NVIDIA's GeForce GTX 580

[Jeffrey]

NVIDIA GeForce GTX 580 is the world's fastest DirectX 11 GPU. With screaming performance and an innovative new thermal design, GeForce GTX 580 rips through the latest games while running quieter than the previous generation. GeForce GTX 580 – don't play DX11 without it. To refresh its high-end, Nvidia has not really developed new GPU, but rather refined the GF100, the GPU GeForce GTX 480/470/465. If Nvidia has decided to name this new revision of GF110 GF100 for reasons of image, it actually corresponds to what has been mentioned often as GF100 B1 is an optimized version and partly corrected GF100 used on A3 previous GeForce. During the development of GF100, Nvidia has encountered many problems, the most important of course have been corrected through several revisions of the chip. Jssen-Hsun Huang, Nvidia's CEO, moreover, we noted a few weeks ago that poor coordination between different teams working on the chip was the source of some of these problems, particularly in terms of communication between units executions and memory controllers.

It is of course not possible to refine a design forever, the day comes when we must launch a product even if it means some compromises from the original plan. This was due to Nvidia with GF100 and the A3 GeForce GTX 480. With a yield lower than expected level of production, Nvidia had to castrate the ultra high end by disabling one of its 16 blocks of threads, a first. The sequence of several revisions (A1, A2, A3) and a higher consumption than expected have made difficult the development of an efficient cooling system. This is probably the biggest mistake by Nvidia with the GeForce GTX 480 which has suffered from a poor image due to its noise and heat.

Since the GF110 is a revision of GF100 it logically includes the internal organization. This GPU is composed of four main structures, which include the GPCS each 4 DMs, the blocks of threads. Each of these DMs with 32 computing units (the "cores" in a more commercial) and 4 texturing units. This little world is attached to 6 of 64-bit memory controllers which include each 8 ROPs, responsible for writing the pixels into memory, and 128 KB of L2 cache. So we get a total of 512 units of calculation, 64 texturing units, a 384-bit memory bus, 48 ROPs and 768 KB of L2 cache. All for a huge die 3 billion transistors. If the support of computing in double precision is always present in GF110, it is not active on the GeForce derivatives, are reserved to the professional world.

For details on the architecture of GF110, our analysis related to that of GF100 remains valid. What has changed between the GF110 and GF100? Few things in the end, the main result of fine tuning a design that can more easily produce the GPU, up slightly higher in frequency and consume less, equivalent specifications, by reducing leakage currents. To do this, Nvidia said it had revalued each block of the GPU to replace, wherever possible, high-performance transistors with transistors more efficient.

 
Some small fixes were also made. Thus, the texturing units are now able to filter FP16 textures (and therefore FP10/11 and RGB9E5) at full speed while on the GF100 A3 they had to settle for a rate cut in half. Since this revision of the GPU had already adapted to these textures data path, which is what is most expensive to implement, we assume that filtering at full speed was indeed planned but did not work for some reason unknown to us. More than adding to the GPU, so it is rather a fix.

Nvidia also said to have added support for additional formats tiles to maximize the effectiveness of culling, the removal of hidden surfaces, in many situations. The video engine and the audio controller are not changed by cons. Unlike small GeForce 400 and GeForce GTX 480 as the GeForce GTX 580 and other possible models based on the GF110 does not bear the bit streaming HD audio formats.

Features :

