The GPU Fantastic! Does it really exist? Every now and then we may think so, but check back six months later and there's always something more light; more fantastic. Another GPU has dawned, and James Wang is here to explain why.
'We fucked up.' That was basically NVIDIA's executive summary on the GeforceFX 5800 Ultra (NV30) during the NV35 launch at E3. For a company with a reputation in PR stunts, such a clear confession is hard to come by. But obviously its intention at E3 wasn't just to admit to the previous failure, but to show the world its weapon of redemption: the GeforceFX 5900 Ultra (NV35).
Competition is goodLooking at the graphics market today, one truly can't help but appreciate the immense benefit of competition. From flagship launch to refresh, this must be NVIDIA's fastest product cycle -- all thanks to ATI. Of course there is question of whether the NV30 even counts as a product 'cycle', given how fast it came and went with few boards available in retail channels. Putting the past behind, I'm glad to say that the NV35 is the real deal. Gone is the dust buster, however, the new 'brick-sink' solution still takes two slots.
Drivers have matured considerably and overall the whole board resembles more what you can expect to buy, rather than a crude engineering sample.
It is interesting to note that NVIDIA and ATI's fabrication strategy is totally reversed. At the March 2003 Morgan Standley investor briefing, NVIDIA's CEO said its strategy to use 0.13-micron for NV35, but employ 0.15-micron for low-end was 'one of our smartest decisions.' His argument is because the low-end requires larger volumes, employing the cheaper 0.15 production lines will allow it to churn out more chips than if it did on the few 0.13 production lines available. ATI has instead put its flagship product on a mature process while migrating to 0.13 for their RV350 mainstream product. Due to the lower power and smaller die, ATI managed to instantly produce a killer mobile DX9 GPU. Both strategies have been rewarding to their respective companies but if 0.13 yields are as healthy as ATI is claiming, then it is likely to edge out NVIDIA in the mainstream race.
What is a GPU anyway?Before jumping into the tech, it's important to keep a clear view of how GPUs are evolving. Sure, everyone knows they're now programmable and function more like CPUs, but how does this change the entire paradigm of a 'PC'? Will there be a day when you can buy a GPU in socket form and plug it right next to the CPU with both processors sharing a common pool of high speed memory? Will programmers eventually have to make a choice on where they want their intensive floating point (decimal) calculations to be done? To answer those questions, let's have a look at how innocent graphics accelerators have evolved into their current monstrous forms.
It all began with drawing pixels. Hardwired circuits were used and that was all they did, draw pixels. 3D graphics weren't too much different. Simply, more specialised hardware blocks were added to fulfil the functions of 3D, such as texturing, filtering and converting 3D to 2D display. The early problems of graphics were specific and hence were resolved with specific hardware.
Fast forward to today, simple and specific graphics problems are all resolved. The problems now are in simulating uncountable amounts of visual phenomenon whose solution under the old paradigm would require infinite hardware blocks. Obviously this is not a solution. By examining many of the 3D problems (reflection, water, shadow) in detail, it just happens that the vast majority of them fall into the category of vector and floating point calculations. Hence the natural solution would be to produce a hardware solution that hosts multiple units dedicated to such operations. Whereas before you can point at one of the Voodoo2's chips and say: 'this chip textures pixels in Tomb Raider,' today you can smash open a GeforceFX and say: 'this block is a vector engine. It can be used to animate vertices, find lighting vectors, or do four pages of my maths homework in parallel.' In a nutshell, today's GPUs have few hardwired parts to solve particular 3D problems. They're simply a specialised math calculator whose pool of resource is specific to the needs of 3D graphics. This is probably the best spiritual description of the NV35.
Refreshing bandwidthNV35 to NV30 is what the GeForce 2 GTS is to the Geforce 256; it's the refresh product of a new architecture. This is often the best product to buy, having matured on a relatively new manufacturing process with most bugs ironed out. The refresh is almost always more balanced, efficient and cooler. Like the first GeForce, the NV30 suffered a major imbalance from the lack of memory bandwidth; although it sported the fastest memory at launch, the 128-bit memory interface was too narrow to unlock the potential of the chip. The NV35 is given a health boost in bandwidth from its predecessor, leaping from 16GB/s to 27.2GB/s. This is the highest bandwidth for any GPU to date, outstripping its closest competitor by a large margin. Just to put it into perspective, an NV35 has the equivalent bandwidth of ten VSA-100 chips or five Voodoo5 boards in parallel!
ATI and NVIDIA have upped their latest flagship products to an insane 256MB of memory. But when you consider that 1,600 x 1,200 with 6x antialiasing uses around 95MB of frame buffer, it doesn't seem too crazy anymore. With vertex data, high resolution textures and longer shader programs all competing for memory; it never hurts to have more. The memory chips on the NV35 also happen to be arranged octagonally, as first seen on the Matrox Parhelia. The symmetry and equidistant spacing between chip and memory using this arrangement provides better signal routing and synchronisation.
NVIDIA has divulged no information on the pipeline organisation of its NV3x series. It is however the general consensus that it has separate execution units for integer, FP16 and FP32 colour formats. The inclusion of the legacy integer format is probably to strengthen workstation performance where dedicated integer hardware has optimal performance.
However according to unofficial channels, the integer units have been removed from the NV35. This has made room for great improvements in shader performance which was severely handicapped on the original NV30. Although this still doesn't stand up to NVIDIA's claim of two-fold increase, shader performance is now shoulder to shoulder with the RADEON 9800 PRO (R350).
Squished bugsThe first GeforceFX had a broken anisotropic filter and a sub-par AA algorithm. New drivers have arrived with the NV35 to rectify these problems. Anisotropic filtering has indeed received a face lift; at equivalent settings the results are directly comparable to the R350. Antialiasing however is another matter. While NVIDIA stresses the importance of their colour compression technique, the actual quality of their antialiased samples is not comparable to ATI's. As it stands, NVIDIA's current AA algorithm is pretty much identical to the Geforce4. When we brought this issue to their Chief Scientist David Kirk, he emphasised the importance for more real samples over hackery algorithm improvements. Although NV35 can perform eight samples, the end quality is hardly improved from its 6x implementation. ATI's gamma-corrected rotate grid sampling technique is still the best we've seen.
ABOVE: The bottom i
Issue: 137 | June, 2012