Thursday 7 January 2016

Nvidia, raising standards

Recently looking into the changes from the current card I've got (Nvidia GTX 680) vs the newer ones for the hardware and architectural changes is interesting and where the next phases lead.

This public doc is good as a showcase for some of the main reasons Kepler-GK110-Architecture

and compares what the goals and changes that have been done from Fermi to Kepler. I quite like some of this including the Atomic operation improvements for one.

Gives some simplified points then in the next going from Kepler to Maxwell - Top 5 things to know about Maxwell

Overview - Maxwell

This is great because you can be a card now with almost 1,000 cores for under £100 without requiring separate power connectors either. So saves money on buying a PSU to support it as you had to before and the electric bill will be lower. Doing this on a large scale simply means many more people will find upgrading their older systems/desktops more of a viable option whilst in the HPC and other distributed computing projects can spend less £ and save running costs on a larger infrastructure and now combined with this Micron 8GB GDDR5 are working to produce greater amounts of memory that can be used with a graphics card, combine this with Nvidia improvements will also benefit those who love high end games will probably like to transition to 4K monitors. So much more processing required across many more pixels, we also have things like new video compression/decompression HEVC being introduced as well. Just need GPU to decode...

GeForce 1000 series  - the next gen, named "Pascal"

Also this NVLINK is great! - nvlink  what is NVlink?"5-12x higher bandwidth" etc

Excellent news for HPC, I'd like to see progress made with nuclear fusion reactors for defo. - exascale-supercomputing


The other point to mention is the clock speeds are lower in the newer architecture whilst still giving better performance. As the frequency of a clock is higher the amount of heat generated is higher too so therefore cooling in this case will be less noisy (if using a fan) as it won't be as hot. A CPU or GPU that runs at a higher clock rates require substantially more power. i.e. a 4GHz core will use much more than double power of a 2GHz core. (If assuming cores are of the same architecture). Can always look up "cpu clock speed vs power consumption" along with the charts, docs etc if anyone disagrees.

Forgot to mention that it is also helpful to programmers & developers.. and it is recommended you read all the way through this doc due to some key considerations - Is parallel programming hard?