1. Microsoft DirectX 11 Support : DirectX 11 GPU with Shader Model 5.0 support designed for ultra high performance in the new API’s key graphics feature, GPU-accelerated tessellation.
2. NVIDIA PhysX Technology : Full support for NVIDIA PhysX technology, enabling a totally new class of physical gaming interaction for a more dynamic and realistic experience with GeForce.
3. NVIDIA 3D Vision Ready1 : GeForce GPU support for NVIDIA 3D Vision, bringing a fully immersive stereoscopic 3D experience to the PC. A combination of high-tech wireless glasses and advanced software, 3D Vision transforms hundreds of PC games into full stereoscopic 3D. In addition, you can watch 3D movies and 3D digital photographs in eye popping, crystal-clear quality.
4. NVIDIA 3D Vision Surround Ready2 : Expand your games across three displays in full stereoscopic 3D for the ultimate “inside the game” experience with the power of NVIDIA 3D Vision and SLI technologies. NVIDIA Surround also supports triple screen gaming with non-stereo displays.
5. NVIDIA SLI Technology3 : Industry leading NVIDIA SLI technology offers amazing performance scaling for the world’s premier gaming solution.
6. NVIDIA CUDA Technology : CUDA technology unlocks the power of the GPU’s processor cores to accelerate the most demanding tasks such as video transcoding, physics simulation, ray tracing, and more, delivering incredible performance improvements over traditional CPUs.
7. 32x Anti-aliasing Technology : Lightning fast, high-quality anti-aliasing at up to 32x sample rates obliterates jagged edges.
8. NVIDIA PureVideo HD Technology4 : The combination of high-definition video decode acceleration and post-processing that delivers unprecedented picture clarity, smooth video, accurate color, and precise image scaling for movies and video.
9. PCI Express 2.0 Support : Designed for the new PCI Express 2.0 bus architecture offering the highest data transfer speeds for the most bandwidth-hungry games and 3D applications, while maintaining backwards compatibility with existing PCI Express motherboards for the broadest support.
10. Dual-link DVI Support : Able to drive industries largest and highest resolution flat-panel displays up to 2560x1600 and with support for High-bandwidth Digital Content Protection (HDCP).
11. HDMI 1.4a Support5 : Support for HDMI 1.4a including GPU accelerated Blu-ray 3D support, x.v.Color, HDMI Deep Color, and 7.1 digital surround sound.

[Jeffrey]

The Chipset :

The GeForce GTX 580 benefits from the use of a full GPU, but no units are off, and frequencies revised upward to show a capacity to load and texturing as well as flows of triangles and pixels up to 18% compared to the GeForce GTX 480. Memory bandwidth by moving against more modestly with a gain of 8%. Like AMD's 40 nm reuse TSMC for its second generation of DX11 chips, Nvidia was forced to do the same with her newborn. By cons, although the name GF110, this GPU is not actually much again and may well have known GF100 A4 B1 or rather given the magnitude of the revision. Indeed difficult to imagine that Nvidia GPU has designed a very complex and finally consciously decide to amputate a few million transistors (in its most upscale designs that are down in line to sell the units partially defective during etching) by marketing. The area of silicon used is still very important, but this time TSMC and Nvidia have been able to make reliable enough production to provide fully functional chips.

With the revision effective architectural Fermi interview on the GF104 GTX 460 which has brought a different distribution of transistors for the benefit of texturing units (TMU) and Cuda cores, many have imagined that the successor to the GF100 would go back on these bases. Always engraved in 40nm by TSMC, the new GeForce GTX 580 chip embeds three billion transistors, thus includes more "CUDA cores" and displays higher operating frequencies while having a lower thermal envelope thanks to various improvements made by the chameleon . The TDP of the GeForce GTX 480 which was going from 250W to 244W and the GTX 580 so that the frequencies are now 772/1544/1002 MHz for the core, the SP and the GDDR5 memory, cons 700/1401/924 MHz ago . In absolute terms, the TDP of 244W is very high, but this is where the new cooling solution nVIDIA.

This large number of DMs allows GF100/110 to be very effective in many situations but of course once more per unit of raw computing power. It can also give a clear advantage in terms of performance relative to the geometric midrange which traditionally was as swift as the high end. Coupled with the presence of 4 small rasterizers, a GPC, it is able to maintain a good speed when cutting small primitive pixels, where heavy rasterizers are necessarily less effective. All the ingredients are there to process the tessellation with an excellent level of performance and Nvidia has obviously take this opportunity to stand out from the competition, which leads to intense lobbying from developers for the use of this technology, This can lead to some questionable results, as is the case in HAWX 2, which seems to have overlooked many aspects of graphics in favor of a tessellation of extreme terrain that we are not asking if it is really justified from the point of qualitatively. Anyway, the next all-terrain "of GF100/110 allows him to swallow these mountains of triangles without worries.

