In recent years, we have become accustomed to the fact that AMD invariably occupies a catch-up position in the competition with NVIDIA. The last time GeForce and Radeon-branded gaming cards were equal in the high end was during the GeForce GTX 980 Ti vs. Radeon R9 Fury X. The weakness of AMD’s graphics division and the Graphics Core Next architecture, which has been around for nearly a decade , became especially prominent after NVIDIA released the GeForce RTX series devices. The brainchild of Jensen Huang not only reigns supreme in the high-performance GPU market, but is already setting the course for the entire industry with real-time ray tracing and machine learning technologies.
To AMD’s credit, its products are still in demand in the budget and middle class of gaming video cards, and the models of the Radeon RX Vega family have proven themselves as general-purpose computing accelerators. And now the Reds have another reason to be optimistic: the senior graphics processor of the Vega family has migrated to the progressive 7nm FinFET standard. The updated Vega 20 chip, which was presented to the public in autumn as part of the Radeon Instinct computing accelerators, became the first discrete GPU produced using this technology, while NVIDIA only relatively recently debugged production at the 12 nm standard.
Considering how far ahead AMD is in terms of technology adoption, only the company’s most dedicated supporters could have expected a gaming graphics card based on Vega 20 silicon. But sometimes even the wildest dreams come true, because AMD decided to do just that. Is it a show of strength or a gesture of desperation? Let’s try to find out.
The major innovations that characterize second-generation Vega silicon are directly driven by the move from the 14nm FinFET node to 7nm. Taiwanese TSMC is manufacturing the Vega 20 chips — it seems that AMD had no other choice in this, because GlobalFoundries, which now produces most of the Graphics Core Next architecture graphics chips, curtailed work on its own 7nm line, and Samsung launched it later than Taiwanese.
«Upgrade» of graphics processors due to a new process technology or simply refinement of circuitry for AMD has already become a tradition. Let’s take the family of Polaris chips: the eldest of these GPUs — Polaris 10 — recently experienced its third birth, already using 12 nm technology. However, 14 and 7 nm chips differ from each other much more than 14 and 12 nm. TSMC has never had a 14nm production line, but in order to appreciate the scale of the changes that the Vega GPU has undergone, we can take the performance of the 16nm process in TSMC’s factories as a starting point. Compared to them, the 7 nm standard increases chip density by 3.2 times, and power consumption is reduced by 61%.
Such numbers are unattainable in practice, because not all microcircuits scale so well, especially as complex as modern processors, whether it be CPU or GPU. However, AMD managed to reduce the area of its flagship chip by exactly one and a half times — from 495 to 331 mm.2, despite the fact that it has 0.7 billion more transistors. How this affected clock speeds and power density of the chip, we will discuss on the example of the Radeon VII video card itself, but the miniaturization of the graphics core gave Vega 20 another important advantage.
In high-end accelerators, AMD has long abandoned GDDR memory in favor of HBM chips, which are placed together with the GPU on a silicon substrate (interposer). The latter is actually a reduced microcircuit, in which there are only lines of conductors and no transistors. As a consequence, the substrate is produced on a standard photolithographic conveyor, and its size is limited by a photomask. The Vega 10 image clearly shows that the giant GPU simply did not allow AMD to place more than two HBM2 stacks on the substrate. That is why Vega 10 has half the memory bus width compared to the Fiji chip (the first GPU in which AMD used HBM memory), even if the increased frequencies of the second generation HBM compensate for an involuntary step back.
On the other hand, there is already enough space on the substrate around the Vega 20 for four HBM2 builds. AMD took advantage of this opportunity to simultaneously double the bandwidth (which now reaches an incredible 1 TB / s) and the amount of RAM: Radeon Instinct compute accelerators on Vega 20 chips received up to 32 GB of RAM, and Radeon VII — 16 GB.
