Unlike the market for central processors, where a serious struggle unfolded in the past year, the segment of consumer video cards was almost completely given over to the «green» GPU manufacturer. While Nvidia was churning out the Pascal family of GPUs, expanding its lineup of GeForce GTX 10-series accelerators, the Radeon Technologies Group continued to work on Radeon RX Vega adapters. These 3D cards had to fight with the flagship solutions of the «green», although in the end the older Radeon RX Vega 64 was only able to catch up with the GeForce GTX 1080, and the title of «best gaming video card» still belongs to the GeForce GTX 1080 Ti released in the first quarter, but about everything in order.
New game king
The last day of February was marked by the announcement of the GeForce GTX 1080 Ti video card, rumors about which have been circulating on the Web since the summer before last.
The novelty received a stripped-down version of the Nvidia GP102 GPU, already familiar from the 2016 Titan X accelerator, with 3584 CUDA cores and equipped with 11 GB of GDDR5X memory with a 352-bit interface. By the way, the main difference between the novelty and the Titan was the absence of one video memory chip and, as a result, the bus reduced by 32 bits.
Quite expectedly, the GeForce GTX 1080 Ti significantly raised the performance bar and turned out to be about 25–45% faster than the GeForce GTX 1080 that debuted a year earlier. Among the models available for sale from Nvidia’s AIB partners, we highlight products from EVGA and Zotac. The first, with the participation of well-known overclockers Vince «K|ngp|n» Lucido and Ilya «TiN» Tsemenko, prepared the GeForce GTX 1080 Ti K|ngp|n Edition, addressed to computer enthusiasts and lovers of sports benchmarking.
This card received a 13-phase power system, a massive three-fan cooler, as well as three VGA-UEFI chips, one of which contains settings for extreme bench sessions using liquid nitrogen. By the way, Vince started the search for the extreme operating modes of the GeForce GTX 1080 Ti on video cards of the reference design. Then he managed to overclock the 16-nm Nvidia GP102 graphics processor, with 12 billion transistors, to 3024 MHz.
Meanwhile, the Hong Kong company Zotac decided to go the other way and released the GeForce GTX 1080 Ti Mini card, which is very compact in size. With a length of only 211mm, this model is one of the highest performance solutions for small form factor systems.
Immediately after the announcement of its new gaming flagship, Nvidia announced the release of updated GeForce GTX 1080 and GTX 1060 graphics cards, equipped with faster memory. The first received GDDR5X chips capable of operating stably at an effective frequency of 11 GHz, and the second received GDDR5 chips with an effective frequency of 9 GHz.
Unfortunately, these modifications are not widely used and the vast majority of solutions available on the market are still equipped with «slow» memory.
Titan X is dead, long live Titan Xp!
One of the most interesting events in the first half of April was the unexpected announcement of the Nvidia Titan Xp semi-professional accelerator. Unlike Titan X on the Pascal architecture, which debuted in 2016, the novelty received a fully functional version of the GP102 chip with 3840 CUDA cores, thanks to which its raw computing power level crossed the mark of 12 Tflops.
Photo by Geeks3D
The memory subsystem also received a slight upgrade and is represented by 12 GDDR5X chips with an effective frequency of 11.4 GHz, communicating with the GPU via a 384-bit bus, which provides a throughput of 548 GB / s. Like its predecessor, Titan Xp is positioned as a low-precision professional computing solution, however, nothing prevents it from being used as a gaming graphics card.
Titan Xp Collector’s Edition Star Wars Jedi Order
According to the results of gaming benchmarks, the Nvidia Titan Xp is on average 10% faster than the GeForce GTX 1080 Ti in the reference version. Whether a small increase in frame rate is worth the extra $500 is still a question. Recall that the recommended cost of the GeForce GTX 1080 Ti video card is $700, and the Titan Xp accelerator was priced at $1200 by Nvidia.
In the absence of competition in the high-end video card segment, AMD is doing its best to maintain market share in the mid-range adapter market. This year, the Sunnyvale chipmaker sent a line of Radeon RX 500-series cards based on advanced Polaris GPUs to fight the younger Pascals this year.
