Topic: Mining performance of Nvidia RTX 2080 Series - page 11. (Read 9684 times)

Utility of the new features depends on the algorithm, for example:

Tensority algorithm - Uses quad 64-iteration 256x256 matrix multiplications, input matrices are composed of 8-bit quantities, results are compressed from 16-bit back to 8-bits prior to the next matrix multiply operation. This allows use of:

1) DP4A 8-bit vector instructions (present in Pascal GP102/104/106, present in Volta V100, unknown currently for Turing - still waiting for CUDA 10 SDK to be available to verify this feature).

2) Tensor core FP16 multiply with FP32 accumulate. FP16 allows for 8-bit integer input values to be stored with no loss in precision, FP32 accumulate allows for 16-bit integer result from 8-bit multiply to be stored with no loss in precision.

3) Tensor core INT8 mode (new in Turing, absent from Volta). If accumulate is in INT16 (likely as this is a matrix multiply operation with 8-bit inputs, should be verifiable once CUDA 10 SDK is available) this would then allow for use of this mode in this algorithm. Throughput would be twice that of FP16 multiply with FP32 accumulate from #2.

4) Parallel-issue FP32+INT32 CUDA cores (introduced with Volta, also in Turing). Similar to #2, FP32 execution path in cores will allow for 8-bit integer multiply and 16-bit integer accumulate.

Groestl algorithm - (Note: I have not coded a CUDA kernel for this, so I am just speculating based on the specs for this algorithm). Uses matrix multiplication in it's "MixBytes" transformation step, input matrices are 8-bit values, hence the same 4 CUDA features discussed above are possibly applicable to this algorithm as well).

As the tensor cores are FLOATING POINT, while mining is an INTEGER function, as I mentioned back when the new Tesla was introduced I don't see them being a factor AT ALL for miners.

Interesting theory, but I'm still making enough to live on from mostly GPU mining.
Profitability IS down to where only folks with very cheap electric are making any money though.

With that said, I don't see the announced RTX 2070, 2080, or 2080 ti being good mining cards vs the 10xx series at their VERY HIGH PRICE, especially with the older cards starting to drop to pricing at or below where they were at 2 years ago before the big cryptocoin price runup happened - and the move of Bitmain, Innosilicon and others to put ASIC into most of what used to be the more popular GPU mined coins hasn't helped that any.

It will be interesting to see how these newer cards perform on Folding (which CAN be mined indirectly), but I suspect they're not going to be price/performance competitive even there.

there could be a surprises with RTX series if mining software devs find a way to utilize the tensor cores, if not the performance bump wont be that impressive considering the price point.   

https://devblogs.nvidia.com/programming-tensor-cores-cuda-9/

absolutely right
or series 10

Yes and no. RTX 2080 series will likely see very strong resale value due to its freshness. I think you can sell your 2nd hand card at near MSRP if you get it first.

Now that RTX 2080 series have hit the market go and just do a simple check on Ebay for the Rx Series, the 570/580 all cost less than 200 dollars and we can build a miner with 1000 dollars or even less if we can wait. That would be a 6 card miner which would give us 180 mhs stable compared to single RTX 2080 ti it is clear who is the winner here.

Not worth to use RTX 2080 series for mining when with that price of a RTX 2080 ti can build a big 6 cards miner now.

It would be interesting to know the performance of the 20*0 series in other algorithms (verge-scrypt/blake, equihash, CryptoNight and forks).

Problem with Nicehash is that it really affects small-cap coins where there isn't a whole lot of hash power. The difficulty adjustment algorithm makes it hugely unprofitable for existing miners when Nicehash switches TH/s worth of power in and out of the network. 

Nicehash is a huge problem yeah, but if nicehash was down for a while it would just be picked up by all the other profit switching pools

Nonsense. Bring down nicehash and you`ll see how profitable everything is again to mine. Nicehash is the doom of home miners, asics not so much.

considering that mining is UNPROFITABLE RIGHT NOW, i don't see the point of spending a huge amount of money to buy a gpu to mine ETH (because there is no other coin that has not been invaded by ASICS and has been made unprofitable for GPU's) specially when Difficulty bomb is already on the plans of the DEVS as stated in the last stream of 24 Agost on youtube.


The block reward is also going to be made, EIP-1234 would reduce block rewards to 2 ETH and EIP-858 would reduce block rewards to 1 ETH per block.

i'm going to say one important thing: unless ETH goes x2 or x3 in the next month, you´ll not even get close to ROI when difficulty bomb is applied.


you have to be really mad, like really really mad to buy GPU's for mining right now.

what about other algo i'm interested, but at that price is a stupid buy

What an absolute scam site.  Why would they be selling the FE card at the same price as Nvidia?  There aren't even drivers available for the 20 series yet.  News sites can't post benchmarks until mid-september, yet this site already knows the mining performance?!

This product for games is not specifically for miners. if you are looking to mine, please use ASIC or FPGA.

Yeah at the current price point, the 2080ti is definitely a NO-GO. This is for both gaming or mining. Like the quoted post says, you'd better off buying 2 cheap 1080tis then linking then with SLI (if you are into gaming). Ray tracing RTX and GDDR6 have yet to have software built for them do so this may change in the future, but at least for now, the 2080tis are complete shit and are just a marketing stunt by NVIDIA.

No matter how beastly the 2080ti is, it will not beat 2x 2070 or 2x 1080ti.

I don't see how 2080ti will be price competitive for mining. The only exception is rendertoken performance where 11gb may provide advantage over 8gb, and raytracing cores > 1080ti.

I think ultimately we need to see how good is the GDDR6 at overclocking. Afterall, it is new technology that is still largely unproven. If GDDR6 turns out to love being overclocked, then RTX 2080Ti will be a BEAST.

Given that this model (RTX 2080Ti FE) costs $1,225.00 , the increase in hashing rate from 55 to 69 MH/s is not enough to justify the price. The other question important for mining application is how much one can overclock this card.

Looks like they probably used the eth-pill rate of a 1080ti at ~55 MH/s, then multiplied by 1.27 to factor in the +27% memory bandwidth gain of GDDR6 in the 2080ti, to get the ~69 MH/s projected value (by comparison, the theoretical zero-latency rate of 616 GB/s bandwidth would be ~75 MH/s, so their estimate seems reasonable assuming "mining-friendly" latencies).