scrypt question, would faster and bigger ram improve hash rate?

[solareclipse64236]

if I bought 768 gigabytes of ram at 1600 mhz, would that improve hashing power?

[1715065041]

1715065041

[1715065041]

1715065041

[1715065041]

1715065041

[DeathAndTaxes]

if I bought 768 gigabytes of ram at 1600 mhz, would that improve hashing power?

No Scrypt parameters used by alt-coins use a negligible amount of RAM, about 128KB.  There is more than that in cache on the GPU (not main memory on graphics card, the registers and cache inside GPU itself).

[rumbitla]

if I bought 768 gigabytes of ram at 1600 mhz, would that improve hashing power?

768 gigabytes of ram is a lot of RAM.
You should buy it and give it a try.

[solareclipse64236]

So the RAM in the GPU is what matters?

[mufa23]

You'll lose more money in the extra electricity it requires to power the RAM than the hashspeed boost it would give you (if it even gave you an extra percent of hash speed). Reaper needs a lot of RAM, but CGminer doesn't need much. If you're running Cgminer (which you should) you only need a gig or two for each rig.

[FlungSpun]

oddly I have had better results clocking my RAM down to 1333 from 1600 with 2x 4GB sticks
has worked on three different rigs

[DeathAndTaxes]

So the RAM in the GPU is what matters?

Well RAM is still incorrect.  It is the cache inside the GPU itself.  RAM generally refers to the graphics card main memory (i.e. 1GB, 2GB, etc).  The memory usage for Scrypt is in KB.  Don't feel bad most people assume "memory hard" actually means "memory hard" but the memory hardness of the Scrypt used in LTC (and clones) is 1% of the default Scrypt requirements and those are considered for use in low security (real time) applications.  It is about 1/6,000th of the memory requirements recommended for high security applications.

[FlungSpun]

So the RAM in the GPU is what matters?

Well RAM is still incorrect.  It is the cache inside the GPU itself.  RAM generally refers to the graphics card main memory (i.e. 1GB, 2GB, etc).  The memory usage for Scrypt is in KB.  Don't feel bad most people assume "memory hard" actually means "memory hard" but the memory hardness of the Scrypt used in LTC (and clones) is 1% of the default Scrypt requirements and those are considered for use in low security (real time) applications.  It is about 1/6,000th of the memory requirements recommended for high security applications.

Interesting, I get the bits about scrypt 1% lite
So why does various monitoring software report GPU memory usage of between 1-1.5GB per card and CG miner using 80+ MB per instance?

[solareclipse64236]

we'll see

[rob143]

I found a rather interesting article on a possible use for a new developer board called the parallela the discussion can be found here : http://forums.parallella.org/viewtopic.php?f=8&t=276 .

[DeathAndTaxes]

So the RAM in the GPU is what matters?

Well RAM is still incorrect.  It is the cache inside the GPU itself.  RAM generally refers to the graphics card main memory (i.e. 1GB, 2GB, etc).  The memory usage for Scrypt is in KB.  Don't feel bad most people assume "memory hard" actually means "memory hard" but the memory hardness of the Scrypt used in LTC (and clones) is 1% of the default Scrypt requirements and those are considered for use in low security (real time) applications.  It is about 1/6,000th of the memory requirements recommended for high security applications.

Interesting, I get the bits about scrypt 1% lite
So why does various monitoring software report GPU memory usage of between 1-1.5GB per card and CG miner using 80+ MB per instance?

It has to do with the way cache works in GPUs.  Developers don't have access to the cache directly.  Scrypt kernel gets around this by loading the data into main memory to ensure it will be loaded into cache. The copies in main memory are not necessary.

As for cgminer uses system RAM. Well that is no different than any other program including your webrowser. 

[J35st3r]

I found a rather interesting article on a possible use for a new developer board called the parallela the discussion can be found here : http://forums.parallella.org/viewtopic.php?f=8&t=276 .

Some interesting analysis on that thread. Don't go selling your GPU's yet though as they are already optimised for hundreds (thousands) of parallel threads, while the 1024-core Epiphany chip is still just vaporware.