From the point of view of the GPU configuration, the new video accelerators turned out to be completely identical to those of the Radeon RX 400 series. The only exception is the ultra-budget Radeon RX 550 video adapter, which received a Polaris 12 (Lexa Pro) crystal with 512 stream processors. However, chip design optimization along with the “running in” of the 14nm FinFET process technology was not in vain, and the Radeon RX 500 series cards received a noticeable increase in GPU operating frequencies, albeit at the cost of increased power consumption.
As for the flagship of the new family of graphics accelerators in the face of the Radeon RX 580, it can be called a worthy alternative to the GeForce GTX 1060 6GB video cards. The advantages of the AMD solution include flexibility in choosing the amount of video memory (4 or 8 GB), the ability to create CrossFireX graphic bundles (recall, cards younger than the GeForce GTX 1070 do not support multi-GPU configurations), as well as often higher frame rates in games, using the DirectX 12 and Vulkan APIs.
Unfortunately, AMD Radeon RX 500-series graphics cards were not destined to find great popularity among gamers. Even before their release, in early April, the first symptoms of another “cryptocurrency fever” began to appear, caused by a sharp increase in the rate of a number of electronic money. In early June, enterprising miners began to sweep almost all models of mid-range and high-end graphics adapters from store shelves, causing a shortage and, as a result, an increase in retail prices for video cards.
In addition to AMD accelerators, which are traditionally in demand among cryptocurrency miners, solutions based on Nvidia chips were also sold out, starting with the GeForce GTX 1060 3GB models. This excitement bypassed only entry-level adapters like the Radeon RX 460 and GeForce GTX 1050, which do not have a high hash rate.
The sharp increase in the popularity of cryptocurrency mining significantly improved the “financial health” of many video card manufacturers and led them to the idea of producing specialized accelerators. Big companies like ASUS, Gigabyte, and Sapphire started offering stripped-down versions of their cards to potential customers in mid-summer, labeled «Mining Edition».
Sapphire Radeon RX 470 Mining Edition
New products are mostly devoid of video outputs, which had a positive impact on the final cost, but put an end to their possible use as part of gaming PCs.
It is fair to say that at the beginning of autumn the retail price of video cards began to gradually decrease to its recommended level, although it never reached it.
The recurrence of «cryptocurrencies», over and over again devastating store shelves, continues to spoil the mood of gamers who are trying to update their gaming system in an already difficult time.
In the absence of worthy competition in the consumer video accelerator market, Nvidia behaves rather inertly and is in no hurry to retire the Pascal “family” adapters that are in high demand. The same cannot be said about the rapidly developing segment of hardware for the needs of artificial intelligence. Here the chipmaker feels like a fish in water, and strives to provide his customers with the most productive hardware.
In early May, the California company introduced the first accelerator based on the Volta graphics architecture called the Tesla V100. It is based on the 815 sq. mm, numbering over 21 billion transistors. This chip includes 5376 CUDA cores (of which 5120 are active in the Tesla V100) and 640 Tensor cores designed specifically for the needs of machine learning. On the same substrate with the GPU, there are four HBM2 multi-layer memory stacks with a capacity of 4 GB each, characterized by a throughput of about 900 GB / s.
The debut of the Nvidia Titan V graphics accelerator in early December can be called quite unexpected. Similar to the aforementioned Tesla V100, the new adapter is based on the GV100 die in a configuration with 5120 CUDA cores, but only has 12 GB of HBM2 multi-layer memory.
As befits prosumer accelerators of the Titan family, this video card can be used not only in professional computing, but also in games. True, the expediency of such an application of Nvidia Titan V is under big question due to the high recommended price of $ 3,000.
Nvidia GV100 GPU. Photo GamersNexus
As for the GeForce GTX line of video cards, if you believe the information floating around the Web, Volta GPUs will not be used in consumer devices at all — Nvidia is developing the Ampere graphics architecture specifically for them. Be that as it may, in order to find out for sure, you will have to wait for comments from the representatives of the “green” chipmaker themselves.
Through the efforts of Raja Koduri, now the former head of the Radeon Technologies Group, in August, the Radeon RX Vega 64 and Radeon RX Vega 56 video cards, expected by many supporters of the “red camp”, saw the light of day. power, amounting to almost 14 Tflops, along with high power consumption.
