Results of 2017: Graphics Accelerators / Video Cards

In 2016, a new era in the history of discrete GPUs began – both manufacturers of graphics processors for desktops and laptops began manufacturing chips with a progressive process technology 14/16 nm FinFET, which allowed to resume the growth of performance and energy efficiency, slowed down during a long period of rule 28 nm. In the next year, 2017, the capabilities of the new technology have already been fully revealed – both NVIDIA and AMD released the first-tier GPUs previously reserved for the compute accelerators to the mass market. The most attention was attracted exactly to AMD, because the company has finally restored its presence in the segment of high-performance game video cards due to the Radeon RX Vega family. On the other hand, NVIDIA has already paved the way for the Volta architecture chips, which are manufactured on the basis of the next photolithography node using FinFET – 12 nm.

It's time to take the subtotals in the field of discrete graphics cards for desktop computers and list the events that marked this rather fruitful year.

The RX 500

The RX 500 The RX 500

The Radeon accelerators with the indices RX 460 RX 470 and RX 480 came in 2016 to replace the obsolete GPU of the 28 nm that managed to survive one, or even two rounds of rebranding, in the initial and average performance segments. Due to attractive prices, AMD put serious pressure on the competing devices of the GeForce 10 series, but from a technical point of view, the Polaris family did not live up to its expectations, because the competitor had much better disposed of the opportunities that the new 14/16 nm FinFET process technology opened to increase energy efficiency and frequency potential of the GPU.

The next year after the debut of the 400 series Radeon there could be no question of AMD suddenly filling all price niches with graphics processors Vega or Navi, so the Polaris chips were destined to be reborn in the Radeon 500 family. the lifetime of the Graphics Core Next Next architecture has used this technique once when the next generation GPU is not yet ready for mass production and the "old" silicon is transferred to the next series of video adapters. But Polaris is a special case. Instead of simple re-labeling, AMD released updated Polaris 20 and Polaris 21 crystals, which optimized the circuitry and production process.

The GPU revision in the Radeon 500 series significantly influenced the performance of the older models of the family. Thus, the standard models Radeon RX 570 and RX 580 in the regular mode reach the same frequencies as the top modifications of their predecessors – RX 470 and RX 480 – with selected crystals and advanced cooling system. As a result, the Radeon RX 580 competes on an equal footing with GeForce GTX 1060 6 GB for gaming performance, and the RX 570 takes its place in the confrontation with GTX 1060 3 GB due to the larger amount of RAM. Alas, the victory of Radeon RX 570 and RX 580 in benchmarks does not compensate for the situation with the prices of video cards on the older chip Polaris. Crypto-currency fever keeps them at a significantly higher level than those recommended by AMD retail values.

SAPPHIRE Pulse ITX Radeon RX 570

As for the younger models of the family – Radeon RX 550 and RX 560 AMD focused more on energy efficiency than speed. In 2016, when the Radeon RX 460 went on sale, AMD had to compromise: let the accelerator yield in speed to its rivals ( GeForce GTX 1050 and GTX 1050 Ti ), but it is cheaper, which is important in this price categories. But the Polaris 11 GPU remained in the reserve part of the logic that was not involved in the Radeon RX 460, and under the name Polaris 21 the chip returned as part of the Radeon RX 560 with a fully active set of computing units.

Suppose that in the older models of the 500 series AMD completely invested the released resource of energy efficiency in increasing the frequencies, but the specifications of the Radeon RX 560 allow to produce video cards with the TBP (Typical Board Power) from 60 W, and some manufacturers boldly dropped to 45 W , because the expanded conveyor Polaris 21 to some extent compensates for the limited frequency range. On the other hand, the powerful versions of the Radeon RX 560, reaching 80 W, the updated GPU helps in the fight against the GeForce GTX 1050 and GTX 1050 Ti. Be that as it may, in this niche of the video card AMD still lose to NVIDIA devices on speed and attract the buyer at the expense of relatively low prices.

