▷ Amd vega

AMD Vega is the name of AMD's most advanced graphics architecture, it is the latest evolution of GCN, its GPU architecture that has accompanied us since 2011. This evolution of GCN is the most ambitious of AMD to date.

Do you want to know more about AMD VEGA graphics cards and all their features? In this post we review all the keys to the GCN architecture and all the secrets that Vega hides.

Index of contents

The birth of GCN architecture and its evolution until reaching Vega

To understand AMD's history in the graphics card market, we have to go back to 2006, when the Sunnyvale company took over ATI, the world's second-largest graphics card manufacturer, and which had been in business for years. Fight with Nvidia, industry leader. AMD purchased all of ATI's technology and intellectual property in a transaction worth $ 4.3 billion in cash and $ 58 million in shares for a total of $ 5.4 billion, completing the action on October 25, 2006.

At that time ATI was developing what would be its first GPU architecture based on the use of unified shaders. Until then, all graphics cards contained different shaders inside for vertex and shading processing. With the arrival of DirectX 10, unified shaders were supported, which means that all shaders in a GPU can work with vertices and shades indifferently.

TeraScale was the architecture that ATI was designing with support for unified shaders. The first commercial product to make use of this architecture was the Xbox 360 video console, whose GPU, called Xenos, had been developed by AMD and was much more advanced than what could be mounted on PCs of the time. In the PC world, TereaScale brought graphics cards from the Radeon HD 2000, 3000, 4000, 5000, and 6000 series to life. All of them were making small improvements continuously to improve their capabilities as they progressed in the manufacturing processes, from 90 nm to 40 nm.

The years went by and the TeraScale architecture was becoming outdated compared to Nvidia. TeraScale's performance in video games was still very good, but it had a great weak point compared to Nvidia, this was a low capacity for computing using GPGPU. AMD understood that it needed to design a new graphic architecture, capable of fighting with Nvidia both in games and in computing, a section that was increasingly important.

We recommend reading our best PC hardware and component guides:

• AMD history, processors and graphics cards of the green giant

GCN is the graphical architecture designed by AMD from the ground up to succeed ATI's TeraScale

Graphics Core Next is the name given to the first graphic architecture designed 100% by AMD, although logically everything inherited from ATI has been key to making its development possible. Graphics Core Next is much more than an architecture, this concept represents the code name for a series of graphic microarchitectures and a set of instructions. The first GCN-based product arrived at the end of 2011, the Radeon HD 7970 that has given such good results to all its users.

GCN is a RISC SIMD microarchitecture that contrasts with the VLIW SIMD TeraScale architecture. GCN has the disadvantage that it requires many more transistors than TeraScale, but in return it offers much greater capabilities for calculating GPGPU, makes the compiler simpler, and makes better use of resources. All this makes GCN an architecture clearly superior to TeraScale, and much better prepared to adapt to the new demands of the market. The first GCN-based graphics core was Tahiti, which brought the Radeon HD 7970 to life. Tahiti was built using a 28nm process, representing a huge leap in energy efficiency compared to 40nm for the latest TeraScale-based graphics core, the Radeon HD 6970's Cayman GPU.

Thereafter, the GCN architecture has evolved slightly over several generations of Radeon HD 7000, HD 8000, R 200, R 300, RX 400, RX 500, and RX Vega series graphics cards. The Radeon RX 400s ushered in a manufacturing process at 14nm, allowing GCN to take a new leap in energy efficiency. The GCN architecture is also used in the APU graphics core of PlayStation 4 and Xbox One, the current video game consoles from Sony and Microsoft that offer exceptional performance for their price.

The GCN architecture is organized internally into what we call computational units (CU), which are the basic functional units of this architecture. AMD designs GPUs with a greater or lesser number of computing units to create its different ranges of graphics cards. In turn, it is possible to deactivate computing units in each of these GPUs to create different ranges of graphics cards based on the same chip. This allows us to take advantage of the silicon that has come out of the manufacturing process with problems in some of the computing units, it is something that has been done in the industry for many years. The Vega 64 GPU has 64 computing units inside and is the most powerful GPU manufactured by AMD to date.

Each computing unit combines 64 shading processors or shaders with 4 TMUs inside. The computing unit is separate from, but is powered by, the Processing Output Units (ROPs). Each Compute Unit consists of a Scheduler CU, a Branch & Message Unit, 4 SIMD Vector Units, 4 64KiB VGPR files, 1 scalar unit, a 4 KiB GPR file, a local data quota of 64 KiB, 4 texture filter units, 16 texture recovery load / storage units and a 16 kB L1 cache.

