2 + 2 = 4, er, 4.1, no, 4.3... Nvidia's Titan V GPUs spit out 'wrong answers' in simulations

originally posted by: Maxatoria
After a good think I would say that we could be hitting a bottleneck in chip performance and that could be the actual chip(s), its memory or even the circuitry between the chips, normally this would be more of a set of chips overclocked to a point where errors start to creep in due to lack of heat displacement especially as its a premium product and thus at the max and running it somewhere in Texas may make it fail a lot easier than somewhere in Canada.

We've got another couple decades before we hit it, but it's going to happen. There's a minimum size on transistors, it's physically impossible for them to become smaller than an atom, we're also approaching a point where speed of light delays are affecting the chips.

Taking a multicore approach can still let us get more performance, and that's where the industry has been heading because the chips themselves can't become much faster, but threading has it's own performance limitations because not all calculations can be run in parallel, some have to be run in serial.

At this point, the speed of cpu calculations for everyday use doesn't have much to do with the hardware, it has more to do with how the software is written.

On that note, there is a massive, and I do mean massive lack of global talent that's capable of optimizing runtimes. Unfortunately, it's not an easy thing to teach because doing it properly requires an extremely deep knowledge of how the chipset operates as well as selling your soul for superhuman coding prowess.

a reply to: Aazadan

Shrinking cpu size has pretty much reached EOL, multi core(s) allows for better performance but only if the software allows it and working out code to use such cores has been a challenge for decades and i've worked on mainframes to try and balance workloads.

Most of the software written is very much if X then do Y and we've seen the problem recently with pre branch execution as well but for this sort of event where its giving wrong results i'd say it probably needs either under clocking from its current values or better cooling or perhaps a temperature sensor to sort it out.

This event is probably someone overclocking it to the max for PR purposes in a nice environment and once its let out in the wild to find out its a little too sporty for just sitting in someones office.

it's not a bug it's a feature


this way it is more compatible with Microsoft
which also only functions 10% of the time

a reply to: njord

Probably these sort of systems will be more used with Linux and from memory of doing comp sci and having to do bits of physical design of stuff the main things would be

1. They stuffed up the basic board design such as signal timings not quite hitting at the right timing intervals due to small errors in resistance of the tracks on the board meaning that delays can happen and the faster the board the less margin of error.

2. Thermal problems, those can cause similar problems to No.1 as tracks at different temp's can respond differently and thats before we head into the next phase.

3. Crappy cooling as a problem is a good place to start, chips have thermal limits as well and pushing them beyond that can create random instability so all the solution may be is a few beefier heatsinks and a new slab of thermal paste and perhaps make sure the machines not in direct texan sunlight.

But after many years of messing around with computers the moment you take anything to even near the limit you need to look at heat as a matter of course, drop the cards frequencies/voltage and see if it works and if so then the manufacturer has probably made a balls up on the specs.