Due to a wide range of power and RAM (2 or 4 GB) under the same brand Radeon RX 560 there are devices with completely different levels of performance. But not only that, AMD allows the use of RX 560 graphics processors with "reduced" configurations of computing units – as in Radeon RX 460. As a result, RX 560 from all video cards in the modern catalog of AMD and NVIDIA has the most diffuse specifications and requires special attention when buying.


The last representative of the 500 series – Radeon RX 550 brought to the desktop desktop GPU brand new graphics processor, which expanded the Polaris family towards the most compact and energy-efficient solutions. In its price niche Radeon RX 550 replaced the retail model Radeon R7 350 and several OEM-items of the lower level from the 300th and 400th series. For high performance in AAA-games such cards do not pretend, instead the RX 550 trumpet supports modern display interfaces (DisplayPort 1.3 / 1.4 and HDMI 2.0) and hardware decoding of high-definition video. However, at its cost, the Radeon RX 550 is also good for games (not the least role in this is the adequate amount of local memory – 2 or even 4 GB), in particular – undemanding multiplayer projects


The Vega The

AMD initially made it clear that the Polaris family of graphics processors was only an intermediate step on the way to Vega – a truly massive upgrade of the Graphics Core Next architecture that will allow the company to return to the GPU market for enthusiasts. And before it finally happened, the GPU Vega 10 was tested in the Radeon Instinct accelerator and the Radeon RX Vega Frontier Edition.

In the consumer market, the Radeon RX Vega family is represented by two models with indices 56 and 64 the latter of which exists in air-cooled or liquid-cooled configurations. Chip Vega 10 – the only offspring of Vega architecture at the moment – is the ideological successor of the Fiji graphics processor Radeon R9 Fury and Fury X ). These GPUs have a similar configuration of the computing units, and both are equipped with high-speed memory of the HBM type. Vega has the honor to bring to the market second generation consumer HBM video cards and the technology of page access to address space that allows to use the local GPU memory as a third level cache and access the "far" memory (system RAM) at the driver level – just as it is arranged in the central processors since the time of Intel 80286.

While NVIDIA, since the Pascal generation, has allocated a GPU for non-graphical computing (GP100) to a special category, the GCN 1.4 architecture underlying Vega, like the previous iterations of Graphics Core Next, is both rendering oriented simultaneously, and for general purpose computing (GP-GPU). Compared with Polaris, Vega's computing capabilities are expanded by "sealing" operations of reduced bit capacity (Packed Math, in AMD terminology): the GPU can simultaneously perform twice as many operations on FP16 format numbers as compared to FP32 operations. Let the mathematics of reduced accuracy in discrete GPUs is still in demand mainly in the tasks of machine learning, but in time it can find wide application in graphics at those stages of the rendering pipeline, which do not require a more accurate representation. The benefit of floating-point operations over 10- and 16-bit numbers is supported by the dominant gaming API – DirectX 12.

Let AMD's chips stay focused on the design load, the Vega creators did not ignore the fixed functionality blocks related to 3D rendering. Vega acquired a completely new, partially programmable geometry engine and a separate scheduler for operations on geometric primitives, and relatively small for such a large GPU, the number of texture overlay units compensates for tile rendering (this method is used by NVIDIA in the Maxwell and Pascal architectures). Finally, Vega boasts the most complete support for DirectX 12 rendering functions "in hardware" among discrete GPUs. Vega capabilities not only correspond to the highest feature level in Direct3D 12 specifications, but also within the framework of individual functions that imply a more detailed separation, Vega outperforms Pascal and Volta architectures from NVIDIA.

AMD Radeon RX Vega 64 Limited Edition

In general, AMD has created a balanced, versatile and in many ways innovative product. Alas, many of the innovations of Vega 10 (as it has already happened with AMD chips) represent an investment in the future, instead of being useful here and now, or used only by professional software. According to rough estimates – crystal size, clock speeds and power consumption – the GPU Vega 10 is a full rival GP102 from NVIDIA ( GeForce GTX 1080 Ti TITAN X and TITAN Xp), and these expectations fully justified in the tests of general-purpose computing. At the same time, in games, the desire of developers to get a universal chip, suitable for both graphic and computational tasks, only burdens Vega: the new flagship of AMD can compete only with the frame rate in AAA projects with GeForce GTX 1080 , and the performance level of the GeForce GTX 1080 Ti in most cases is out of the question.