Some of the 0.7 billion transistors that appeared in the new version of the processor are concentrated in additional memory controllers, but the GPU architecture has undergone other changes. AMD has worked on pipeline efficiency, both in 3D rendering and general purpose calculations. In particular, the performance of ROP units has been increased so that they can take full advantage of the extended memory bus, and on the part of shader ALUs, the execution speed of FMA (Fused Multiply-Add), perhaps the most important operation for many computing tasks. Otherwise Vega 20 still belongs to the fifth generation of GCN and does not contain changes at the ISA (Instruction Set Architecture) level. The last feature of the chip worth mentioning is the execution of double precision calculations (FP64) at 1/2 of the peak performance in FP32. In theory, any accelerators of the GCN architecture have this ability, but at the design stage of the chip, AMD decides how much the FP64 performance will be limited in accordance with the market positioning of the device. But since the main purpose of Vega 20 is GPGPU (General Purpose GPU) applications, this function has not been cut at the silicon level.
Although the Vega 20 graphics processor matches the set of main computing units from the older first-generation Vega, the Radeon VII did not get a fully functional version of the GPU. So far, the Vega 20 chip in a configuration with 64 active CUs (Vega 20 XT) can only be found on the Radeon Instinct MI60 board, while the Radeon VII uses the same version of the chip as in the Instinct MI50 — Vega 20 XL. It has 60 functional CUs left, which means 3840 shader ALUs (stream processors, in AMD terminology) and 240 texture mapping units. If we compare with the previous generation model range, then the Radeon VII is exactly in the middle between the Radeon RX Vega 56 and Vega 64, which means that AMD still has the opportunity to release a more powerful video card if the company deems it appropriate.
The fact that the Vega 20 graphics processor appeared on the consumer market only in a “stripped down” version is not surprising. The output of suitable microcircuits with such a number of transistors at the rate of 7 nm still leaves much to be desired, and AMD saves fully functional crystals for its server products. On the other hand, there is no reason to be disappointed, because the gaming version of Vega 20 retained 94% of its full processing power. Judging by how Vega 56 and Vega 64 are correlated in games, and they differ much more in terms of hardware, Radeon VII will definitely survive the loss of four Compute Units.
|Model||Radeon RX Vega 56||Radeon RX Vega 64||Radeon RX Vega 64 Liquid Cooled||Radeon VII|
|Name||Vega 10 XL||Vega 10XT||Vega 10 XTX||Vega 20XL|
|microarchitecture||GCN 1.4||GCN 1.4||GCN 1.4||GCN 1.4|
|Process technology, nm||14nm FinFET||14nm FinFET||14nm FinFET||7nm FinFET|
|Number of transistors, million||12 500||12 500||12 500||13 200|
|Clock frequency, MHz: Base Clock / Boost Clock||1156/1471||1247/1546||1406/1677||1400/1750|
|Number of shader ALUs||3584||4096||4096||3840|
|Number of texture overlays||224||256||256||240|
|Number of ROPs||64||64||64||64|
|Bus width, bit||2048||2048||2048||4096|
|Clock frequency, MHz (bandwidth per contact, Mbps)||800 (1600)||945 (1890)||945 (1890)||1000 (2000)|
|I/O bus||PCI Express 3.0 x16||PCI Express 3.0 x16||PCI Express 3.0 x16||PCI Express 3.0 x16|
|Peak performance FP32, GFLOPS (based on maximum specified frequency)||10544||12665||13738||13440|
|RAM bandwidth, GB/s||410||484||484||1024|
|Image output interfaces||HDMI 2.0, DisplayPort 1.4||HDMI 2.0, DisplayPort 1.4||HDMI 2.0, DisplayPort 1.4||HDMI 2.0, DisplayPort 1.4|
|Retail price (USA, without tax), $||From 329 (newegg.com)||From 399 (newegg.com)||From 599 (amazon.com)||699 (recommended at the time of release)|
|Retail price (Russia), rub.||From 23 000 (market.yandex.ru)||From 35 910 (market.yandex.ru)||From 38 650 (market.yandex.ru)||ND|
The transition to the «thin» norm of 7 nm allowed a radical increase in clock frequencies: Vega 20 operating range on a Radeon VII board is 1400-1800 MHz. It is curious that it is the last number that appears in official materials when it comes to the upper frequency available for a video card, while first-generation Vega accelerators are more often characterized by the term Boost Clock (maximum frequency for a typical computing load), and the new Vega is heavier 0.7 billion transistors. It looks like AMD is more confident than ever in the merits of the new process technology.
