A few days after the release of the HD 6990, GTX 590 wants to be the response of the Chameleon to AMD Monster. According to NVIDIA this map should be launched more than a month ago, but the brand has chosen to not coming in second on the market to refine its product: one of the stated goals is to do better than HD 6990 at the level of noise.
It almost would feel that this objective is the prank as the Fermi architecture is known for a consumer and a very important thermal dissipation. The simple fact to imagine that the firm to the Chameleon could involve two GF110 on a same PCB without a still more audible than the HD 6990 map could pass for an April fool. And yet this is precisely what NVIDIA claims to have succeeded.
However, compromises were necessary: found a fan in the centre of the map as in AMD, and the frequency of the GPU has the stand down from that of the GTX 580 to remain within acceptable thermal limits. Despite this, the GTX 590 is required such as the the most demanding that we have tested all the time!
But the fact is that the map is silent. Really silent. It is even the more discrete bi - GPU card since the ATI Rage Fury Maxx, exit 12 years ago already. Mission accomplished on the noise, therefore, but remains a key question: is that NVIDIA has of surrender in terms of performance to achieve this result?
Two full GF110 under the hood
Our discussions with the trade mark, it is clear that NVIDIA has always wanted to use two full GF110 on GTX 590. As we have seen with the GTX 580, the GF110 employs 4 GPC Graphics Processing Clusters (), each 4 SM (Streaming Multiprocessors). Each SM contains 32 CUDA cores, 4 texture units and 1 Polymorph engine is a total of 512 CUDA Cores, 64 units of textures and 16 Polymorph engines by GPU.
This set is connected to a 384-bit memory controller that breaks down into six units of 64-bit, which take all 256 MB of GDDR5 to reach 1.5 GB GF110 and therefore 2 x 1.5 GB in the case of the GTX 590. At the other end of the memory controller is six partitions ROP capable of generating eight pixels in 32-bit or 48 pixels per cycle in total.
NVIDIA connects the two GF110 of the 590 GTX with its own PCI Express NF200 bridge to multiplexer interface PCIe 2.0 16 lines at the base in 16 lines per GPU.
GeForce GTX 590
GeForce GTX 580Radeon HD 6990
Radeon HD 6970
Radeon HD 6950
Stream Processors (CUDA Cores)
At the macro level, the GPU is still engraved in 40 nm by TSMC and are about 3 billion transistors on 520 mm2. The GF110 is more complex but also more massive than the Cayman of AMD since the latter did on-board only 2.64 billion transistors in an area of 389 mm2.
The architecture of the GF110 thus was fully preserved, however it cannot not say the same frequency: while the GTX 580 enjoys a GPU to 772 MHz and a memory 1002 MHz, dual-core GTX 590 should respectively only 607 and 853 MHz. Result that performance is not as good as that could have been hoped, even though the relative unimportance: GTX 580 is a map without equivalent, since the HD 6970 is positioned opposite the GTX 570. If HD 6990 displays excellent performance over HD 6970 gain, NVIDIA will therefore may not need as good performance so its 590 GTX is competitive.
0 Comments:
Post a Comment