Another disadvantage of Vega is its extremely high power consumption. The new architecture can operate at much higher frequencies than Polaris, but in terms of performance per watt in gaming applications Vega 10 with a devastating account lost to Pascal chips of comparable performance. If at the time of Fiji AMD almost caught up with NVIDIA in the energy efficiency race, then today the gap in favor of an opponent is greater than ever.

Compared with Radeon RX Vega 64, the younger model of the series – Vega 56 – was more convincing. In the games, she won a confident victory over GeForce GTX 1070 and demonstrated a remarkable overclocking potential, due to which it easily reaches the level of Vega 64 and is threateningly matched to GeForce GTX 1080. GeForce GTX 1070 is not capable of such achievements, although NVIDIA already Vega 56 with the help of the newest representative of the GeForce 10 series – GTX 1070 Ti .

The AMD Vega 10 graphics processor

Be that as it may, we welcome the return of the Radeon brand to the high-performance graphics market. The main problem of AMD now is not due to the speed and power consumption of the GPU, but to the lack of video cards of the Radeon RX Vega series in the market that spurs prices and pushes Vega 56 and Vega 64 into a price category where comparison with NVIDIA products ends not in favor of the "reds" ". Sooner or later, Vega crystal shipments will reach the proper level, and non-reference design models have already appeared, but because of the deficit, AMD continues to lose revenue, so necessary in order to overcome the growing technological backwardness from NVIDIA in the next generation.

# NVIDIA GeForce GTX 1080 Ti

The debut of the Vega family from AMD did not shake the position GeForce GTX 1080 Ti as the leader of high-performance gaming accelerators. Given the fact that Vega has not claimed a faster graphics rendering performance than the GeForce GTX 1080 since the promulgation of the GPU specifications, NVIDIA did not even have an urgent need to update the GeForce 10 series lineup so early, but as a result, the new flagship GTX 1080 Ti , was released in March last year.

Before entering the consumer market, the GP102 GPU announced itself as part of Tesla's "countable" accelerators and a request for the Titan X graphics card. This is the standard tactic of NVIDIA: the top position in the new line of gaming graphics cards is taken up by the device on a chip for a while the second echelon (in this case it is the GeForce GTX 1080), and the flagship GPU first removes the cream in the professional category and only then, with the loss of computing units and functions, reaches the gaming devices.

NVIDIA GeForce GTX 1080 Ti Founders Edition

The GeForce GTX 1080 Ti actually represents a TITAN X at a discount. In comparison with the "Titan" game graphics card missed only one controller RAM, eight ROP and 1 GB of local memory. In games, these losses are completely compensated by the increased frequencies of GPU and RAM. Even the GP102's computational capabilities in the GTX 1080 Ti are fully preserved: the graphics card can aggregate four operations on int8 data in one clock cycle of a 32-bit CUDA core – just like TITAN X, which is positioned as an accelerator for deep learning tasks.

With the release of the GeFoce GTX 1080 Ti, NVIDIA has completed the transition from the 28 nm GPU to the new generation, since the difference in the game speed between the new flagship and its predecessor, the GeForce GTX 980 Ti, is a convincing 54-67%. The GTX 1080 Ti was the first among consumer graphics cards, which is unconditionally suitable for playing in 4K resolution. This achievement was claimed by the GeForce GTX 780 Ti and Radeon R9 290X, but before the advent of the GTX 1080 Ti, the task was only for TITAN X or the GTX 1070/1080 in SLI mode.

# NVIDIA GeForce GTX 1070 Ti

In the family of Radeon RX Vega accelerators, it was the younger model, Vega 56, that proved to be the most dangerous product for NVIDIA. In opposition to the GeForce GTX 1080 and Vega 64, we can state the equality of forces (and even with a lot of reservations), but this is all that AMD could achieve without violating the reasonable limits of power consumption of a single video card. On the contrary, due to the fact that the distance between the GTX 1080 and the next highest NVIDIA card is a long distance, Vega 56 dealt a sniper strike on the GeForce GTX 1070. The AMD accelerator not only outperforms the GTX 1070 in terms of speed at nominal frequencies, but also easily accelerates to the level GeForce GTX 1080. But the same reason (significantly weakened in comparison with the GTX 1080 GTX 1070 configuration) enabled NVIDIA to quickly close the gap in its ranks by releasing GeForce GTX 1070 Ti .

