The curtain’s been dropped on much of Nvidia’s upcoming Fermi-based graphics cards, and the five-hundred-and-twelve-core GF100 looks like a behemoth, indeed. A completely overhauled architecture is all about three things: scalability, parallelism and geometry. Oh, and ripping your eyeballs out.
Tom’s Hardware, Anandtech, HotHardware and others go pretty deep on the new architecture, which is now eminently scalable. Here’s the overall structure of the GF100, which should give you an idea of the scalability – the GF100 is made up of four graphics processing clusters (GPC), themselves composed of four streaming multiprocessors (which are made up of 32 CUDA cores and texture units) and a raster engine:
To go deeper on architecture, you’re better off reading the 10-page reports from any of the sites linked above, but bottom line, Tom’s Hardware is predicting something like double the performance of Nvidia’s current GTX 285. Anandtech also points out that Nvidia’s geometry performance only went 3x between the NV30 engine in the ancient GeForce FX 5800 and current GT200 in the GTX 280, but the Fermi-based GF100 has 8x the geometry performance of the GT200. The endgame being that ” it allows them to take the same assets from the same games as AMD and generate something that will look better. With more geometry power, NVIDIA can use tessellation and displacement mapping to generate more complex characters, objects and scenery than AMD can at the same level of performance.”
There is a cost. Even though it’s at the 40nm process, those three billion transistors are going to run hot, and the GF100 maybe the hottest single-card GPU ever. It’s also not going to be cheap. At all. [Tom's Hardware, Anandtech]
matt
Tuesday, January 19, 2010 at 11:11 AMmeh, it will just be bigger, louder and more power hungry for not alot of noticeable improvement. just like the ATI ones.
also, that stuff about the geometry speeds is misleading. don’t expect twice the difference in performance than the difference between the 5800 and the 280…
however, I think its a good way to go, with tessellation. The biggest bottleneck with Gfx cards these days is memory bandwidth and latency, that’s why the increase in performance with the latest ATI cards was directly proportional to the increase in memory bandwidth, of around %50, not the 100% increase in processing power. things like tessellation should mean less geometry read from memory and thus less of a reliance on memory bandwidth. the same was true for the geometry shader, with procedurally generated geometry actually faster than already generated geometry because that had to be loaded from mem.
With processing power far outpacing memory performance, its actually QUICKER to generate everything on the graphics card than it is to have it pre-generated and sent to the graphics card!
the trick is DEVELOPERS need to take these new features and use them. and with them all busy with outdated console tech, this might take a while, the good news is the tessellater actually does something developers have been doing manually for a while now anyway, LOD, this should encourage developers to adopt it faster, tessellation is also in the 360, so, hopefully people have been adopting it.
Thomas Hambleton
Tuesday, January 19, 2010 at 3:59 PMWill be fantastic for running other calculations on though. Nvidia are obviously positioning their cards as a “co-processor” of the 2010s to do big analytic work on.