Samsung starts producing GDDR6 RAM for next-gen graphics cards GDDR6 is twice as fast, uses less power, and will be coming in 16gb packages roughly 20% more expensive at first. Very likely, these are an order they have for Nvidia's new Ampere graphics cards which will be assembled and stockpiled over the next few months. Sometime in late summer or early fall the Ampere cards are expected to come on the market and help drive prices down lower, with other memory manufacturers also ramping up production to higher levels then they had anticipated. The rumors for Ampere suggest it has roughly the same number of cuda cores as a gtx 1080, which means it can likely deal with almost any rasterized engine easily enough, but it also has a tensor core and twice as much faster memory. For serious real time ray tracing, you might want 32-64gb of HBM4 which is still too expensive, but this is Nvidia's first introduction of the technology which they are certain to improve upon significantly within the first few years. Among other things, they need to see what uses for their ray tracing the game developers can figure out to use that are less computationally demanding, and all the myriad other uses developers can put the same circuitry to, such as AI applications and physics. With any luck, video game developers will find a significant number of cheap easy ways to use the new circuitry that cut way back on the requirements for rendering everything in real time ray traced lighting. No doubt Nvidia has already been working with developers in this area, but once the hardware is available the developers and moders alike can share their ideas with each other by programming them into the their games. That might sound like pie-in-the-sky wishful thinking, but HMB was invented specifically because affordable high speed memory has become a major bottleneck in producing affordable graphics cards, but a temporary one. In the meantime, the developers now have two very different types of circuitry for crunching the numbers faster for different things in the their video games.
NVIDIA GeForce GTX 1180 - Specs, Performance, Price & Release Date (Preliminary) Here is confirmation of the Ampere specs and release date. It will have the same number of cuda cores as a gtx 1080 ti with up to 16gb of GDDR6. That puts it somewhere over 13 teraflops of performance, which is the estimated requirement for running just about any existing video game in 4k at high frame rates. For UHD games you want at least 80fps if you can get them, for a greater feeling of connection between the mouse and display and you need the same high frame rates for most VR headsets. All the extra memory in an Ampere means it can store that many more frames in advance and can work the cuda cores and tensor cores simultaneously using the same memory pool. This is the kind of tiny incremental increase we might expect from Nvidia for the indefinite future, similar to Intel doling out minor improvements they make you pay through the ass for. However, having essentially two different computers on the same chip with all that extra memory makes the actual power the graphics card difficult to assess. The tensor cores are analog circuitry and it could easily be another twenty years before people figure out all the basic ways they can be used to enhance graphics. A digital computer is easier to correct for errors, while an analog one like their tensor cores can be outrageously faster and more efficient, but more prone to errors.