The "titanium" version is an extremely generous upgrade of the GeForce GTX 1070. The GTX 1070 Ti is equated with the GTX 1080 for power reserve and base frequency, and as part of the NVIDIA GPU, it activated additional computational units, which in total raised the design computing power of the chip to 27%. In fact, the main factor that differentiates the GTX 1070 Ti and GTX 1080 is the memory – GDDR5 or GDDR5X – with the associated clock type.

NVIDIA GeForce GTX 1070 Ti Founders Edition

GeForce GTX 1070 Ti achieved its goal – at the nominal frequencies to provide a practically insignificant, but still indisputable advantage over the competitor's video card. At the same time, the GTX 1070 Ti proved to be no less promising for overclocking than Vega 56. In fact, due to the overclocking potential, the GeForce GTX 1070 Ti undermines the positions of the more expensive model of its own family – if Vega 56 at higher frequencies only moves to the level of the GeForce GTX 1080, then the updated GTX 1070 Ti easily outperforms it.


The appearance of the GeForce GTX 1080 Ti totally devalued TITAN X – these video cards are almost equal in speed, so the release of the updated version of the pre-accelerator accelerator, TITAN Xp, was only a matter of time. However, let TITAN Xp have a fully functional GP102 GPU, the performance gain compared to the GeForce GTX 1080 Ti due to unlocked computing units and an additional gigabyte of RAM was rather modest.

However, TITAN Xp has not recently claimed the exclusive status that TITAN X once enjoyed. Only a year after the first Titan, Pascal architecture saw the light of TITAN V, which already uses the GPU of the Volta family. The peculiarity of the situation is that TITAN V got the oldest and so far the only Volta-GV100 chip, while the higher-level Pascal (GP100) did not reach the request of the graphics cards, not to mention the gaming ones.

The GV100 is an extremely ambitious project. The crystal is characterized by a record area and the number of transistors (21.1 billion). Фактически это микросхема с наибольшим транзисторным бюджетом среди всех когда-либо выпущенных на сегодняшний день, считая гигантские FPGA и многокристальные процессоры AMD EPIC/Threadripper. Этому достижению поспособствовал техпроцесс 12 нм FinFET, на который NVIDIA переключилась с проверенного 16 нм FinFET.

Хотя GPU в составе TITAN V не является полностью функциональным (часть вычислительных блоков и одна из сборок памяти HBM2 была отключена), все равно TITAN V располагает на 33% большим набором шейдерных ALU по сравнению с TITAN Xp. Но игры и 3D-рендеринг как таковой не являются основной сферой применения TITAN V. Главное нововведение архитектуры Volta — специализированные ядра для тензорных операций (произведение матриц А и B складывается с матрицей C), которые используются главным образом в обработке данных сетями машинного обучения (inference). Кроме того, NVIDIA существенно модифицировала организацию стандартных ядер CUDA с целью разместить на площади GPU как можно больше вычислительных ресурсов.

Графический процессор NVIDIA V100 на плате Tesla V100

В расчетных задачах TITAN V полностью оправдывает свою беспрецедентно высокую цену — $3 000. С другой стороны, в рендеринге (в частности, в играх) возможности GV100 пока под вопросом: по сравнению с TITAN Xp прирост производительности с текущей версией драйвера явно не соответствует как технологической, так и ценовой разнице двух видеокарт.

#Прогноз на будущее

Подрядчики по выпуску микросхем, с которыми сотрудничают AMD и NVIDIA, — TSMC и GlobalFoundries — обещают безболезненный переход от нормы 14/16 нм FinFET к следующему узлу, 12 нм. Пользуясь официальными заявлениями производителей, равно как и неподтвержденными слухами, можно с уверенностью утверждать, что в 2018 году оба разработчика дискретных GPU обновят ассортимент игровых видеокарт за счет чипов нового поколения.

