Looking back at the past point in time, Intel officially announced the high-performance graphics product brand - "Intel Arc" on August 17, 2021, Chinese name sharp, until March 30, 2022, Intel officially unveiled its mystery.
Since Intel announced its unique graphics initiative, Intel-made graphics has been a hot topic in the technology industry. For consumers, finally having a third choice in addition to NVIDIA and AMD means that the new competitive landscape will bring new trends and trends to this market. According to the previous disclosure, intel Xe-HPG ruixuan graphics card will use TSMC 6nm process, abandon the original EU basic module, the use of a new Xe core, the first generation of Xe HPG micro-architecture core codenamed Alchemy (alchemist), the second and third generations will be named Battlemage, Celestial, and then the new next generation Xe architecture of Drid. As you can see, Intel's planning on Xe-HPG has been quite complete.
Part 1. Interpret Intel's accumulation in the GPU field
If you look at the historical process, Intel is not actually a new player in the "graphics club", as early as more than two decades ago, Intel has tried to open discrete graphics, and has never given up on high-performance graphics. In fact, in 2017, Intel came up with a graphics card code-named "DG1", which made users feel extremely excited. But in fact, Intel officially launched a mobile discrete graphics in the market - Intel Iris graphics codenamed "Alchemist".
According to the interpretation of Tom Petersen, an Intel academician (mainly responsible for graphics innovation): "Intel's biggest advantage is platform-level innovation. The reason why we started to release our graphics card from the mobile platform is because this has always been our strength and strength, and we let users feel the cooperation between the platforms through notebooks and mobile terminals. On the basis of mobile notebooks, we can also provide more differentiated features to highlight different experiences. ”
Intel's advantages in the mobile terminal are very obvious, old users should remember that Intel since the launch of the Centrino platform in 2003 has unified the bus interface, wireless specifications, universal interface and other notebook platform standards; then in the traditional high-speed data lightning interface, Intel Quick Sync Video for video codec, as well as Ultrabook, two-in-one innovative form, thin and light Evo platform, have been enhancing the platform advantages of the mobile terminal.
And from the Xe Iris core display to the Xe-HPG Iris monochrome, Intel has actually experienced several generations of accumulation in graphics, in fact, last year Intel has launched a SoC composed of 8 Xes on the architecture day - the server-oriented Ponte Vecchio GPU has left a deep impression on us.
The new A series of mobile discrete graphics follow Intel's naming convention on the CPU, and are divided into three series: 3, 5, and 7, of which 3 series are mainly for mainstream games, 5 series for performance games, and 7 series for hardcore performance games. Three of the series are on sale with the release date of March 30, while the 5 and 7 series will be available in the early summer of 2022. Regarding Intel's performance division, I will analyze them one by one.
Regardless of the family, Intel A-Series mobile discrete graphics will use the same graphics architecture, which is the graphics core we often refer to as the various functional cores of graphics cards. Today, graphics cards place more and more emphasis on the parallelism of graphics cards in the architecture, and it can be seen that Xe-HPG Razor Display is dividing the hardware layer into AXe core with built-in XMX (Matrix Engine), Xe Media Engine that supports next-generation codec standards, and Xe Display Engine to solve high output specifications and new graphics pipelines to handle a variety of different display tasks. Therefore, on top of the A series mobile discrete graphics card, the support for the DX12 Ultimate Ultimate Edition, AI operations, video codec, display output and rendering shading and other tasks are all responsible for different engines, which is conducive to sharing the computing load of the Xe core.
Part 2. See the benefits of Xe HPG microarchitecture
Xe kernel
Xe core is the Xe HPG microarchitecture on the smallest module, just mentioned, it replaces the previous integrated graphics card on the EU concept, it is the most basic execution unit of the graphics card, the performance of a single Xe core is powerful or not, related to the overall performance and energy consumption ratio of this generation of graphics cards. The Xe core of the A series of mobile discrete graphics cards contains 16 256-bit wide SIMD vector engines (XVE), which perform most of the operations for traditional graphics shaders; the kernel also contains 16 matrix engines (XMX), each of which is 1024 bits wide, that is, specifically for AI plus operations; in order to meet the high bandwidth requirements of matrix, vector and ray tracing units, each Xe core has a large local memory of 192KB. It can be dynamically allocated between L1 cache and shared local memory (SLM) according to the needs of each workload.
