91% of theoretical global memory bandwidth performance for the Maxwell Titan X), the very late release of the final version of CUDA 8.0 and the lack of support for Visual Studio 2015, their priorities are apparent. NVIDIA currently has a myopic focus on ‘machine learning’ and if a bug (or set of bugs) does not impact that sub-set of their market they have little interest in fixing the problem.īetween the “not as advertised” memory bandwidth performance of the high-end Pascal GPUs ( 76% of theoretical global memory bandwidth for the Pascal Titan X vs. There is only so much one can tweak by fiddling with memory controller settings in the VBIOS or driver.Īccording to Wikipedia, September 8 marks the one-year anniversary of CUDA 7.5, and I am still having some hope that NVIDIA will use that opportunity to finally ship the final version of CUDA 8.0.Īt this point it is clear that for some CUDA applications this GDDR5x memory has problems.
#PUREBASIC LIBRARY SUBSYSTEM FULL#
My expectation is that first-generation DDR5X products will simply not reap the full benefits of this new memory type. The symptoms we all observe seem consistent with a company that has bitten off a bit too much, relative to the size of their engineering staff, and the lack of bug fixing (the mfaktc folks for example have been waiting for a fix to a compiler bug introduced with CUDA 7.0 it’s supposed to be fixed in CUDA 8.0 final) can easily be explained by a lack of engineering bandwidth, rather than a lack of good will.Īs far as DDR5X is concerned, it seems to me that this is technology that has been brought to market so rapidly that neither supplier(s) nor users have had adequate time to adjust and work out all the kinks.
Or specifed power state, I can not understand what the problem isĪs annoying, troubling, and frustrating the delays and bugs are to CUDA users, I do not see evidence that NVIDIA is ‘blaming the user’ for any of these problems. May be i need some other way to initialize the GPU, not the same as for the old architecture… On older architectures everything works fine (Kepler, Maxwell)Īs I understand it, the reason GDDR5x memory On the gtx 1070 I get performance 31Mh/s and on gtx 1080 only 4Mh/s. Look, I wrote a miner, a program for mining cryptocurrency ethereum based dagger Hashimoto algorithm (ethash) Last I checked there were already more than 500 different cryptocurrencies … That game has been going on for years, I really wonder why anybody still cares. If the problem is low performance AND this is a new algorithm, the reason is probably that the people who cook up new coin algorithms purposefully try their best to make them as inefficient for accelerators (GPUs and FPGA) as possible.
What do you mean by “bad result”? What is “dagger-Hashimoto”? Since you mention “mining” I assume it is some sort of cryptocurrency thing.
0 kommentar(er)
