rate is a cache-related thing. The entire point of having a cache is to reduce the amount of times a process has to access off-chip memory. There is no GPU that I'm aware of that has more than 128MBs of cache so what are you comparing it to? Whatever the hit-rate may be, it would be much higher than other chips and its better for it.

your second point about FP64 yes there is no blockchain coin in FP64 YET... talk a littler bit more about this and they create one coin for you.

And there probably never will be because they'd be sacrificing mining speed and crypto are a get-rich quick scheme where they've managed to revolutionize the idea of stocks by adding insane amount of power usage. The sooner people realize how dumb crypto is, the better. These cards are made with the intention to do scientific research that will actually be of some good to humanity.

well these parts are big enough that AMD created CDNA and cut them all out.

Sure they're big enough that there's a decent savings by removing them but they're being removed because they wouldn't be used in the environments that deploy them, and that space would be better used for CUs. Does removing them mean they can fit twice as many CUs in the same space? No, not even remotely close.

[bridgman]

Ethereum miners don't like the infinity cache so they buy other cards like radeon7 or nvidia.
for other tasks without stupid limitations like ethereum the 128mb cache is super nice...

Can I suggest you tweak your default wording a bit and focus on video memory bandwidth rather than Infinity Cache ?

Miners do not have a problem with Infinity Cache itself, they have a problem with the lower video memory bandwidth which Infinity Cache makes possible.

Vega20 is great for mining because the HBM bandwidth is so high - if we went back and added Infinity Cache it would not get worse.