However, if we ignore the absolute figures and compare Radeon VII with Vega 64 in percentage terms, then AMD managed to overclock the GPU by only 10% at maximum frequency, and the intergenerational difference in the throughput of standard precision shader operations (FP32) and texture mapping speed is 12 and 7%, respectively — not so much even adjusted for the four missing Compute Units. But of course, in recent years it has become more obvious than ever that the title “nanometer” (transistor gate length) is far from a complete characteristic of the photolithographic process. Therefore, we cannot say with confidence how much silicon produced at the 7nm standard at TSMC factories should differ in theory from the 14nm Vega 10 chips manufactured by GlobalFoundries.
Whatever the case, AMD claims that the Radeon VII is 25% faster than the Vega 64 at the same power consumption, and the average increase in FPS in games is from 20 to 42%. Whether this is true or not, we will find out from the tests, but we know where these numbers come from: thanks to four HBM2 memory channels, the Vega 20 chip has more than twice the memory bus bandwidth compared to Vega 10. And the largest gap between new and Radeon RX Vega 64 reaches when the application can master more than 8 GB of VRAM. However, for the Radeon VII, 16 GB of memory is mostly a reserve for the future, because so far only a few games and only in 4K mode with aggressive graphics quality settings go beyond 8 GB. Another thing is working applications for 3D rendering, video editing and similar resource-intensive tasks, in which the Vega architecture has already shown itself well, but was sometimes limited by the amount of local memory.
When it comes to general purpose computing, it’s worth noting that the Radeon VII has the ability, unique for its price, to execute double precision (FP64) operations at 1/4 of the throughput in FP32. As mentioned above, for the Vega 20 chip itself, this characteristic is equal to 1/2 of FP32. AMD decided to artificially limit the consumer version of silicon, but even in this form Radeon VII is unmatched in terms of price/performance ratio in double precision calculations. At the same time, in order to leave an appropriate distance between Radeon VII and the Radeon Instinct family of accelerators, the new product was relieved of a number of other regalia of a server product: in particular, there is no Infinify Fabric bus connector for combining several GPUs into a cluster and supporting fourth-generation PCI Express — the absence of the latter is rather discouraging in light of the fact that the upcoming Zen 2 platform will take full advantage of it.
The Radeon VII went on sale for a suggested price of $699. The price clearly indicates the video card of a competing company, with which the new product will have to fight in benchmarks — the GeForce RTX 2080. AMD assures that these devices are equivalent in terms of gaming performance. Like it or not, we will find out today. But whether AMD has the ability to satisfy the demand for the Radeon VII from hungry fans is still an open question. The price of almost $700 is a record for single-processor Radeon accelerators. However, according to unofficial, but quite convincing calculations, the production of one copy of Radeon VII (taking into account the release of suitable chips at a rate of 7 nm and current prices for HBM2 memory) costs the company $650, and at first the video card will be in severe deficit.
Reference accelerators on «red» chips, intended for mass production, have never been distinguished by exquisite design. AMD engineers (or rather, SAPPHIRE, because it is this company that always produces reference samples for AMD) cannot be reproached for neglecting quality, but, as a rule, they think about the aesthetics of the product last — even if we are talking about such expensive and in some in the sense of status things, like high-performance graphics cards. The only exceptions were certain modifications of the Radeon RX Vega (limited series Frontier Edition, Limited Edition and Vega 64 Liquid Cooled), the high price of which justifies the manufacturer’s design costs and expensive materials. However, Radeon VII is also not intended for the mass market; for AMD, this is primarily a signal of readiness to fight for the top-tier video card market. So, the device should make a favorable impression when meeting.
In the guise of the Radeon VII, the same “design language” is heard, but which is spoken by the “premium” versions of the Radeon RX Vega. The graphics card shell is entirely made of anodized aluminum, and only the Radeon logo and the R cube in the corner of the case remain of the plastic elements — both of which glow red when the device is receiving power. It is noteworthy that even the cooling fans have aluminum covers.
