Some Of Nvidia's Chips Are Getting The Incorrect Response For Math Problems (Nvda)

This doesn't audio equally earthshaking equally the Intel  Pentium FDIV põrnikas which was discovered inwards 1994 but it is troubling.
From The Register:

 2 + 2 = 4, er, 4.1, no, 4.3... Nvidia's Titan V GPUs spit out 'wrong answers' inwards scientific simulations
Fine for gaming, non thence much for modeling, it is claimed

Nvidia’s flagship Titan V graphics cards may get got hardware gremlins causing them to spit out unlike answers to repeated complex calculations nether sure enough conditions, according to estimator scientists.

The Titan V is the Silicon Valley giant's most powerful GPU board available to date, together with is built on Nv's Volta technology. Gamers together with casual users volition non notice whatever errors or issues, notwithstanding folks running intensive scientific software may come across occasional glitches.

One engineer told The Register that when he tried to piece of occupation identical simulations of an interaction betwixt a poly peptide together with enzyme on Nvidia’s Titan V cards, the results varied. After repeated tests on 4 of the top-of-the-line GPUs, he works life ii gave numerical errors close 10 per cent of the time. These tests should hit the same output values each fourth dimension ane time to a greater extent than together with again. On previous generations of Nvidia hardware, that to a greater extent than ofttimes than non was the case. On the Titan V, non so, we're told.

We get got repeatedly asked Nvidia for an explanation, together with spokespeople get got declined to comment. With Nvidia kicking off its GPU Technology Conference inwards San Jose, California, side past times side week, perchance together with thence we'll kicking the bucket merely about answers.

All inwards all, it is bad tidings for boffins equally reproducibility is essential to scientific research. When running a physics simulation, whatever changes from ane piece of occupation to merely about other should endure downwards to interactions inside the virtual world, non rare glitches inwards the underlying hardware.

Collisions
Take for instance software that models molecular interactions. This kind of code uses Newtonian equations to predict the nation of a organization at whatever given time, such equally calculating the grade of particles afterwards collisions. If a simulation has the same surroundings together with starts alongside the same conditions, the output should endure the same, ane time to a greater extent than together with again. But that isn’t ever the instance when using Nvidia’s Titan V GPUs to squelch the numbers.

An manufacture veteran, who alerted us to the issue, reckoned this is due to a retention issue. Chip companies commonly force their high-end silicon to the restrain to maximize performance. Nvidia may endure overclocking or red-lining its Titan V inwards merely about way, causing read errors from memory. These mistakes are carried forrard inwards calculations, resulting inwards numerical errors. Another crusade could endure a blueprint blunder.

It is non downwards to random defects inwards the chipsets nor a bad batch of products, since Nvidia has encountered this type of cockup inwards the past, nosotros are told. The moneybags biz released patches for merely about of its older GeForce together with Titan models that exhibited like problems to address these errors. There was no number alongside its Titan X menu based on its Pascal architecture, we're told....

 ...MORE