While the GeForce GTX 1080 will be on the market within 10 days, NVIDIA has raised the NDA on testing. It is thus possible to learn more about this high end card, waiting round the GTX 1070 and the response from AMD.
A few days after the announcement of the GeForce GTX 1080 NVIDIA has left Press published the first reference card tests recovered during his lecture.
the opportunity to make a first review of this model to be proposed at no less than 789 euros in France (699 dollars announced), pending those partners and the sale, scheduled for May 27 As a reminder, the 1070 GTX is not expected before June.
Pascal delivers its last secrets
This is also an opportunity to discover some additional details concerning architecture compared to what had already been revealed in our previous analysis.
We learn that the 2560 CUDA Cores are divided into four GPC each containing five SM 128 units. The texturing units are almost at the 980 Ti (160 against 176 and 128 to 980) but that the ROPs are at the 980: 64 (96 for the Ti model). The L2 cache is still 2MB The GP104 chip, engraved in 16 nm, measuring 314 mm² (7.2 billion transistors) against 398 mm² for GM204 (5.2 billion transistors).
on the side of the memory interface, the 256 bits are divided into 8 x 32 bits (rather than 4 x 64-bit) and NVIDIA Pascal shows have reinforced the compression level of 4: 1 and 8: 1 against 2: 1 above. This last mode is also improved. In practice, the necessary bandwidth reduction would be 11 to 28% depending on the games, while the available (320GB / s) is higher compared to a 980 (224 GB / s). Nevertheless, it remains lower than that of a 980 Ti (336 GB / s).
in terms of Async compute, NVIDIA has revised its architecture to improve things with Dynamic Load Balancing to take full advantage of the GPU resources for graphic calculation or compute as needed without waiting that one or another of the tasks are completed. So far he had to go through a static allocation, which was not to the advantage of GeForce.
Then with preemption, which is possible at the pixel level, but also threads and instructions for the Compute. Thus, a calculation can be paused in order to perform another deemed more critical, before resuming the first. It remains to see if it really will gain a noticeable compared to AMD architectures on this point, especially against Polaris should not be long till the end of his nose.
The SLI is also entitled to a facelift with a new bridge” HB “for High Bandwidth, according to NVIDIA take this technology on larger definitions. In practice, it remains to be seen, especially the multi-GPU is not necessarily the trend of the moment.
In addition to the arrival of the DP 1.4 in addition to HDMI 2.0, we learn that the HDR will be arriving in several games and can be used via the service local streaming GameStream and Shield Android TV, subject to availability HDR of a TV with HDMI port 2.0.
for fans of a fluid display, a new software solution comes to replace G-Sync for those who do not have a compatible screen: Fast Sync. This promises (again) to take advantage of the vertical synchronization without tearing and with little latency. The result remains to be discovered in practice.
The results appear at the rendezvous, AMD is preparing to answer
As for performance and overclocking, the promise seems held according to the tests of our brothers since it is generally well beyond the GTX 980 Ti and R9 Fury X, with a much lower consumption and the possibility of passing the 2 GHz. But again, it will be interesting to wait for other manufacturers models, more accessible, to see what it is in practice.
Besides, AMD will not let and promises of the new Polaris aside for tomorrow. So the battle has just begun, and should continue at least until the end of Computex which will open at the end of the month Taipei
in French.
English
David Legrand
editorial director and head of L @ bs Nancy. Geek extreme specializing in the analysis of high-tech products, social networks and trio of screens. Adept free.
No comments:
Post a Comment