But, of course, the main difference between Radeon VII and previous reference cards based on large chips of the GCN architecture is the open-type cooling system. SAPPHIRE turns out solid coolers with a radial fan — even Vega 64 with a nominal power consumption of around 295 W is served by a similar CO. The weak point of the “turbine” is the level of noise, and as soon as the Radeon VII crossed the 300 W mark, AMD had to create a more efficient design for the reference version.
The length of the printed circuit board and the case of the Radeon VII — in these parameters the new product does not differ from the Vega 64 — made it possible to place three impellers with a diameter of 75 mm above the radiator. Compared to most fans used by AMD and NVIDIA partners in devices of the original design, SAPPHIRE chose an unusually powerful model that can reach speeds above 3000 rpm — looking ahead, we admit that this is very suitable for the new product, but about the temperature regime of the Vega chip 20 we will talk a little later.
Compared to the reference versions of Vega 64, AMD’s new flagship has grown in width: now the case protrudes beyond the boundaries of the panel with video outputs. This is primarily due to the increased area of the printed circuit board, and at the same time, under the front panel of the casing, there is more space for the GPU heatsink. The Radeon VII cooling system is well-designed, but frustratingly simple: the GPU and HBM2 memory chips are covered with a vapor chamber, and five flattened heat pipes 10 mm wide help distribute heat over the radiator fins. AMD has done away with such frills as a vapor chamber on the entire area of the printed circuit board, and the voltage regulator components are cooled by a flat metal frame. The last part is separated from the main radiator: there is contact between it and the bottom surface of the heat pipes, but it hardly contributes to heat transfer, because. There are no thermal pads in between.
On the whole, we can state that the cooler of the new design is quite worthy of cooling such a powerful and hot GPU as Vega 20, but it is unlikely to show extraordinary efficiency in tests. Whatever one may say, for 300 watts of power, the Radeon VII has a rather small heatsink, and the direction of the heatsink fins and the closed shape of the shroud allow air to be released only perpendicular to the motherboard.
However, these shortcomings are compensated to some extent by the unusual thermal interface between the heatsink and the GPU chip. Instead of thermal paste, the base of the evaporation chamber is covered with a special dense composition. According to AMD, it is known that this is a Hitachi TC-HM03 thermal pad based on graphite fibers. But with the usual thermal pads that are glued to the VRM field effect transistors in the Radeon VII, it has little in common. Passport thermal conductivity of Hitachi TC-HM03 (25-45 W/m·K) is significantly higher than that of high-quality mineral thermal pastes (12.5 W/m·K). The only disadvantage of this solution is that the gasket cannot be reused after removing the heatsink — the material will definitely tear and it is not a fact that it can be replaced with ordinary thermal paste. The fact is that the gap between the base of the evaporation chamber and the surface of the GPU is designed for the thickness of the gasket — after the user removes the scraps from the Hitachi TC-HM03 chip, the thermal paste will have to be applied in an excessively thick layer, and no one guarantees uniform heatsink pressure. In theory, you can increase the pressure with the screws on the back of the PCB, but they seem to be already screwed to the ground.
Among the first owners of Radeon VII, there will definitely be daredevils who dismantle the cooling system, and then it will become known whether the new product allows a safe change of thermal interface. But we decided to save an expensive and potentially scarce device, leaving the radiator in place.
⇡#Printed circuit board
Since we did not disassemble the test Radeon VII due to fears of disrupting the cooling system or splitting the GPU as a result of replacing the thermal interface, and AMD did not provide photos of the internal structure of the video card, we had to resort to third-party sources in this matter. However, the PCB of various Radeon VII samples hardly differ within the same batch, and the only photo of the front surface contains almost all the information we are interested in.
Due to the fact that the HBM2 RAM stacks are located on a common silicon substrate with a graphics processor (unlike the first samples of the Radeon RX Vega 64, the gaps between the chips are already filled with epoxy compound), a significant part of the textolite in Radeon VII is empty, but it is immediately clear with what The goal was to make the PCB wider than the Radeon RX Vega reference samples. Additional area was needed in order to accommodate the components of a reinforced power system. There are landing pads for chokes, capacitors and MOSFETs of sixteen separate phases. AMD has not only ditched dual FET drivers, but is using expensive integrated driver FETs (DrMOS). In addition, the Radeon VII’s PCB is unusually thick — probably thicker than any other graphics card we’ve dealt with. And since the Vega 20 chip does not need tracks on the board to communicate with the RAM, the reason is probably the voltage regulator.
