From Hexus.net

Compute machine and multitasking. But GTX 4x0 is crippled

As much of a general-purpose computer as a GPU, the parallel architecture is also designed for the high-performance computing segment in mind. The enhanced cache structure, detailed above, helps with general computations, and GF100’s adherence to the IEEE 754-2008 floating-point standard means that it can run high-accuracy tests (double-precision support) at an increased rate when compared to anything NVIDIA has designed before.

Delve a little deeper, handily not mentioned in any briefing, and NVIDIA is limiting the double-precision speed of the desktop GF100 part to one-eighth of single-precision throughput, rather than keep it at half-speed, as per the Radeon HD 5000-series. We'll have to wait for the Tesla parts before that's restored to Radeon-matching levels.

If true, it would mean only about two times the performance for a GTX480 compared to a GTX285. While it definitely helps, it won't close the gap to ATI's offerings (factor 6 or so would have been needed). And I doubt most here are willing to shell out thousands of dollars for the Tesla versions.

That report looks dodgy. The author seems to be confused about the benefits of IEEE compliance and quotes ATIs cards as having 1/2 the sp speed in dp. It should actually be 2/5, which is 40% instead of 50%.

MrS

---

Scanning for our furry friends since Jan 2002

...quotes ATIs cards as having 1/2 the sp speed in dp. It should actually be 2/5, which is 40% instead of 50%.

FYI, AMD claims (on their web site) 0.544TFlops double precision vs. 2.72TFlops single precision for the HD5870. Similar ratios are shown for the HD5850 and HD5830.

That report looks dodgy. The author seems to be confused about the benefits of IEEE compliance and quotes ATIs cards as having 1/2 the sp speed in dp. It should actually be 2/5, which is 40% instead of 50%.

In fact, it's 2/5 only for adds, for multiplications or combined multiply-add operations it's only 1/5 (for fused multiply-adds it's 1/4) on ATI cards. But as the theoretical SP power is that much higher on ATI compared to nvidia (roughly 2 times as high) and with DP there are no problems to fill the 5 parallel slots of the VLIW units, it effectively compensates this theoretical disadvantage to Fermi.

A Tesla C2050 with 448 SPs and 1215 MHz shaderclock (like the GTX470) would have exactly 544 GFlop/s theoretical DP peak throughput. I guess it is pure coincidence that a HD5870 with its 320 VLIW units (5 slots each to arrive at 1600 SP ALUs) running at stock 850MHz arrives at the exact same theoretical 544 GFlop/s ;)

But if nvidia disables a large part of that for consumer GeForce cards (hedging the Tesla market?), it won't end up nowhere near a HD5870. A GTX480 class GPU could be able to do 480 SPs/2 * 2 Flop * 1,4GHz = 672 GFlop/s. If nvidia cuts this down to a fourth, it would be only 90% faster than a GTX285 (+ maybe some internal improvements) with double precision, but it would need more like a factor of six to threaten ATI's top model here at MW.

But we will see. Theo Valich from BSNews announced to do some GPGPU testing (presumably including some BOINC projects) on a GTX480 in the next week.