It can be seen that the Xe kernel is quite comprehensive, and the vector operation and AI operation are placed in the same core, which can flexibly call local memory according to the response load type, and the operation efficiency is higher.
Vector Engine (XVE)
The Vector Engine (XVE) is an important part of the Xe core, and it can be seen that Intel has improved the engine's ALU (Arithmetic Logic Unit) to provide a dedicated execution port for floating-point operations (FPs). FP instructions can be run simultaneously with integer operation (INT) instructions, such as DP4a's fast IN8 calculations are executed by this engine.
Matrix Engine (XMX)
Another part of the Xe kernel is the new XMX matrix engine, which is tasked with enhancing AI acceleration for high-throughput matrix multiplication, covering the most common AI data types including BF16 and INT8.
Just now I mentioned the benefits of putting the Vector Engine (XVE) and matrix engine (XMX) in the same Xe kernel can flexibly call local memory, let's see the execution efficiency of different engines in the implementation of data execution. Mac is the basic SIMD vector instruction used in the graph, equivalent to performing 16 Ops per clock (8 parallel operations multiplication + 8 parallel additions) ;DP 4a is optimized for AI computations that do not require 32 bits of precision, for a total of 32 parallel multiplications; the matrix engine is optimized for a total of 256 operations by accumulating 4 deep pipelines of multiplications.
Therefore, different engines can realize the parallel processing of floating-point FP, integer INT and XMX instructions, and realize the parallelism and resource sharing of the two engines in the form of lock steps to obtain higher computing power and execution efficiency.
Rendering Slice
Render Slice can be understood as the upper unit of the Xe kernel, which is the basic building block of reusable IP. In the Xe-HPG microarchitecture, every 4 Xe cores form a render slice. In each render slice, functional modules such as mesh shaders, sampler feedback, and four hardware light tracking accelerators are included for support for the DX12 Ultimate and for real-time ray tracing.
Further up, the render slice can be used to build different SoCs, that is, different A Series mobile discrete graphics cards, and the currently released product contains at least two render slices and a maximum of eight. It can be seen that the Xe-HPG microarchitecture does not have the visual sense of building blocks, which is the advantage of the architecture - flexible configuration. In addition, its energy consumption ratio has also been significantly improved, compared to the previous generation of product Xe-LP microarchitecture, its performance per watt is improved by 1.5 times.
Part 3. An important feature of A-Series mobile discrete graphics
After talking about the composition and advantages of Xe HPG microarchitecture, in fact, many friends basically understand some important features of A series mobile discrete graphics cards. Of course, I'll expand on it further to get an idea of Intel's progress in these areas.
XeSS supersampling
Intel announced the XeSS supersampling technology when it released the Arc brand on the architecture day, and from the architectural point of view, this technology is also rooted in the Xe core, with the XMX matrix engine as its hardware execution unit. From the technical characteristics disclosed by Intel, XeSS is different from the ordinary Spatial upscaler technology on the market, which is to re-render the low-resolution picture through neural network auxiliary motion vector operations and generate more realistic high-resolution pictures. The technology is similar to NVIDIA's DLSS.
XeSS supersampling technology requires the support of game manufacturers, on top of this year's CES 2022, Intel announced 505 Games' "Death Stranding: Director's Cut Edition" to support this technology, and by the day the A series mobile discrete graphics card was released, 14 games had completed support for XeSS. As you can see, this progress is relatively fast, and as more and more notebooks loaded with A-series mobile discrete graphics cards are on the ground, there will be more game manufacturers who will support this feature.
Xe Media Engine
Xe Media Engine is an important technological advancement for A Series mobile discrete graphics, which integrates Intel's so-called "one of the most advanced media accelerators" can be seen, in addition to H.265 / HEVC, H.264 / MPEG-4 / AVC, VP9 and other major media formats, the engine is the industry's first hardware acceleration support for AV1 encoding and decoding, and its hardware encoding speed is 50 times faster than CPU software encoding. This also means that laptops with A-Series mobile discrete graphics cards for media codec work are likely to free up CPU performance and help the computer perform more tasks at the same time.