The key features of the Radeon RX Vega cards were the use of HBM2 multi-layer memory along with an impressive number of stream processors: the RX Vega 64 model operates 4096 units at once. Compared to the solutions of the Radeon R9 Fury line, there was a slight downgrade in the memory subsystem of the new video cards: the high cost of HBM2 chips forced AMD engineers to limit themselves to only two 4-gigabyte stacks that provide a memory bandwidth (PSP) of 484 GB / s. Whereas the four HBM chips used in the Radeon R9 Fury X are characterized by a memory bandwidth of 512 GB / s.
Nevertheless, despite the completely immodest «appetites», comparable to overclocked video cards based on the Nvidia GP102 chip, the GeForce GTX 1080 and GTX 1070 adapters using the GP104 chip became the main competitor of the Radeon RX Vega graphics accelerators. In general, AMD’s new products have established themselves as a possible analogue of Nvidia’s pre-top solutions, but at the same time, the real price / performance ratio was not in favor of the «reds». The expensive and difficult-to-manufacture Vega 10 graphics processor has negatively impacted the availability of new graphics cards, and retail prices for the reference Radeon RX Vega 64 in overseas stores are approaching $ 700 instead of the recommended $ 500.
Sapphire Radeon RX Vega 64 Nitro+
Attention also deserves the situation with the presence, or rather the lack of sales of Radeon RX Vega cards performed by AIB partners of the chipmaker from Sunnyvale. Even a few months after the debut of the reference design models, many manufacturers are still not ready to start delivering their own variants of video cards, samples of which have already been in the test labs of selected web publications. Of course, this does not apply to all vendors, but even those Radeon RX Vega models of alternative design that are available in foreign online stores can be safely classified as exclusive products.
GeForce GTX 1070 Ti
In the fight for a potential buyer, Nvidia continues to expand its GeForce GTX 10-series graphics card lineup, offering solutions for gamers with any budget. Taking into account the level of performance of the Radeon RX Vega 56, located between the GeForce GTX 1070 and GTX 1080, the company decided to launch its own adapter on the market with similar performance. They became the video card GeForce GTX 1070 Ti, presented at the end of October.
The novelty is based on the already familiar Nvidia GP104 graphics processor, the first 3D cards based on which debuted more than a year and a half ago. The GeForce GTX 1070 Ti operates with 2432 CUDA cores, 152 texture units and 64 render units, and the memory subsystem includes 8 GB GDDR5 with a 256-bit bus and an effective frequency of 8 GHz.
Interestingly, the GPU operating frequencies of absolutely all GeForce GTX 1070 Ti models are reduced to a common denominator and range from 1607-1683 MHz. Nevertheless, the lack of factory overclocking in no way upset the avid fans of free megahertz and can be easily fixed with the help of specialized utilities. According to the test results in our test lab, the GeForce GTX 1070 Ti was about 5-7% slower than its «big sister» in the face of the GeForce GTX 1080.
Among other interesting events in the discrete graphics segment, we should highlight the release of the super-budget GeForce GT 1030 video card, as well as the new version of the professional Radeon Pro Duo accelerator.
In mid-May, Nvidia announced the release of the youngest representative of the «family» of Pascal video cards — the GeForce GT 1030. This card is designed to replace the aging GeForce GT 730, is equipped with two gigabytes of GDDR5 memory with a 64-bit bus and is built on a 14-nanometer GP108 crystal, in which includes 384 CUDA cores, 40 texture units and 32 render units.
According to the test results, the new product turned out to be a good solution for the low-end segment of video cards, which can be recommended to gamers with a limited budget who prefer undemanding online games.
In the new version of the Radeon Pro Duo accelerator, AMD engineers decided to use a pair of 14-nm Polaris 10 crystals. Each GPU operates 2304 stream processors and communicates with 16 gigabytes of GDDR5 memory via a 256-bit bus.
Compared to the previous version, which uses two Fiji XT crystals with HBM memory, the new product has a moderate TDP of about 250 W, which made it possible to use a relatively simple air cooler. Recall that the original Radeon Pro Duo, due to the huge power consumption (about 375 W), was equipped with a maintenance-free CBO.
Unfortunately, for gamers, the updated Radeon Pro Duo is not of particular interest, both because of the presence of two graphics processors and related shortcomings, and because of the high cost — the recommended cost of the device is $ 1,000. The target audience of the novelty are professionals working with CAD applications, video content and 3D graphics.
In the next part, we will look at other segments of the computer market and note the main trends that have emerged this year.
Overview of the main events of 2017. Processors and platforms
Overview of the main events of 2017. Memory, overclocking and new trends