However, of the sixteen VRM phases in the Radeon VII, only twelve are soldered — one less than on the first-generation Vega reference boards. Judging by the wiring, ten phases are for the GPU and two for the HBM2 memory chips. Power is supplied through two eight-pin connectors, which means that the power that the Radeon VII can master within the current that these conductors are designed for along with the PCI Express slot is 375 watts. Judging by the fact that the BIOS of the video card allows you to increase power consumption from the standard 300 W by 20% (up to 360 W), AMD complies with these specifications.
Concluding the physical examination of the novelty, we note that the Radeon VII board lacks two components that no high-performance AMD card has been without in recent years — a power indicator in the form of an LED bar next to the power connectors and, more importantly, a backup BIOS chip.
⇡#Test stand, testing methodology
|Test bench configuration|
|CPU||Intel Core i9-9900K (4.9 GHz)|
|Motherboard||ASUS MAXIMUS XI APEX|
|RAM||G.Skill Trident Z RGB F4-3200C14D-16GTZR, 2 x 8 GB (3200 MHz, CL14)|
|ROM||Intel SSD 760p, 1024 GB|
|Power Supply||Corsair AX1200i 1200W|
|CPU cooling system||Thermaltake Water 3.0 Ultimate|
|Frame||CoolerMaster Test Bench V1.0|
|Operating system||Windows 10 Pro x64|
|AMD GPU software|
|All video cards||25.20.15015.2003 Press|
|NVIDIA GPU software|
|All video cards||NVIDIA GeForce Game Ready Driver 418.81|
|Synthetic 3D Graphics Benchmarks|
|Test||API||Permission||Full screen anti-aliasing|
|3D Mark Fire Strike 1.1||DirectX 11 (feature level 11_0)||1920×1080||Off|
|3DMark Fire Strike 1.1 Extreme||2560×1440|
|3DMark Fire Strike 1.1 Ultra||3840×2160|
|3D Mark Time Spy 1.1||DirectX 12 (feature level 11_0)||2560×1440|
|3DMark Time Spy Extreme 1.1||3840×2160|
|Game (in order of release date)||API||Settings, test method||Full screen anti-aliasing|
|1920×1080 / 2560×1440||3840×2160|
|gta v||DirectX 11||Built-in benchmark. Max. graphics quality||MSAA 4x + FXAA + Reflection MSAA 4x||Off|
|Ashes of the Singularity: Escalation||Vulkan||Built-in benchmark. Max. graphics quality||MSAA 4x + TAA 4x|
|Total War: WARHAMMER II, built-in benchmark||DirectX 12||Built-in benchmark (Battle Benchmark). Max. graphics quality||MSAA 4x|
|Wolfenstein II: The New Colossus||Vulkan||OCAT, Roswell mission. Max. graphics quality. Deferred Rendering Off, GPU Culling Off, Adaptive Shading Off.||TSSAA (8TX)|
|Final Fantasy XV||DirectX 11||Built-in benchmark + OCAT. Max. graphics quality. GameWorks Off, DLSS Off||TAA|
|Far Cry 5||DirectX 11||Built-in benchmark. Max. graphics quality||TAA|
|F1 2018||DirectX 11||Built-in benchmark. Max. graphics quality||TAA|
|Strange Brigade||Vulkan||Built-in benchmark. Max. graphics quality||AA Ultra|
|Shadow of the Tomb Raider||DirectX 12||Built-in benchmark. Max. graphics quality||SMAA 4x|
|Assassin’s Creed Odyssey||DirectX 11||Built-in benchmark. Max. graphics quality||TAA High|
|Battlefield V||DirectX 12||OCAT, Liberte mission. Max. graphics quality. DXR Off||TAA High||TAA High|
|General purpose computing, video encoding/decoding|
|Adobe Premier CC 2019||Rendering and encoding 8K video||Export to H.265 (HEVC) 8K@24p|
|Blender 2.8 Beta, Cycles||BMW Demo||—|