AMD озвучила свои намерения предельно ясно — на базе технологии 12 нм FinFET грядет обновление флагманского графического процессора Vega 10. Действительно, несколько серверных ускорителей с GPU Vega 20 и объемом памяти 16 либо 32 Гбайт уже зарегистрированы в базе устройств Евразийской экономической комиссии. Ходит довольно много слухов о GPU под названием Vega 11 с памятью HBM2, а соответствующую видеокарту прочат в соперники GeForce GTX 1060. AMD поспешила заявить, что Vega 11 является не дискретным чипом, но составной частью грядущих APU с процессорным ядром Zen, однако база ЕЭК содержит данные и о двух потребительских устройствах на основе двух конфигураций данного GPU — Vega 11 XT и Vega 11 PRO. Там же фигурирует и следующий по старшинству графический процессор, Vega 12.

Совершенно ничего не известно о модельном ряде графических процессоров Navi — следующей после Vega итерации GCN. Неофициальные источники поместили грядущий анонс новой архитектуры AMD на август 2018 года. Говорят, что чипы Navi будут производиться по норме 7 нм, и уже в этом поколении AMD может отказаться от выпуска монолитных GPU высшей категории производительности в пользу модульных конструкций из нескольких ядер, объединенных кремниевой подложкой.

В планах NVIDIA официальным наследником архитектуры Pascal является Volta, и недавний выход TITAN V указывает на то, что Volta повторит траекторию развития семейства Pascal: сначала компания выпустит флагманский игровой ускоритель на основе GPU второго эшелона, затем Volta проникнет в среднюю и начальную ценовую категории, и наконец появится видеокарта с большим GPU, которым станет либо GV100, либо его аналог, лишенный специфических вычислительных функций.

На текущий момент Volta является последней вехой в обнародованном NVIDIA манифесте развития GPU, но интернет полнится слухами о новой архитектуре под названием Ampere. А вот что именно скрывается под этим именем, пока совершенно неизвестно. У NVIDIA есть два пути — либо Ampere станет следующей универсальной платформой после Volta, которая найдет применение и в ускорителях вычислений, и в игровых видеокартах, либо NVIDIA планирует зарезервировать Volta для профессиональных продуктов, а грядущее поколение GeForce будет сформировано именно процессорами Ampere. В качестве возможного атрибута GPU Ampere называют оперативную память типа GDDR6, которая отличается скоростью передачи данных вплоть до 16 Гбит/с на контакт и повышенной энергоэффективностью по сравнению с GDDR5X.

Кроме того, NVIDIA тоже изучает перспективы GPU, собранных из нескольких дискретных ядер на кремниевой подложке. Похоже, что это неизбежный путь развития для графических процессоров, но коммерческой стадии многочиповые GPU вряд ли достигнут раньше следующего, 2019 года.

