Post
Topic
Board Mining (Altcoins)
Re: AMD Vega frontier released!
by
AKRO
on 01/07/2017, 17:17:33 UTC
I'm not an expert so take this with a grain of salt, do oyur own reading and analysis

I remember when RX 480s came out, and they were getting ~19Mh/s on an 8gb card, then people understood the card's specs and now they hash at 29.
I don't think there is any single miner that is optimized for Vega yet, so it's using generic kernel. And it all depends on the algorithm and how it uses the card's resources.

The r9 380 4gb can hit like 25mh/s maybe even higher with expert modding and cooling (29 is possible but highly unstable), whereas stock it's like 18.5mh/s and with the case of increasing DAG size for ETH, it shouldn't hit too many snags, since it technically has 384 bit bus, but only 256 accessible(physically), which should indicate that memory access was intended to be optimized for a higher bus speed but Radeon didn't want it being better than the 290/390 price to performance in gaming.
In my rough bro-science opinion, it's effectively the same bottleneck on the 470 and to a lesser extent the 480 when the Dag jumps again in the coming months, those cards should see significant performance drops closer to their release date numbers, because they're rocking 256-bit buses and they were made for it.

Vega is a pretty much Fiji meets 14nm, with some improvements. Higher core clock by 60%, faster memory clock by 100%, but half the bus width. I reckon we would have to hit about 50mh/s+ before exhausting the 512 bit memory bus, so unless Vega can hit over 100mh/s (which i doubt) it's going to suffer from poor optimization, and once that's cleared up, it should fetch low 50s at a conservative estimate based on what it's currently doing now to maybe low 70s based on specs, and even higher in pure theoretics regarding how the memory is utilized which might be a hard NO. it should be very resilient to DAG changes, and the gaming cards will be 400-500 USD, 225W, single slot, and according to AMD, even faster, better etc. not a horrible deal. I feel like it should do pretty well in memory hard algorithms where bus or bandwidth are key, like Zcash, Zero(N=192, K=7 ) or Zcoin

You can think of GPU specs as a shipping warehouse, as far as i can tell anyway.
Core: how fast the employees are fetching and packing shipping pallets
Memory speed: how fast other employees are relaying orders to the core and loading the containers into trucks
Memory Bus: how many truck loading docks you have
Memory Bandwidth: how fast you are able to load the trucks.
Hashrate: how fast you are shipping.