Slashdot Mirror
NVIDIA Drops Surprise Unveiling of Pascal-Based GeForce GTX Titan X (hothardware.com)
MojoKid writes from a report via HotHardware: Details just emerged from NVIDIA regarding its upcoming powerful, Pascal-based Titan X graphics card, featuring a 12 billion transistor GPU, codenamed GP102. NVIDIA is obviously having a little fun with this one and at an artificial intelligence (AI) meet-up at Stanford University this evening, NVIDIA CEO Jen-Hsun Huang first announced, and then actually gave away a few brand-new, Pascal-based NVIDIA TITAN X GPUs. Apparently, Brian Kelleher, one of NVIDIA's top hardware engineers, made a bet with NVIDIA CEO Jen-Hsun Huang, that the company could squeeze 10 teraflops of computing performance out of a single chip. Jen-Hsun thought that was not doable in this generation of product, but apparently, Brian and his team pulled it off. The new Titan X is powered by NVIDIA's largest GPU -- the company says it's actually the biggest GPU ever built. The Pascal-based GP102 features 3,584 CUDA cores, clocked at 1.53GHz (the previous-gen Titan X has 3,072 CUDA cores clocked at 1.08GHz). The specifications NVIDIA has released thus far include: 12-billion transistors, 11 TFLOPs FP32 (32-bit floating point), 44 TOPS INT8 (new deep learning inferencing instructions), 3,584 CUDA cores at 1.53GHz, and 12GB of GDDR5X memory (480GB/s). The new Titan X will be available August 2nd for $1,200 direct from NVIDIA.com.
I thought it had been surpassed by C++, but this is great for everyone.
Yes, but it had to shoot first.
The card is designed for data mining and neural network research; it's not for games or even remotely intended to be used for them.
A lot of researchers are using GPUs for things very different than graphics. A professor was telling me just last week that the boundary between a machine learning training algorithm being interesting was to train to deal with a problem in a week or less [note one trained it does its job much faster, that's just the get-it-ready-to-go time], and that GPUs were often used for that training. The bit width requirements are modest, but the amount of data to process is huge.
Of course, he went on to show how the approaches his students had come up with were faster and more power efficient by orders of magnitude for many common algorithms, but still they were trying to improve a normal way of doing things, which is to get up and running fast using GPUs are a source of number crunch.
In summary, people who don't actually need so more horsepower buying it helps keep it being developed for the smaller number of people who get it who are actually doing something useful with it.
Nvidia uses scientists names for their products: Tesla, Fermi, Kepler, Maxwell, Pascal and Volta (next version after Pascal).
If the leading edge of "consumer" graphic cards is of any interest to someone, they'd know what Pascal was since it's been announced now for over 2 years.