In addition to the plans for the release of central processors, the Advanced Micro Devices Corporation revealed some details about its intentions in the graphic chips market. Apparently, in the next few years, AMD plans to represent a new graphic architecture once every 1.5–2 years, as well as refuse the services of the main manufacturer Radeon until recently — TSMC.
GPU developers traditionally reveal even less information about their microcircuits than the creators of the CPU and other chips. In this regard, there is nothing surprising in the fact that the key slide about the GPU development of Advanced Micro Devices for the coming years is actually limited to the names of three microarchitectures, as well as the references of four technological processes that will be used to manufacture new GPUs. However, this does not mean that AMD plans are unclear at least in general terms.
Globalfoundries for CPU and GPU
The transition to technological processes with Finfet transistors (14LPP, CLN16FF+) made it possible to significantly increase the transistor budgets of graphic processors, thereby radically raising their performance. When switching to Finfet AMD, one more important step took: TSMC services for the production of GPU: all Polaris and Vega chip are exclusively Globalfoundries. Thus, AMD reduces many years of cooperation with TSMC in the field of discrete GPU, which began in the last century.
The decision to translate the production of graphic processors to Globalfoundries is strategic and dictated by the desire to timely and relatively cheaply embedded the latest generation GPU into processors, which are historically produced by this company. It is precisely due to the fact that VEGA discrete chips are made on Glofo, GPU based on VEGA architecture could be built into the AMD Ryzen Mobile processors (Raven Ridge) — The team that was designing this microcircuit needed to take ready-made optimized for 14LPP GPU blocks from Radeon Technologies Group (GPU team).
In the coming years and other discrete GPU AMD development will also be made by Globalfoundries, and nuclei based on their architecture — Timely integrate into hybrid processors, which is designed to increase the competitiveness of the latter.
The transfer of GPU production to Globalfoundries has a flip side. The fact is that Globalfoundries does not have the historical experience in the production of huge microcircuits, and the 14LPP technological process licensed by Samsung was developed for mobile systems on a crystal of 100–150 mm2, but not «large» GPU whose area exceeds 600 mm2. We do not know how the GF 14LPP+ technology process was finalized and whether the future 7-nm DUV and 7-nm EUV with the sight was created on «Big» GPU (most likely it is — IBM Microelectronics engineers with extensive experience in the production of giant microcircuits took part in their development), but it is very likely that when designing «Big» VEGA (14LPP) AMD was somewhat limited in terms of nucleus size.
Vega, Navi and NEXT GEN
As is already known, in the near future AMD will produce a number of products based on a graphic processor based on VEGA architecture for professional and domestic use. This GPU C 4096 stream processors and performance at 12–13 FP32 TFLOPS will be manufactured at Globalfoundries capacities using the 14LPP technological process, the one that is already used for the manufacture of CPU and GPU. This GPU will not be the only one in the Vega family — A little later, AMD will present at least another processor made by improved 14LPP+technology, modified by Globalfoundries. At present, we do not know when exactly the Vega 14nm+ processors appear and what market segments they will be designed, but, apparently, we are talking about the nearest quarters (the deadline — The first half of 2018).
The next graphic architecture of AMD after VEGA will be NAVI, the GPU based on which will be produced using 7 Nm technology. The transition to a process of 7 nm technology itself will allow AMD to increase the density of transistors by half, reduce energy consumption by 60 % (with the same clock frequency and complexity) or increase the frequency by 30 % (with the same energy consumption and complexity). How exactly the AMD will use the transistor budget, we do not know, but it is logical to expect that the new production technology will make it possible to significantly increase GPU productivity for different market segments. In addition, we remember that in the case of NAVI AMD promised the use of a new type of memory (or new types memory?), which is once again intended to increase performance.
Considering that in the case of NAVI, AMD promised two key innovations: scalability and use of new memory, it is logical to assume that the architectural Navi will resemble GCN, becoming its sixth generation. As for the time of the appearance of NAVI in the market, this is the second half of 2018 (optimistic scenario), when Globalfoundries will be ready to start mass production of microcircuits using the first -generation 7 Nm technology.
Advanced Micro Devices has not yet come up with the name for the follower of Navi in architectural plan, and therefore the GPU based on this architecture is now called NEXT GEN (next generation). We do not know whether this means a departure from the GCN concept, or is it something else to talk about. The only thing that AMD says about the new architecture is that the GPU will be made using the second generation of the 7-nm Globalfoundries technology using EUV. What practical advantages EUV will bring in the second half of 2019 – The first half of 2020, in addition to reducing the time of the production cycle and a possible increase in the output of suitable crystals, is not yet known. Therefore, NEXT GEN should expect the same thing that we expect from all new GPUs in the first place: increasing performance on watts.