AMD Vega is GCN's most ambitious evolution

The differences between the different generations of the GCN architecture are quite minimal and do not differ too much from each other. An exception is the fifth-generation GCN architecture, called Vega, which has greatly modified shaders to improve performance per clock cycle. AMD began releasing details of AMD Vega in January 2017, causing high expectations from the first moments. AMD Vega increases instructions per clock, reaches higher clock speeds, offers support for HBM2 memory and a larger memory address space. All of these features allow you to significantly improve performance over previous generations, at least on paper.

Architectural improvements also include new hardware programmers, a new primitive discard accelerator, a new display driver, and an updated UVD that can decode HEVC at 4K resolutions at 60 i frames per second in 10-bit quality per color channel..

The computing units are heavily modified

The AMD Vega development team, led by Raja Koduri, modified the fundamental plane of the calculation unit to achieve much more aggressive frequency targets. In previous GCN architectures, the presence of connections of a certain length was acceptable because the signals could travel the full distance in a single clock cycle. Some of those pipeline lengths had to be shortened with Vega so that signals could traverse them in the span of clock cycles, which are much shorter in Vega. AMD Vega's computing units became known as NCU, which can be translated as a new generation computing unit. To the reduction of the pipeline lengths of AMD Vega were added modifications in the logic of search and decoding of instructions, which were reconstructed in order to meet the objectives of shorter execution times in this generation of graphics cards.

On the L1 cache texture decompression data path, the development team added more steps to the pipeline to reduce the amount of work done in each clock cycle to meet the goals of increasing the operating frequency. Adding stages is a common means of improving the frequency tolerance of a design.

Rapid Packet Math

Another important novelty of AMD Vega is that it supports the simultaneous processing of two operations with less precision (FP16) instead of a single one with greater precision (FP32). This is technology called Rapid Packet Math. Rapid Packet Math is one of the most advanced features in AMD Vega and is not present in previous GCN versions. This technology allows a more efficient use of the GPU's processing power, which improves its performance. The PlayStation 4 Pro is the device that has benefited the most from Rapid Packet Math and has done so with one of its star games, Horizon Zero Dawn.

Horizon Zero Dawn is a great sample of what Rapid Packet Math can bring. This game uses this advanced technology to process everything related to grass, thus saving resources that can be used by developers to improve the graphic quality of other elements of the game. Horizon Zero Dawn impacted from the first moment for its overwhelming graphic quality, to the point that it is impressive that a console of only 400 euros can offer such an artistic section. Unfortunately, the Rapid Packet Math has not yet been used in PC games, much of the blame for this being that it is an exclusive feature of Vega, as developers do not want to invest resources in something that very few users will be able to take advantage of..

Primitive shaders

AMD Vega also adds support for new Primitive Shaders technology that provide more flexible geometry processing and replace vertex and geometry shaders in a render pipe. The idea of ​​this technology is to eliminate non-visible vertices from the scene so that the GPU does not have to calculate them, thereby reducing the level of load on the graphics card and improving the performance of the video game. Unfortunately, this is a technology that requires a lot of work on the part of the developers to be able to take advantage of it and it finds a situation very similar to that of Rapid Packet Math.

AMD had the intention to implement the Primitive Shaders at the driver level, which would allow this technology to work magically and without the developers having to do anything. This is something that sounded very nice, but finally it was not possible due to the impossibility of implementing it in DirectX 12 and the rest of the current APIs. The Primitive Shaders are still available, but it must be the developers who invest resources for their implementation.

ACE and Asynchronous Shaders

If we talk about AMD and its GCN architecture we have to talk about Asynchronous Shaders, a term that was talked about a long time ago, but about which almost nothing is said anymore. Asynchronous Shaders refer to asynchronous computing, it is a technology that AMD devised to reduce the deficiency suffered by its graphics cards with geometry.

AMD graphics cards based on the GCN architecture include ACEs (Asynchronous Compute Engine), these units consist of a hardware engine dedicated to asynchronous computing, it is a hardware that takes up space on the chip and consumes energy so its Implementation is not a whim but a necessity. The reason for the existence of ACEs is the poor efficiency of GCN when it comes to distributing the workload between the different Compute Units and the nuclei that form them, which means that many nuclei are out of work and therefore wasted, although they remain consuming energy. The ACE are in charge of giving work to these nuclei that have remained unemployed so that they can be used.

The geometry has been improved in the AMD Vega architecture, although it still lags far behind Nvidia's Pascal architecture in this regard. GCN's poor efficiency with geometry is one reason AMD's larger chips don't deliver the expected result from them, as the GCN architecture becomes more inefficient with geometry as the chip grows larger. and include a greater number of units of computation. Improving geometry is one of AMD's key tasks with its new graphics architectures.

