NVIDIA's high-end GPU 'Titan V' has calculated results different each time and the engineers scream

NVIDIA's high-end GPU "Titan VA simulation engineer who uses it in the simulation screams that it's okay as the result of calculation changes every time it is calculated. The Register is killing the whimsical phenomenon of Titan V, which is why "2 + 2 = 4, yea, 4.1 ... ... 4.3 after all".

2 + 2 = 4, er, 4.1, no, 4.3 ... Nvidia's Titan V GPUs spit out 'wrong answers' in scientific simulations • The Register
http://www.theregister.co.uk/2018/03/21/nvidia_titan_v_reproducibility/

NVIDIA'sTitan VIs a GPU adopting the Volta architecture, including 5120 CUDA cores and 12 GB HBM 2 memory, is the highest graphics board for consumers. Since the single precision floating point operation performance is 13.8 TFLOPS and the Tensor calculation performance is 110 TFLOPS and it is designed for deep learning, it is assumed to be used not only in gamers but also in a wide field such as machine learning.


According to The Register, an engineer conducted a simulation of the interaction between protein and enzyme and found that different results would appear in exactly the same condition. Calculation test using Titan V was carried out, and it seems that the result of 2 of them resulted in about 10% opening. In this type of simulation, it is natural that the same numerical value is output every time, and it is pointed out that it was a phenomenon peculiar to Titan V that NVIDIA graphic board up to the previous generation Pascal architecture did not result in such a result.

An anonymous technician wary of interference from NVIDIA said to the Register, "I will avoid using Titan V until a software patch corresponding to this mathematically strange problem is released. As Titan V is priced at US $ 2999 (about 310,000 yen) and it is expensive as a graphic board, it seems that Titan V suffered a big pain due to being unable to calculate it although it was introduced for research .

According to The Register, some industry experts familiar with the GPU think that there are some memory problems. Titan V has the possibility of causing memory reading error in a technique to improve performance such as overclocking. In addition, it is pointed out that the possibility of error in design is also in the first place. In addition, it seems that the error that the output of the calculation result found this time fluctuates is hardly a problem in the use of general game.


The Register inquired about NVIDIA to output different calculation results by Titan V, "NVIDIA has confirmed that using Amber of biomolecule simulation software resulted in at least 1 strange impact on Titan V Although they are aware of the report, I do not think that Titan V has a design problem, and users who experienced the error are requesting information to "[email protected]" I have answered.