Summer of GPU Love incoming.
What Groetsl has that x11 does not though is some massive NVIDIA lurve too (3x performance incoming, see below). Nasty!
I expect equally impressive AMD improvements and further bumps to CPU ops.
Also Groestl could scream on ARM: we'll have to wait and see. I cant see x11 getting any such optimizations cause it has taken the mish-mash science algo approach.
Go Groestl!
I'm doing some temp testing with 2x290s and I don't believe what is going on.
The temp outside is getting close to summer temps. The 2 r9 290s are in a garage. With scrypt, even during the winter, I needed to have some cooling fans blowing air through the garage.
With GRS, with temps that are approaching summer, I need to keep the garage ventilation closed. This brings the cards up to good operating temps around 75c. I have a couple of small mobo powered fans blowing air into the GPUs path, but that is about it.
wow. Along with x11, GRS is very impressive. Why are people still using scrypt?
1. Are the hardware issues experienced with the random access from many threads something like "parallelisztion resistance"? Or in other words would that make this algo offer low(er) ASIC advantage compared to other algos in use? Would producing ASIC hardware therefor be more difficult and result in a lower / reduced ASIC advantage compared with SHA-256?
2. So NVIDIA will run 3x current speed with your "different approach"!?
3. Can you imagine an efficient ARM miner for this algo (a BOUNTY has been offered)
4) Any other info bout what makes this algo different (quirks, surprises etc) from others used in crypto mining would be enlightening
thank you Christian for your generous contribution to this conversation
1) the implementer of the SPH library chose a table based approach because it works fast on CPUs. This doesn't work as fast on GPUs though. On GPUs explicit computations are usually faster than table lookups. It's just this particular approach taken that is somewhat resistant to parallelization.
2) yes.
3) maybe on 64 bit ARM platforms with 4 and more cores, using wide 128 bit SIMD operations (NEON?). I am not very familiar with ARM, so I can only speculate. In fact, our alternative approach would also apply here (and to Intel SSE2/AVX2)
4) it's very similar to AES, but somewhat upsized. The algorithm uses interesting maths based on Galois Fields GF(2^8).
and thanks so much for your considered response.
just as a follow up to 1)… Do Bitcoin ASIC's use a table-based approach or specific explicit computations? I guess it's the latter. And can you speculate on how a Groestl ASIC might vary from a SHA-256 ASIC and the possibility that Groestl might offer a reduced ASIC advantage compared to SHA-256?
So…
ARM mining has potential (speculation appreciated) Lets get that ARM BOUNTY up!!
and…
your
3x NVIDIA performance news (that's insane!) is
HUGE for Groestl!
When can we expect to see this?So many many thanks
No more questions after this I promise
But this will be for ccminer I believe. AMD cards will get optimzations too I expect but I dont have a time-frame (I'll do some digging).