|CompuBench 2.0||Ocean Surface Simulation||—|
|N-Body Simulation 1024K||—|
|DXVA Checker 4.1.2, Decode Benchmark||H.264||1920 × 1080 (High Profile, L4.1), 3840 × 2160 (High Profile, L5.1). Microsoft H264 Video Decoder|
|H.265||1920 × 1080 (Main Profile, L4.0), 3840 × 2160 (Main Profile, L5.0), 7680 × 4320 (Main Profile, L6.0). Microsoft HEVC Video Extensions|
|VP9||1920×1080, 3840×2160, 7680×4320. Microsoft WebM MF VP8 Decoder|
|ffmpeg 4.0.2, H.264 encoding||1920×1080||-c:v h264_amf -quality speed -coder cabac -level 4.1 -refs 1 -b:v 3M||-c:v h264_nvenc -preset fast -coder cabac -level 4.1 -refs 1 -b:v 3M|
|3840×2160||-c:v h264_amf -quality speed -coder cabac -level 5.1 -refs 1 -b:v 7.5M||-c:v h264_nvenc -preset fast -coder cabac -level 5.1 -refs 1 -b:v 7.5M|
|Ffmpeg 4.0.2, H.265 encoding||1920×1080||-c:v hevc_amf -quality speed -level 4 -b:v 3M||-c:v hevc_nvenc -preset fast -level 4 -b:v 3M|
|3840×2160||-c:v hevc_amf -quality speed -level 5 -b:v 7.5M||-c:v hevc_nvenc -preset fast -level 5 -b:v 7.5M|
|7680×4320||—||-c:v hevc_nvenc -preset fast -level 6 -refs 1 -b:v 20M|
|Lux Mark 3.1||Hotel Lobby (Complex Benchmark)||—|
|SiSoftware Sandra Titanium (2018) 2018.8.28.26||GPGPU Scientific Analysis||OpenCL, FP16/FP32|
We register the power of video cards separately from the CPU and other PC components using an ammeter MingHe VAC-1050A. To simultaneously measure the current passing through the additional power connectors and the motherboard slot, a PCI Express x16 riser is used, in which the power lines are broken and brought to a separate cable.
As a test load, FurMark is used with the most aggressive settings (3840 × 2160 resolution, MSAA 8x) and Crysis 3 (maximum graphics quality, 3840 × 2160 resolution, MSAA 4x). Power measurements are taken after the graphics card has warmed up, when the GPU temperature and clock speeds have stabilized. We also log a number of other variables during the test using MSI Afterburner software: clock speed, GPU voltage and temperature, and cooling fan speed.
The following video cards took part in performance testing:
⇡#Clock speeds, power consumption, temperature, overclocking
Opportunities for overclocking the Radeon VII and analyzing its performance are currently very limited. The AMD WattMan utility allows you to log only the frequency of the GPU, RAM, GPU temperature and fan speed. And overclocking software — such as MSI Afterburner or GPU-Z — does not yet have access to this video card at all. But even in such conditions, it is still possible to experiment with Radeon VII, and quite fruitfully.
The curve connecting the voltage of the graphics core and its clock speed ends at 1802 MHz and 1.08 V (while the maximum voltage that can be set manually is 1.218 V) — Radeon VII works quite close to these parameters. So, under load in the 3DMark Time Spy stress test, after the GPU temperature stabilizes, the GPU frequency fluctuates around 1765 and reaches 1788 MHz. When it comes to power consumption, the Radeon VII isn’t doing nearly as badly as we feared. The power of the video card turned out to be 29 W less compared to the Radeon RX Vega 64, although the energy efficiency of the latest NVIDIA chips still does not shine for AMD products, despite the nanometer advantage.