#Приложение. Актуальные линейки видеоадаптеров AMD и NVIDIA для настольных ПК

Производитель AMD
Модель Radeon RX 550 Radeon RX 560 Radeon RX 570 Radeon RX 580 Radeon RX Vega 56 Radeon RX Vega 64 Radeon RX Vega 64 Liquid Cooled Edition
Графический процессор
Название Polaris 12 Polaris 21 PRO/XT Polaris 20 XL Polaris 20 XTX Vega 10 XL Vega 10 XT Vega 10 XT
Микроархитектура GCN 1.3 GCN 1.3 GCN 1.3 GCN 1.3 GCN 1.4 GCN 1.4 GCN 1.4
Техпроцесс, нм 14 нм FinFET 14 нм FinFET 14 нм FinFET 14 нм FinFET 14 нм FinFET 14 нм FinFET 14 нм FinFET
Число транзисторов, млн 2200 3 000 5700 12 500 12 500 12 500
Тактовая частота, МГц: Base Clock / Boost Clock 1100/1183 1175/1275 1168/1244 1257/1340 1156/1471 1247/1546 1406/1677
Число шейдерных ALU 512 896/1024 2048 2304 3584 4096 4096
Число блоков наложения текстур 32 56/64 128 144 256 256 256
Число ROP 16 16 32 32 64 64 64
Оперативная память
Разрядность шины, бит 128 128 256 256 2048 2048 2048
Тактовая частота, МГц (пропускная способность на контакт, Мбит/с) 1750 (7000) 1750 (7000) 1750 (7000) 2000 (8000) 800 (1600) 945 (1890) 945 (1890)
Объем, Мбайт 2048/4096 2048/4096 4096/8192 4096/8192 8096 8096 8096
Шина ввода/вывода PCI Express 3.0 x8 PCI Express 3.0 x8 PCI Express 3.0 x16 PCI Express 3.0 x16 PCI Express 3.0 x16 PCI Express 3.0 x16 PCI Express 3.0 x16
Пиковая производительность FP32, GFLOPS (из расчета максимальной указанной частоты) 1211 2611/2284 5095 6175 10544 12665 13738
Производительность FP32/FP64 1/16 1/16 1/16 1/16 1/16 1/16 1/16
Пропускная способность оперативной памяти, Гбайт/с 112 112 224 256 410 484 484
Вывод изображения
Интерфейсы вывода изображения DL DVI, HDMI 2.0b, DisplayPort 1.3/1.4 DL DVI, HDMI 2.0b, DisplayPort 1.3/1.4 DL DVI, HDMI 2.0b, DisplayPort 1.3/1.4 DL DVI, HDMI 2.0b, DisplayPort 1.3/1.4 DL DVI, HDMI 2.0b, DisplayPort 1.3/1.4 DL DVI, HDMI 2.0b, DisplayPort 1.3/1.4 HDMI 2.0, DisplayPort 1.4
TDP/TBP, Вт 50 60–80 150 185 210 295 345
Рекомендованная розничная цена (США, без налога), $ 79 99 (4 Гбайт) 169 (4 Гбайт) 199/229 399 499 699
Рекомендованная розничная цена (Россия), руб. 5 369 НД 11 299 (4 Гбайт) 13 449 / 15 299 НД НД НД
Розничная цена на (без налога), $ От 85/109 От 124 От 240/284 От 259/299 От 649 От 699 От 1099
Розничная цена на (без налога), руб. От 5 990/7 437 От 7 500/8 608 От 19 950 / 23 270 От 22 160 / 24 100 От 57 890 От 58 690 От 67 488
Производитель NVIDIA
Модель GeForce GT 1030 GeForce GTX 1050 GeForce GTX 1050 Ti GeForce GTX 1060 3 Гбайт GeForce GTX 1060 6 Гбайт
Графический процессор
Название GP108 GP107 GP107 GP106 GP106
Микроархитектура Pascal Pascal Pascal Pascal Pascal
Техпроцесс, нм 14 нм FinFET 14 нм FinFET 14 нм FinFET 16 нм FinFET 16 нм FinFET
Число транзисторов, млн 1800 3300 3300 4400 4400
Тактовая частота, МГц: Base Clock / Boost Clock 1227/1468 1354/1455 1290/1392 1506/1708 1506/1708
Число шейдерных ALU 384 640 768 1152 1280
Число блоков наложения текстур 24 40 48 72 80
Число ROP 16 32 32 48 48
Оперативная память
Разрядность шины, бит 64 128 128 192 192
Тактовая частота, МГц (пропускная способность на контакт, Мбит/с) 1500(6000) 1750 (7000) 1750 (7000) 2000 (8000) 2250 (9000) 2000 (8000) 2250 (9000)
