AFTER SOME 18 MONTHS of GPU doldrums - no, I don't consider an April 2008 GeForce 9800GTX a real improvement over November 2006 GeForce 8800GTX, and same for HD3870 over the 2900 series - we're finally about to experience a quantum leap in the 3-D card performance. As you all know by now, the middle of next month should see the rollouts of both ATI HD4870 and Nvidia GTX280, the new top-end cards for both vendors.

Compared to the ~10% real 3-D speedup for either Nvidia or ATI lines over the past year, these two should each bring upwards of 50 per cent speedup compared to their previous brethren, with the HD4870 bringing even more percentage-wise: simply, it compares against a slower predecessor.

Talking about the HD4870, a deep throat says it seems to almost double the performance of HD3870 in some cases: no wonder when you combine a ~10% faster clock, above 850 MHz with 50 per cent more shaders clocked independently this time, twice as much texture throughput and upwards of 80 per cent higher memory bandwidth using the GDDR5. Not bad at all - this is a much-welcomed medicine, knowing that ATI has suffered a black eye in the performance battles against Nvidia for quite a while now.

In the computational matters that matter, whether HPC supercomputing simulations or home multimedia Photoshop filters, the users will be glad to know that, for some $350 with a single H4870, they can call their machine a " teraflop PC". Yeah, it is 1008 peak GFLOPs on the GPU - but keep in mind, it's a single precision FP figure, and the real IEEE double precision FP will more likely perform at around a third of this peak performance. Even then, it will work only if most of the data is in the graphics memory, not 10x slower going " abroad" to the system memory via high-latency PCI-E.

Now, Nvidia's entry seems even more advanced in some ways: 512 bit memory to give huge bandwidth even when using the current GDDR3, 240 complex shaders, CUDA programming with IEEE double precision FP. With Nvidia's huge marketing machine and such, I expected them to ramrod the ATI entry.

Yet, that Editor's Day a few days ago seem to have put a bit of damper on the GTX280. Instead of a more competitive ~700 MHz GPU clock I expected, where it would pretty much have no competition, the final ~600 MHz GPU clock, taken into account with other performance parameters as well, seems to put the part dangerously close to the HD4870. Still faster, of course, but possibly not by much.

I'll refrain here from DX10, OpenGL, 3DMark Vantage and so on, but look at the computation again. If the clock speeds are right, we'd end up at around 933 GFLOPs for the GTX280 in single precision FP math (no double precision numbers yet). Slower peak number than the HD4870? Yes!

Does it matter in actual performance? Probably not, since the library and application optimisation will determine which GPU steams ahead. Does it matter ego-wise, when you're speccing and creating visualisation or GPU computation clusters, where "teraflop per card" means a lot to impress the ministers signing off the budget? Most likely, yes.

Now, all the explanations for the comparatively slow clock rate - large die of nearly a square-inch, 65nm process instead of 55nm, complex structure with a zillion shader processors and plenty of on-chip memory with wide outside memory controller - do make sense on their own.

However, I feel there could be another dimension to it - those slides could also be a deliberate understatement and the card might end up faster than the early numbers say. First, it would accomplish lulling the ATI gang into a bit of "I strike luck" complacency when the HD4870 launch comes, both performance and driver tuning wise. A couple of days later, Nvidia would announce the GTX280, and, well, maybe there'd suddenly be another higher clock speed there - one that would not only break the teraflop barrier, but also up all the rest of the competitive benchmark results on the first day?

Now, we don't think it will end up MUCH faster anyway, there could always be an "Ultra" edition with 256 shaders and higher clock, before the 55nm shrink comes in. But then, HD4870X2 is not far away, another two months or less, and it is expected to have roughly the same power and price as the GTX280. So, it may be good for Nvidia to do that bit of extra speed-tuning push for the GTX280 anyway, pre or post launch. This time, the competition won't back off so easily...