In the PowerTune algorithms, the new AMD flagship has one feature that radically changes the behavior of the cooling system and the GPU’s response to overheating: along with the GPU temperature at the edge sensor, which the API of past AMD video cards gives to monitoring programs, WattMan reports another value — Junction Temperature. This parameter is formed by 64 sensors scattered over the area of the GPU and reflects the temperature in the hottest zone. Temperature measurements in different parts of the crystal are not new to Vega 20. Previous generation GPUs also have this feature, but before it was needed only for emergency shutdown of the chip in case of catastrophic overheating. Now it is Junction Temperature that signals the GPU that it is time to lower the clock speed or speed up the fans. The cooling system strives to keep the Junction Temperature within 100 ° C, accelerating the fans up to 2725 rpm, and when heated to 110 ° C, “throttling” already occurs — the GPU clock frequencies fall. The graphs clearly show that there is always a big difference between the Junction Temperature and the readings of the edge sensor: with the standard settings of the Radeon VII edge sensor reports 74°C, Junction Temperature at the same time reaches 106 °C.
As the creators of Radeon VII said, the distributed measurement that underlies the concept of Junction Temperature is not needed at all in order to limit overclocking (both using automatic and manual methods), but on the contrary, in order to use the slightest opportunity to raise clock frequencies during periods short-term cooling of the chip. On the other hand, this can be interpreted in such a way that the GPU always strives to approach a temperature value at which current leakage begins to threaten stable operation. To some extent, this seems to be true, because when overclocked by the user, the card really turned out to be very sensitive to cooling.
Our sample Radeon VII allows you to increase the maximum frequency of the GPU from the standard 1802 to 2002 MHz, and the effective frequency of the RAM from 2 to 2.3 GHz. To do this, we had to raise the supply voltage to 1.13 V and expand the power reserve by 20%, but the main thing is that stable operation is possible only at the maximum speed of the cooling fans (3850 rpm)! The «external» temperature of the GPU in this case is even lower than in normal mode (67 versus 74 ° C), but the Junction Temperature is exactly the same — it is obvious that everything depends on it.
Fortunately, with a more conservative overclocking of the Radeon VII, it is not necessary to think about heating. Without exceeding the nominal supply voltage of 1.08 V, we managed to raise the maximum GPU frequency to 1952 MHz, and under load it stabilized at 1914 MHz. It was with these settings that the Radeon VII passed repeated performance tests. With the exception of 3DMark — in this test package, when you try to overclock the slightest, there is not an increase, but, on the contrary, a drop in clock frequencies to a zone of about 1600 MHz and a loss of final points (but for the Time Spy stress test, in which we measured the power, this is paradoxical way does not apply). Needless to say, Radeon VII is currently behaving extremely suspiciously in overclocking. The “raw driver” is probably to blame, and we will have to return to this topic as soon as AMD debugs its software.
Like first-generation Vega-based graphics cards, the Radeon VII benefits from GPU undervolting. The voltage on the GPU at the top of the auto-acceleration curve can be fearlessly reduced by 100 mV — clock speeds do not suffer from this at all, but the Radeon VII is already comparable in power to the GeForce RTX 2080. The CO fan speed drops to 1700-1800 rpm due to undervolting, and this is clearly seen in the noise level.
The results of the Radeon VII, which we received in the synthetic tests of the 3DMark package, correspond to the data that was leaked to the Internet even before the publication of the review: the AMD accelerator outperforms the GeForce RTX 2080 in Fire Strike test points under the Direct3D 11 API, but in the Time Spy benchmark under Direct3D 12 NVIDIA graphics card leads by a wide margin. On average, the advantage of the RTX 2080 over the Radeon VII in «synthetics» is 5%. Moreover, all three NVIDIA accelerators — GeForce RTX 2070, RTX 2080 and GTX 1080 Ti — are ahead of the Radeon VII in Time Spy, but according to the average estimate, the new AMD flagship is not inferior to the GTX 1080 Ti and is 9% superior to the RTX 2070.
⇡#Game tests (1920×1080)
Since this review, we have been testing graphics cards on the updated platform with the Intel Core i9-9900K processor, and this helps to more clearly see the difference between various devices in 1080p mode. However, the second-generation Vega GPU at this resolution still cannot work at full capacity — already due to its own architectural qualities.