Cooler :

At first, the cooling solution of the GTX 580 is no different from that of the GTX 480. The heat pipes are no longer apparent, but overall it retains a dual slot solution with a fan and a plastic shroud. However, great efforts have been made by nVIDIA side cooling. The most notable was undoubtedly the arrival of an evaporation chamber, a process already used by some manufacturers such as Sapphire graphics cards for their high-end overclocked. The principle is simple: A liquid is placed in a sealed chamber, with the heat generated by the GPU, it turns to steam and spreads away from the GPU. On this side, the chamber is in contact with heat sink fins that they receive the air blown by the fan. The steam, once in cooler box turns back into water and returns to the GPU to begin the cycle again.

The fan has also undergone some optimizations that allow it to vibrate and generate less noise less objectionable. Its management has also been revised with a speed that will vary more reactive function of the temperature of the GPU. Finally the rear fascia is beveled, which has the effect nVIDIA according to channel air in such use multi GPU (SLi).

[Dealing with Heat :

One of the other new GeForce GTX 580, rather headlock for testers for that matter, is the power control. With older GeForce, a system was responsible for monitoring the evolution of the GPU to heat, if necessary, to reduce frequencies to prevent potential damage. With his new baby, nVIDIA goes further by checking real-time feed the card through its port and two PCI Express power connectors 8 and 6 pin. If the voltages exceed the expected values, two small chips are responsible for lowering the frequency of 50%. But uh? What happens when I launch a Crysis all details thoroughly with my overclocked GTX 580 and Vision 3D activated? In theory anything since the control equipment is also associated with a software control. For the frequencies are changed, we must firstly that the card exceeds the intensity of energy available and, secondly, that the pilots detected ugly software quarantined.

For us, this is exactly what it is because of "security" does goes so far implemented only in Furmark and OCCT, two stress tests which push the cards to their limits and allow to get an idea of the maximum power of graphic cards. By putting them in quarantine, they are no longer anything on GTX 580 as soon as the drivers detect the frequencies are divided by two.
To give you an idea, with a GTX 480 we passed the 420W supports our test config. With the GTX 580, the same platform and same OCCT test, we do not overcome the 270W, which is approximately what we get with a GTX 460. Of course, in practice few applications "real" cards that push as far and OCCT Furmark. Neither Crysis, STALKER a COP or a Metro2033, three of the most demanding games of our test panel do our config exceed a consumption of 400W. However, we believe it is interesting and important to know the full potential of consumption of a card, even if it is true that we would also result in the past also connect to the consumption of full play

[Jeffrey]

Direct X Support

But nay, nVIDIA persists and signs, its initial distribution was good, the GF104 uses one different because it is designed to ensure a good compromise in terms of performance with both games DX9/10 (more sensitive to performance texturing) and DX11 (more sensitive to the geometrical power by nVIDIA) and given a limited number of transistors while the GF110 does not have this problem. Finally the architecture of the latter is the same as the GF100.

Memory Controller

Knowing that the cost of production depends on the surface of silicon used as much to shoot itself in the foot, the chameleon has been forced to resort to amputation to compensate for these production problems / design of the chip that would have required a new revision to address before marketing. But then, competition being what it was at the time, Nvidia has the resolve to launch a chip "imperfect" to occupy the ground while working on a new revision of the latter who will ultimately name GF110. Physically, it is covered with a heat spreader like its predecessor, the packaging is the same. We thus find four GPC (Graphics Processing Cluster) related to 6 64-bit memory controllers (each with 8 ROPs) and including them within a Raster Engine and 4 SM (Streaming Multiprocessor) each composed of 32 computer units or scalar Cuda Cores , 4 TMU and Polymorph Engine dedicated to geometry and in particular the much talked Tessellation with DX11.