Xe display engine
Unlike other graphics architectures, Intel has built a standalone Xe display engine specifically for A-series mobile discrete graphics for high-specification display outputs. It supports the widely compatible HDMI 2.0b and DP1.4a specifications found in today's notebooks, and supports output to a single 1080p@360Hz or 4 4k@120Hz HDR displays. In order to solve the problem of displaying high frame rate screen tearing, Intel has developed its own Speed Sync, and at the same time supports the VESA standard Adaptive Sync, especially the self-developed Speed Sync, which can effectively avoid the high latency problems caused by V-Sync in the past.
On the Xe display engine, there is also an Intel-created Move Sync feature, which is integrated on top of the Sharp Graphics card as a plug-in. The principle of this feature is to reduce visual distortion by blurring the boundary between two torn frames. This plug-in requires no coding and the system load is much lower than frame synchronization technology.
Part 4. Parameters and performance of A-Series graphics cards
Alchemist, that is, A series mobile discrete graphics card using two SoC packages, can be simply called "big chip" and "small chip", the large chip code name is ACM-G10, a total of 32 Xe cores and optical tracking units (that is, 8 render tiles), providing 16MB L2 cache, 256-bit GDDR6 interface and 16 PCIe 4 interface; small chip codenamed ACM-G11, configured as 8 Xe cores and optical tracking units, 4MB L2 cache , 96-bit memory interface, 8-way PCIe 4 interface. Both chips contain two Xe multifunctional codec engines and a four-way display output engine.
From the product parameters point of view, Rui Xuan 3 series contains a total of A350M and A370M two models, providing 6 and 8 two Xe cores and the number of optical tracking units, power consumption design and graphics card frequency are not the same, using the same 4GB 64-bit wide GDDR6 independent video memory; Rui Xuan 5 series only one A550M, using 16 Xe cores and the number of light chasing units; Rui Xuan 7 series also has two configurations, The A730M with 24 Xe cores and optical chase units and the A770M with 32 Xe cores and optical chase units respectively, of which the A770M has up to 16GB of 256 bits of video memory, and the graphics card power also reaches 120~150W.
As mentioned earlier, the full range of A-Series mobile discrete graphics cards supports DirectX12 Ultimate, supporting ray tracing, variable rate shading, mesh shading, and sampler feedback. In addition to hardware-level light tracing, such as variable-rate shading, allows applications such as games to tell graphics cards which parts of the scene require higher detail and which parts can be appropriately reduced, thereby speeding up the rendering of the current frame and ultimately increasing the game frame rate.
Taking the A370M as an example, this graphics card is basically positioned below 1080p to play mainstream games in medium and high quality, and it has twice the performance improvement compared to the core display of the 12th-generation Core mobile processor. Among the test results released by Intel, games such as "Killer 3", "Doom: Eternity", "Return to German Headquarters: New Bloodline" and other games can achieve frame rate performance of 60 to 80fps with medium and high image quality.
In competitive games like Fortnite and GTA, it can be seen that the A370M can run above 90fps at 1080p resolution.
In the creative software, such as the commonly used Handbrake, Dvinci, Premiere Pro, there are different performance improvements, especially the two typical scene applications in Premiere Pro, which have performance performance of 1.8X and 2.4X more than Ruiju graphics cards.
Part 5. The driving force behind the A-Series graphics cards
Only at the hardware level to see the Intel A series of mobile discrete graphics is not comprehensive, it is behind the many technologies for different levels of performance improvement to provide the driving force, in order to facilitate user understanding, Intel packaged it as Intel Deep Link, it actually contains three aspects of technology: dynamic power sharing, super coding and super computing power.
In terms of dynamic power sharing function, Intel actually launched the first version of the dynamic power sharing function as early as 2016, that is, the Kobe-Lake G era, dynamically distributing power between CPU die and GPU die. Today's A series discrete graphics, the application of Intel's latest algorithms, through the cycle to collect various CPU and GPU temperature, occupancy, respective power consumption and other system information, especially in the game GPU load when the GPU Bound scene is too high, dynamically adjust the gpu and CPU power ratio, more power is allocated to the GPU. Cpu high power scenarios, and vice versa. Currently, the minimum interval between adjustments of this function takes only 100 milliseconds.