In terms of average frame rate in eleven games, the Radeon VII is 20% faster than the Radeon RX Vega 64, which corresponds to the lower limit of the range in which AMD evaluated the performance of the new product. Comparison with other test participants also does not show the Radeon VII in a favorable light. Its main rivals — the GeForce GTX 1080 Ti and GeForce RTX 2080 — lead here with an advantage of 6 and 11%, respectively. Even the GeForce RTX 2070 is only 2% behind AMD’s flagship accelerator.
As in previous reviews, when we compared AMD and NVIDIA devices with comparable performance potential, some games lean towards the «green» side, while others — towards the «red» side. Only «green» benchmarks in this case are clearly more.
⇡#Game tests (2560×1440)
Based on the 1440p test results, it is noticeable that the increase in resolution is gradually tipping the scales in favor of the Radeon VII. So, the new product is separated from the Radeon RX Vega 64 by a distance of 22% of the frame rate. But the balance of power between Radeon VII and other video cards has not fundamentally changed. The advantage of 4 and 10%, respectively, is still on the side of the GeForce GTX 1080 Ti and GeForce RTX 2080, respectively, and the GeForce RTX 2070 is still in threatening proximity to the Radeon VII (5% difference).
⇡#Game tests (3840×2160)
While the first-generation Vega-based graphics cards can’t quite cope with modern games at 4K resolution, the Radeon VII, on the contrary, feels quite confident in 2160p mode: according to the results of most tests, the new product provides an average frame rate above 60 FPS and has a good headroom. at the minimum frame rate. By the way, the minimum frame rate in this case could indicate a possible lack of RAM for a particular video card, however, rivals of the Radeon VII, equipped with 8 GB of RAM, also do not allow an abnormal strong drop in performance.
In a situation where the extremely fast Radeon VII memory bus and upgraded ROPs are actively loaded, the new product is the easiest to compete with the GeForce RTX 2080 — the average advantage of the NVIDIA graphics card has decreased by 7%. On the other hand, the Radeon VII achieved parity with the GeForce GTX 1080 Ti and outperformed the GeForce RTX 2070 by 9%.
⇡#Computing general purpose
From a purely technical point of view, AMD turned out to be an extremely interesting product. Like it or not, Vega 20 is the first GPU on the market built on the 7nm FinFET norm. At this stage in the development of the semiconductor industry, it is AMD that is mastering advanced technology ahead of rivals — on the front of both central processors and discrete GPUs. But in terms of performance in gaming applications, the Radeon VII barely achieved the goal that AMD was counting on.
Due to the radically increased clock frequencies, the novelty develops a level of performance that the first generation Vega chips never dreamed of: we are talking about an advantage of 20-25%, depending on the screen resolution. But at the same time, it is still impossible to consider the Radeon VII as a really equal rival for the GeForce RTX 2080. According to the results of our tests in eleven popular games, the NVIDIA accelerator leads by 9-10% in average frame rate at relatively low screen resolutions (1080p and 1440p), and only in 4K mode the distance between rivals is reduced to 7%. However, it should be noted that victory in this battle largely depends on the optimization of specific games for a particular architecture: in our test method there are projects that do not see the difference between Radeon VII and RX 2080 or even prefer the AMD product.
So let AMD once again fail to play on an equal footing with NVIDIA, Radeon VII has every right to exist in the $699 price category — if only the retail prices of the new product really correspond to this level. However, a position somewhere close to the GeForce RTX 2080 for the Radeon VII is completely unacceptable if the Radeon Technologies Group is going to return the sympathy of computer hardware enthusiasts with this video card. AMD considers the big advantage of the novelty, which outweighs the lag in performance from its main rival, the amount of RAM 16 GB. We tend to think that in the next year or two most games will still not need such a large frame buffer. On the other hand, the amount of RAM, doubled compared to what is now considered the norm for accelerators of this performance category, should be appreciated by those buyers who upgrade every five years.
Finally, the advantage of Vega chips in general-purpose calculations has not gone away. The Radeon VII is an extremely powerful workstation graphics card that AMD is selling at a bargain price in terms of its features and performance.