Объем, Мбайт 2 048 2 048 4 096 3 072 6 144
Шина ввода/вывода PCI Express 3.0 x4 PCI Express 3.0 x16 PCI Express 3.0 x16 PCI Express 3.0 x16 PCI Express 3.0 x16
Пиковая производительность FP32, GFLOPS (из расчета максимальной указанной частоты) 1127 1862 2138 3935 4372
Производительность FP32/FP64 1/32 1/32 1/32 1/32 1/32
Пропускная способность оперативной памяти, Гбайт/с 192/216 192/216 192/216 192/216 192/216
Вывод изображения
Интерфейсы вывода изображения DL DVI, HDMI 2.0b, DisplayPort 1.3/1.4 DL DVI-D, DisplayPort 1.3/1.4, HDMI 2.0b DL DVI-D, DisplayPort 1.3/1.4, HDMI 2.0b DL DVI-D, DisplayPort 1.3/1.4, HDMI 2.0b DL DVI-D, DisplayPort 1.3/1.4, HDMI 2.0b
TDP/TBP, Вт 30 75 75 120 120
Рекомендованная розничная цена (США, без налога), $ 80 109 139 199 249/299 (Founders Edition)
Рекомендованная розничная цена (Россия), руб. НД 8 490 10 490 НД 22 990 (Founders Edition)
Розничная цена на (без налога), $ От 67 От 110 От 151 От 220 От 270
Розничная цена на (без налога), руб. От 4 578 От 7 760 От 9 949 От 16 800 От 19 800
Модель GeForce GTX 1070 GeForce GTX 1070 Ti GeForce GTX 1080 GeForce GTX 1080 Ti TITAN Xp TITAN V
Графический процессор
Название GP104 GP104 GP104 GP102 GP102 GV100
Микроархитектура Pascal Pascal Pascal Pascal Pascal Volta
Техпроцесс, нм 16 нм FinFET 16 нм FinFET 16 нм FinFET 16 нм FinFET 16 нм FinFET 12 нм FinFET
Число транзисторов, млн 7200 7 200 7 200 12 000 12 000 21 100
Тактовая частота, МГц: Base Clock / Boost Clock 1506/1683 1607/1683 1607/1733 1480/1582 1405/1582 1200/1455
Число шейдерных ALU 1920 2432 2560 3584 3840 5120
Число блоков наложения текстур 120 152 160 224 240 320
Число ROP 64 64 64 88 96 96
Оперативная память
Разрядность шины, бит 256 256 256 352 384 3072
Тактовая частота, МГц (пропускная способность на контакт, Мбит/с) 2000 (8000) 2000 (8000) 1250 (10000) 1375 (11000) 1376,25 (11010) 1426,25 (11410) 850 (1700)
Объем, Мбайт 8 192 8 192 8 192 11 264 12 288 12 288
Шина ввода/вывода PCI Express 3.0 x16 PCI Express 3.0 x16 PCI Express 3.0 x16 PCI Express 3.0 x16 PCI Express 3.0 x16 PCI Express 3.0 x16
Пиковая производительность FP32, GFLOPS (из расчета максимальной указанной частоты) 6463 8186 8873 11340 12150 14899
Производительность FP32/FP64 1/16 1/16 1/16 1/16 1/16 1/2
Пропускная способность оперативной памяти, Гбайт/с 256 29/72 320/252 484 548 653
Вывод изображения
Интерфейсы вывода изображения DL DVI-D, DisplayPort 1.3/1.4, HDMI 2.0b DL DVI-D, DisplayPort 1.3/1.4, HDMI 2.0b DL DVI-D, DisplayPort 1.3/1.4, HDMI 2.0b DL DVI-D, DisplayPort 1.3/1.4, HDMI 2.0b DisplayPort 1.3/1.4, HDMI 2.0b DisplayPort 1.3/1.4, HDMI 2.0b
TDP/TBP, Вт 150 180 180 250 250 250
Рекомендованная розничная цена(США, без налога), $ 349/399 (Founders Edition) 449 499/549 (Founders Edition) 699 (Founders Edition) 1 138 2 999
Рекомендованная розничная цена (Россия), руб. 31 590 (Founders Edition) 33 990 45 790 (Founders Edition) 52 990 (Founders Edition) 88 900 221 990
Розничная цена на (без налога), $ От 459 От 459 От 529 От 710 НД НД
Розничная цена на (без налога), руб. От 31 410 От 35 000 От 40 820 От 52 630 НД НД

Если вы заметили ошибку — выделите ее мышью и нажмите CTRL+ENTER.

Leave a Reply