HBCC and HBM2 memory

The AMD Vega architecture also includes a High Bandwidth Cache Controller (HBCC), which is not present in the graphics cores of Raven Ridge APUs. This HBCC controller allows more efficient use of the HBM2 memory of Vega based graphics cards. In addition, it allows the GPU to access the DDR4 RAM of the system if the HBM2 memory runs out. HBCC allows this access to be done much more quickly and efficiently, resulting in less performance loss compared to previous generations.

HBM2 is the most advanced memory technology for graphics cards, it is the second generation high bandwidth stacked memory. HBM2 technology stacks different memory chips on top of each other to create an extremely high density package. These stacked chips communicate with each other via an interconnect bus, the interface of which can reach 4, 096 bits.

These characteristics make the HBM2 memory offer a much higher bandwidth than is possible with GDDR memories, in addition to doing it with a much lower voltage and power consumption. Another advantage of HBM2 memories is that they are placed very close to the GPU, which saves space on the graphics card PCB and simplifies its design.

The bad part about HBM2 memories is that they are much more expensive than GDDRs and much more difficult to use. These memories communicate with the GPU through an interposer, an element that is quite expensive to manufacture, and which makes the final price of the graphics card more expensive. As a consequence, HBM2 memory-based graphics cards are much more expensive to manufacture than GDDR memory-based graphics cards.

This high price of HBM2 memory and its implementation, as well as a lower performance than expected, have been the main causes of AMD Vega's failure in the gaming market. AMD Vega has failed to outperform the GeForce GTX 1080 Ti, a card based on a Pascal architecture nearly two years older.

Current graphics cards based on AMD Vega

AMD's current graphics cards under the Vega architecture are the Radeon RX Vega 56 and the Radeon RX Vega 64. The following table lists all the most important features of these new graphics cards.

Current AMD Vega graphics cards
Graphic card	Compute Units / Shaders	Base / Turbo Clock Frequency	Amount of memory	Memory interface	Memory type	Memory bandwidth	TDP
AMD Radeon RX Vega 56	56 / 3, 584	1156/1471 MHz	8 GB	2, 048 bits	HBM2	410 GB / s	210W
AMD Radeon RX Vega 64	64 / 4, 096	1247/1546 MHz	8 GB	2, 048 bits	HBM2	483.8 GB / s	295W

The AMD Radeon RX Vega 64 is the most powerful graphics card from AMD today for the gaming market. This card is based on Vega 10 silicon, made up of 64 Compute Units that translate into 4, 096 shaders, 256 TMUs and 64 ROPs. This graphics core is capable of working at a clock frequency of up to 1546 MHz with a TDP of 295W.

The graphics core is accompanied by two HBM2 memory stacks, which add up to a total of 8 GB with a 4, 096-bit interface and a bandwidth of 483.8 GB / s. It is a graphics card with a very large core, the largest ever made by AMD, but which is not capable of performing at the level of the GeForce GTX 1080 Ti Pascal GP102 core, in addition to consuming more energy and producing much more heat. This inability of AMD to fight with Nvidia seems to make it clear that the GCN architecture needs a much bigger evolution to keep up with Nvidia's graphics cards.

AMD Vega's future goes through 7nm

AMD is going to breathe new life into its AMD Vega architecture with the move to a 7nm manufacturing process, which should mean a significant improvement in energy efficiency over current designs at 14nm. For now AMD Vega at 7 nm will not reach the gaming market, but will focus on the artificial intelligence sector, which moves large amounts of money. Concrete details about AMD Vega at 7nm are not yet known, the improvement in energy efficiency can be used to maintain the performance of current cards but with a much lower power consumption, or to make new cards much more powerful with the same consumption as the current ones.

The first cards to use AMD Vega at 7nm will be the Radeon Instinct. Vega 20 is the first AMD GPU manufactured at 7nm, it is a graphic core that offers double the density of transistors compared to the current Vega 10 silicon. The size of the Vega 20 chip is approximately 360mm2, which represents a reduction surface area of ​​70% compared to Vega 10 which has a size of 510mm2. This breakthrough enables AMD to offer a new graphics core with 20% faster clock speed and an energy efficiency improvement of approximately 40%. Vega 20 has a power of 20.9 TFLOPs, making it the most powerful graphics core announced to date, even more than Nvidia's Volta V100 core offering 15.7 TFLOPs, although this one is manufactured at 12nm, which puts AMD at a clear advantage in this regard.

This ends our post on AMD Vega. Remember that you can share this post with your friends on social networks, in this way you help us to spread it so that it can help more users who need it. You can also leave a comment if you have something else to add or leave us a message in our hardware forum.