Comparison :

In terms of fill rate, the Radeon has a big advantage over the GeForce GTX 400, especially in FP10 format, processed at full speed while the GeForce he spends half speed. Given the limitation of these GeForce GTX 400 in terms of data path between SMs and ROPs, it is unfortunate that Nvidia has not given her the opportunity to enjoy GPU formats FP10 and FP11. The GeForce still retain some advantages. First of all they can handle the FP32 single channel at full speed without blending. Then they keep blending with maximum efficiency INT8 where the Radeon suffer a loss.

Given the architectural differences of the different GPUs in the processing of geometry, we obviously looked more closely at the subject. First we observed the flow of triangles in two cases: when all triangles are displayed and when they are all rejected. The GeForce GTX 580 is very powerful here, ahead of the GeForce GTX 480 by 30% when the triangles are rendered while theoretically gain should be between 10 and 18%. Since the GeForce are artificially limited at this level compared to Quadros we assume that Nvidia has left here a little more space compared to the GF110 GF100, introducing so an artificial difference between these two revisions of the GPU. When it comes to eject triangle rendering via culling, none comes close Radeon GeForce upscale.

The advantage of the GeForce Radeon is evident here. AMD and Nvidia have very different approaches and GeForce show a rate that varies depending on the number of processing units of the map so that it varies depending on the frequency for the Radeon. The architecture of the past that they are quickly clogged by the amount of data generated, which drastically reduces their speed in this case. Doubling the size of the buffer dedicated tessellation unit in the GPU Radeon HD 6800 allows them to significantly outperform the Radeon HD 5800, but cannot compete with the GeForce.

Strangely, the GeForce GTX 580 only shows a very low gain on the GeForce GTX 480 and is even less efficient in this test when rendering triangles are ejected through culling. We assume that this is related to a different software that favors some cases (probably most important), but few cause a decrease in others, we had already seen this with certain profiles of Quadros. The GeForce cashing here much better the burden tessellation that Radeon 6800 series, although the result shows a gain. Using an adaptive algorithm which will regulate the level of tessellation following areas that will receive more or less detail, depending on the distance or after the resolution of the screen allows in all cases significant gains and is more representative of what will be the developers. The gap between the GeForce and Radeon is then reduced, and even before they begin in some cases.

Conclusion :

You were probably rehashed when you were younger, review is good, and probably nVIDIA retained his lesson. Indeed, with the GTX 580 brand the chameleon has revised its copy (double) and we are not proposing a new architecture, but a new revision of the architecture Fermi GF100 inaugurated with the GeForce GTX 480. Better optimized, as it likes to call GF110 nVIDIA includes the use of new transistors. Sometimes more swift in sensitive areas, sometimes less, where the brand could afford. Then, while the GeForce GTX 480 embarks was "only" 480 cores executions, nVIDIA has been reactivated with its latest enhancements Streaming Multiprocessor (SM) and so we provide a "real" farm with its 512 cores. Each SM is provided with four texturing units, the climb TMUs also from 60 on the GTX 480 to 64 on the GTX 580. Feel free to read our article on the GeForce GTX 480 to get a little more information about the architecture Fermi.

The Radeon HD 5870 is clearly exceeded by the GTX 580, designed exclusively for high-end configurations, with a focus that will provide power to the motherboard. We now await the response from AMD architecture with Caiman and HD 6900, expected in December. Only the HD 5970, dual GPU's condition, still manages to stand up to the GTX 580.

Not surprisingly, the card GeForce GTX 580 Zotact follows the NVIDIA reference design and is a copy / paste the nVIDIA sample we received, the only distinction being located at different stickers attached. Both maps measure 26 centimeters therefore exactly like the GTX 480, they both wear the new cooling system with dual slot nVIDIA evaporation chamber, include two DVI outputs and a mini HDMI output, two PCI Express power connectors 6 and 8 pins, and of course two cards for SLI. Their frequencies are identical, and they both embark 1.5GB GDDR5 on bus 384 Bit. Basically, two GTX 580 all there is more classic. Nvidia announces having improved filtering Z-cull and added support for FP16 textures at full speed. These two changes alone would post the brand to drive up performance significantly in games.