AMD would be about to present AMD Vega, not as a single graphics card, but as a complete graphics family.
A leak of information ensures that AMD would be developing a new generation of graphics cards based on the Vega architecture. This new family of graphics cards would be the Radeon 500 series, which, with the arrival of AMD Zen, will constitute a new range of products, which can pose a problem for competing companies. This second generation of graphics cards will make use of GDDR5 / GDDR5X memories, although it is an unconfirmed data. If these data are true, just like what we have been seeing for weeks and assuming it is the AMD RX 490, it is not actually the RX 490, it is still a graphics card from the RX 500 series and it would not be the top of the range.
According to the Fudzilla media, which has been the one that has uncovered it, AMD is working on a new graphics generation based on the Vega architecture, which will hit the market in 2017 and will be very soon. It would come with a top-of-the-range graphics that seamlessly supports 4K resolutions at a minimum of 60FPS in Triple A games, as we supposedly saw at the New Horizon event and the mid-range and input range would have resolutions of 1440p and 1080p for games. Regarding the memories, the most powerful of this new graphics range would use HBM2 memories, while the mid and low range would have GDDR5 / GDDR5X memories.
Last year Robert Hallock, AMD Marketing Technician, confirmed that the GCN architecture is compatible with HBM and GDDR5 memories. That a new graphic family with both memories is launched on the market, in principle, should not be a surprise. Since Fudzilla it has not been confirmed if AMD uses GDDR5 or the more powerful version GDDR5X in its mid-range and input graphics.
During New Horizon, AMD reportedly showed a Vega graphic in full swing. That graphics card whose name and performance are not known, would have theoretically surpassed the NVIDIA GTX 1080. Apparently this Vega GPU would have 8GB of HBM2 memory and would surpass the GTX 1080 by 10%, in terms of performance in Doom through Vulkan with 4K resolution. (We know that there is an AMD Vega that is being used for benchmarks, but without 100% confirmed data and without seeing characteristics and the physical graphics card, we cannot affirm in a firm way the existence of it or that it may or may not surpass the NVIDIA GTX 1080 in a certain percentage)
Vega architecture
Vega, is the architecture used for its Next Compute Unit units, which were unveiled during a press conference. This architecture has some advantages over the previous generation. The main one is that each SIMD unit within Vega NCU, is now able to simultaneously process processes of variable length. We know that understanding this requires more information.
Currently GCN has a calculation unit composed of four groups of 16 vectors in each SIMD unit, which allows executing four groups of sixteen vectors during four cycles in each group of processing threads. They also have a scaling unit, which is capable of executing one instruction per cycle. It is used for time-critical tasks, performing a four-vector trend reversal of the SIMD unit at times when it is simply not good enough.
They have a drawback, which is that the sixteen SIMD units work the same, regardless of the data changes that are being entered. The SIMD unit has to run all four cycles of each processing thread, regardless of content. This causes that the execution of the package would require the same execution time in a package of four vectors, which would cause that the other twelve ALUs within the SIMD are not useful. The graphic loads are not uniform, which means that it is practically impossible to find any scenario in which all the SIMD units are occupied at any given time.
Wavefront SIMD, variable bandwidth for more performance in fewer cycles
This is not the case with the new GCN implementation within Vega. The V9 architecture has new intelligent programmers and coherence subsystems that allow several fronts of processes, of different widths, which are executed simultaneously within any computing unit capable of accommodating the workload. This avoids process slowdowns, avoiding jobs that do not produce results.
This allows us, using NCU, to finish more work in the same amount of time compared to current CUs. It also allows freeing cache memory resources for other computing blocks. It is not easy to predict the difference in improvement in resource usage and CU occupancy, but performance on graphics workloads fluctuates and good management is important to improve performance. Vega NCUs are not just faster, they are also more energy efficient.
Source: wccftech
