The graphics cards specially developed for the professional sector, the NVIDIA TITAN V, would present errors in the calculations within the field of medicine.
Although we have seen it subjected to a gaming benchmark (I don't know why this nonsense), the NVIDIA TITAN V, based on a Volta GPU, is specially designed for Artificial Intelligence and Deep Learning. The characteristic of these GPUs is that they implement Tensor Cores, a set of cores that are specially designed to perform millions of calculations and predictive tasks, very important for exposed fields. These graphics cards cannot be found on the market, as they are intended for a very specific audience.
These graphics cards, despite being incredibly powerful, seem to be having some problems in some scientific simulations, especially within the field of medicine. This is at least what an engineer who has spoken with The Register, who assures that the TITAN V cannot offer reliable results, under some specific conditions. It seems that this graph is producing an error, which gives different results, when it does the same calculations repeatedly.
The work of simulating the interaction between a protein and an enzyme has been given as an example. Theoretically identical calculations should be produced each time. This engineer assures that two out of every four TITAN V that he has been able to test would be giving errors when they run the same simulation, something that is a major problem.
There is no data in this regard, but the engineer comments that it could be due to some kind of failure when designing the memories. The Register also spoke with a veteran of the hardware industry and comments that NVIDIA may be making the TITAN V work at its technical limits, even in some cases beyond. It seems that NVIDIA would have disabled the memory error protection protocol in these graphics compared to other generations.
This problem, at the moment, has only been detected in very specific cases in the field of medicine, there is no data that it exists for other fields in which they are used. Despite this, it is an important problem that we will have to see how it is corrected by NVIDIA. It is curious that the engineer mentions that two out of four TITAN Volta have this problem. Could it be some kind of problem with the HBM2 memories?