Super encoding technology improves codec efficiency by integrating the existing codec hardware capabilities of notebook computers, while using codec engines with integrated graphics cards and discrete graphics cards. This collaboration is achieved through OneVPL's API interface, where specific API functions are assigned to different multimedia engines in groups for parallel processing.
With the same logic, Intel can also involve the entire system in content creation, and such a machine learning-based service is called MLS. MLS is a framework in OpenVino that intelligently distributes loads to different hash modules. Also based on the latency sensitivity, throughput, performance requirements, power consumption and other load characteristics of the polling graphics card, the load is assigned to the discrete graphics card, integrated graphics card, or CPU. On closer inspection, Deep Link's three functions are intrinsically linked in terms of operational logic and resource mobilization, so it's not surprising that they are a technology as a whole.
It is very interesting that Intel Deep Link technology is based on the existing architecture of the notebook computer, the maximum possible mobilization of all the hardware resources of the notebook computer, after equipped with a discrete graphics card, for a job can also maximize the integration of graphics card and CPU remaining resources to improve the efficiency of the whole machine.
Specifically, I asked Devon Nekechuk, Director of Graphics Product Management at Intel. Devon mentioned: "Intel Deep Link technology is aimed at overall performance in dynamic power sharing, but super coding and super computing power need to work with different ISVs (software developers). Of course, for users, Intel's AI-based MLS engine is transparent, software algorithms are integrated into the lower-level software libraries, and the system is automatically provisioned according to the load, without human intervention. The ISV can further improve efficiency through deep customization to carry out the entire system and open up the software of the upper and lower layers. At present, Intel has achieved deep matching with important content creation software such as DaVinci Resolve and Handbrake. ”
Based on the existing hardware architecture, Intel Deep Link technology can achieve a 30% performance improvement on dynamic power sharing, a 60% performance improvement on super coding, and a 24% performance improvement in combination with the XMX matrix engine. Whether every A-series mobile discrete graphics card has the same performance improvement capabilities, I will test it for you one by one when the real machine starts.
Part 6. A-series graphics notebook preview
The first A series mobile discrete graphics card released is the Ruixuan 3 series, which will also be mainly installed on the thin and light version of the Evo platform. According to Intel, the Evo thin and light book equipped with the Ruixuan 3 unique display and the 12th generation Core CPU still has more than 9 hours of battery life, which is also a confirmation of the energy consumption ratio performance of the A series discrete graphics cards.
The world's first notebook equipped with Ruixuan graphics card has been launched, for the Samsung Galaxy Book2 Pro equipped with Ruixuan A350M, the current domestic temporarily can not see this product, but from the second quarter, including Acer, Asus, Blue Sky Computer, Dell, Gigabyte, Haier, HP, Lenovo, Samsung, MSI and NEC and other major OEM manufacturers will launch a unique product equipped with Ruixuan 3 series, and Ruixuan 5, 7 series of notes will also be available this summer.
With the launch of the Rui Xuan unique notebook, all Rui Xuan unique display will be loaded with a unified graphics card management software - Arc Control. It provides a workload of rapid upgrade, real-time viewing of graphics card performance, and related settings for live broadcasting and streaming, such as opening the live broadcast function to share the wonderful picture of the game to the live broadcast platform, turning on the virtual camera to remove the background, automatically adjusting the screen ratio, and capturing and saving the highlight moments in the game.
Part 7. Write at the end
The interpretation of the Intel Iris A series of mobile discrete graphics will be reported here for the time being, but the report on the product has just begun for us. I believe that everyone, like us, is very interested in the real performance of the A series mobile discrete graphics card on the notebook. And with the listing of Intel's unique display, it will undoubtedly cause a violent chemical reaction to the market, and will also bring a series of uncertainties: for example, how does Intel's thin and light display perform? Is there an advantage in the price? Will it be opened to other manufacturers? How well are game manufacturers and software manufacturers following up on XeSS, Deep Link and other software? We will follow up on all topics at a high level.