Ok that sounds good, so it would even make sense to buy a i7 6700k, 16 GB RAM and then add lets say 2 graphic cards to have a CPU/GPU fully used machine? The GPU mining software would run on Linux or Windows?
By the way I also would throw in 1 BTC for this.
I am afraid all the people with 4 or less GPUs paying for the miner could be shooting themselves in the foot. If medium-size GPU miners have access to this, small GPU miners and CPUs might need to say good bye to getting any blocks. They might be better off with 1 CPU than with 4 GPUs, if this discussion is the difference between an open miner and a closed one. We may be a lot better off if Wolf0 takes the most profitable route like Ampy and sells to 1 buyer.
Despite the current dearth of willing talent, that situation won’t last long.
When we have pools up and running (won’t be long now) then the payout to miners (of any size) will be proportional to the number GPUs and CPUs they have (with GPU’s being weighted about 4.5 times a CPU in the most extreme case).
IMO an open source GPU miner is both an inevitability and also not a threat to CPU mining.
Remember, a GPU with 4,000 cores can use only a handful of those cores (4 or 5) to mine, because the bottleneck is memory bandwidth.
Also recall that GPUs cost more (both in electricity and initial purchase price) than CPUs.
It is not even clear at this point that a GPU miner will provide any cost-per-coin benefit at all – let alone one that is significant enough to make GPU mining a threat to CPU miners.
EDIT: Just to expand on this point a bit: I know someone with almost two decades experience in programming for FPGAs, GPUs, clusters, etc. It’s what he does for a living.
The only reason I haven’t asked him to help yet is because it feels (to me) borderline unethical to sell people a GPU miner for ~1 BTC each that would provide such a small bump in mining power even if we achieved the maximum theoretical improvement possible. Indeed, with electricity costs taken into consideration, your cost per coin may even be worse than if you mined on a CPU.
I know that “a fool and his money are soon parted” and all that, but I’m not sure it’s ethical to participated in that parting – so I try to avoid doing it.
@Austin-Williams - thanks for all the quality information.
So it ‘seems’ a reasonable approach to long term EV+ mining would be multi-core CPU where each core has access to at least 1 GB or RAM?
Say a 8 core CPU, with 16 threads and 16GB+ of RAM would be able to mine fairly efficiently even vs. GPUs
Same question here. How efficient would be CPUs with more then 4 cores? Or is the gain between 4 core with 8 GB RAM to 8 core with 16 GB RAM not nearly linear?
That’s exactly what I’m after. In theory it should be almost 100% linear since data transfer from CPU to RAM, and back scales pretty much linear independently of the number of cores/threads.
It would also be great if that was the case, since it should mean that multi core high end desktops and severs hold an advantage over botnets. At least somewhat.
And even if it doesn’t keep pace with GPU, it would be reasonably viable and not a complete dud since even the most beefy GPUs are capped memory size-wise at around 8-12 solving threads.
As far as I understand it till now multi-thread mining is not enabled or possible at the moment. So at least it is not possible for all the cores to work on one package of data if I understand that correctly? Is this of significant relevance? Perhaps somebody can give some more detailed answers to these questions.
@Prometheus with genproclimit=-1 the miner will use all available cores. Only the benchmark is single threaded. The confusion is that technically, the miner is not multi-thread per se, but the program launches X instances of it in parallel, one for each core. It’s transparent for the user, so for me it’s multi-thread-like 
(but yeah, different set of data on each core)
@DanLar75 about the multi-core advantage, it’s unclear for the moment. Yes, a second core with another Gb of memory can run another miner. The main bottleneck, however, is the memory bandwidth, so if the memory bandwidth is not saturated by one miner, the second one will take full advantage of the multi-core architecture. Otherwise they will compete for an already maxed out capacity, and might perform even worse than a single instance.
More testing is what is needed. The second beta will enable easy multi-core testing (zbenchmark).
For the moment, the best way is to launch two miners, one single and one multicore, on two different machines with the same setup, and count the coins after a few days
(the test cannot occur sequentially because the difficulty is not the same)
@adrian I have a 16 (32 thread) core dual Xeon HP820z with 128 GB RAM I will benchmark on. I might DM you for advice if you don’t mind. I happy to share benchmark data etc.
This is exactly right.
Increasing the number of cores and the amount of RAM on the same motherboard will NOT increase your mining speed.
In practical terms: what you really want to increase is the number of memory busses you have.
To be even more specific (for the hardware buyers in this thread): what you really want to increase is the the number of motherboards you have.
A screaming fast single core CPU (or dual core, that way one core can be preoccupied with non-mining tasks), with > 1 GB of GDDR4 memory, and a motherboard with a fast memory bus. That’s your ideal off-the-shelf mining rig. Build lots of those.
… but you have only ONE memory bus.
If you swap out the 128GB RAM with 2GB RAM (of the same generation/technology) You shouldn’t see any decrease in mining speed.
That CPU is nice though!
Re: PC motherboard(s)… How much bandwidth would you get spending the same amount on Raspberry Pi 3s? Yes, they use DDR2 but the question remains…
That’s a great question. It would be worth looking into.
STEP 1: Fire up a RPi3, get it mining, and do a benchmark.
I suspect that in the RPi3 case, it may well be the speed of the CPU and slow DDR2 that are the bottleneck. But it’s worth looking into to verify. It would be pretty cool if a RPi cluster were an effective way to mine ZEC.
All this low end hardware would work better with a multi-core solver…
@Austin-Williams you were talking about multi-motherboard, this is a nice concept
Does anyone knows of anything that exists, close to this concept?
Do the multi-processor board share the memory between the processors?
Is this possible to imagine an architecture with tens or hundreds of CPU+mem SOCs on the same motherboard, on a single OS? 
Yes they do. And this is exactly why mining with more cores and more RAM does not in anyway increase your mining speed.
You need more memory bandwidth. And unless one plans on designing and manufacturing their own computing equipment, the only way to achieve this is with more motherboards.
Tl;dr: if you want to mine zec faster, you need more entire computers. Not faster single computers.
Grid computers. Like ol’days.
I’ve been trying zcash mining on several machines. I have noticed a small improvement in using a recently built desktop with DDR4 RAM against an older laptop with DDR3 RAM, but the advantage was not huge. This is a lot closer to the original Satoshi vision of “one CPU, one vote” than even I had hoped.
I think memory bus is less of a bottleneck than thread-ram. And since each mining thread takes 768 MB I need to match 32 threads to ram ar least (I think).