Since the beginning of the fight between AMD and NVIDIA Immortals for the high-end market, GPUs under 2000 yuan can play almost none. It wasn't until Intel launched Intel Arc graphics cards in 2022 to return to the discrete graphics market, and the Xe-Core-based GPU was born, and the XE-HPG architecture entered the consumer discrete graphics market, making the confrontation between the two in the market switch to a three-legged stand.
Gaming graphics cards are not meant to be built. This not only requires advanced manufacturing processes and efficient large-scale parallel processing capability design, but also needs to be recognized by game engines and practitioners to obtain excellent compatibility on PC platforms. In particular, it is even more difficult to break through the industry barriers of building a game ecosystem with NVIDIA CUDA Cores, Tensor Cores, RT Cores, and Game-ready drivers.
Even so, Intel ARC still uses advanced design concepts and Intel's own strong industry appeal and R&D capabilities to allow ARC to continue to upgrade, and it is also quite outstanding. Take our protagonist today, Intel ARC A750, as an example, the OC overclocking version is only 1649 yuan. To be reasonable, what more bikes do you want.
It is not enough to have **, only sufficient performance can be worthy of the description of cost performance. Nearly a year and a half after the iCard debuted, what kind of progress has been made in games, creation, AI, and driver optimization, and is the cost of less than 2,000 yuan worth it now? Now let's take the Gunnir Intel ARC A750 Photon 8G OC W as an example.
The Alchemist appears
In XE-HPG, the GPU codenamed Alchemist was the first to be put on the market stage, and according to the plan, the Intel ARC brand will include Alchemist, Battlemage, Celestial and Druid in the future.
In terms of design, the XE-HPG is not an expanded version of the XE-LP used in the previous 13th Gen Core GPUs, but has a new design, that is, the XE-Core core is introduced. Xe-Core can be thought of as a collection of vectors and tensor ALUs, with L0 and L1 cache units. Logically close to XE-LP subslicing, NVIDIA SM (Streaming Multiprocessor). If you know a little about GPUs, you may know that the unit level is not set in stone, for example, NVIDIA has modified the SM level when updating the architecture.
Each XE-Core will contain 16 Vector Engines (VE) and 16 XE Matrix Extensions (XMX). Each of these vector engines can process 256 bits per cycle. If you break it down, each vector engine contains 8 FP32 ALUs, which is about the same as the XE-LP EU. Since the 16 vector engines are capable of handling 128 FP32 operations per clock, which is an FMA throughput of 256 FLOPS, it is also the same SM as the NVIDIA Ampere GPU in terms of throughput per clock.
In XE-Core, every 16 vector engines are paired with 16 matrix engines for matrix and tensor computation, and here Intel uses a proper noun for its name, XE Matrix Extensions, abbreviated XMX, which shows its importance. XMX is mainly used for AI-accelerated, matrix tensor computation, and each XMX engine uses an 8-depth pulsation array. XMX performs eight 512-bit-wide matrix calculations per clock cycle. These vector and matrix engines are supported by a wide load memory unit that can retrieve 512b of data per clock cycle, while each xe-core has 512kb of data cache for l1.
Although SM and Xe-Core match in vector throughput, Intel has twice the throughput of matrix operations and can perform twice as much as ALUs, which means that Intel GPUs still tend to invest more resources in matrix computing and AI computing.
On top of XE-Core, the logic of XE-HPG is Render Slices, which, like XE-LP, provide most of the functionality to Intel GPUs. For Alchemist, a slice contains 4 xe-cores, 4 ray tracing units, 4 texture samplers, geometric rasterization frontends, and 2 pixel backends. This 4:4:4 layout means that within the Alchemist GPU, each XE-Core has its own texture sampler and ray tracing unit.
Since the Alchemist GPU contains up to 8 slices, the full GPU state contains 32 Xe-Cores, 4096 FP32 ALUs, supports DirectX 12 Ultimate, and has an XMX matrix engine. And then cut down from this to form a discrete graphics card product with different positioning.
This is the case with the Intel ARC A750. It uses the GD2-512 GPU, codenamed ACM-G10, built on TSMC's 6nm process, with 21.7 billion transistors and a core area of 406 mm. Compared to the A770, just one render tile unit is removed, and the 7 render tile units have a total of 28 xe-cores, 28 ray tracing units, 448 xmx engines, and a base frequency of 205GHz, the highest frequency can reach 24ghz，tdp 225w。
Not only that, as one of the AIC manufacturers of Intel ARC GPU, the Blue Halberd also adds a lot of color to the A750. As a review, the Gunnir Intel ARC A750 Photon 8G OC W adopts a white exterior design that is more in line with the aesthetics of the white host, and with a set of multi-dimensional cooling system called ICICLE, it can better ensure the stable performance of the graphics card.
For example, the trirotor fan itself supports intelligent start-stop technology, which can effectively control the heat dissipation noise of the graphics card, and five nickel-plated heat pipes + high-density heat dissipation fins can provide a good heat dissipation auxiliary effect. In the stress test state during the actual test, it can be seen that the GPU core temperature is up to 58, and the external temperature of the graphics card is concentrated in the power supply part, and the temperature is about 45 in a room temperature environment of 20.
At the same time, the Gunnir Intel ARC A750 Photon 8G OC W power supply part uses a dual 8-pin design, which is well compatible with ATX 30 pre-power supply design.
In terms of interface configuration, the Gunnir Intel ARC A750 Photon 8G OC W provides enough for 1 HDMI 21 and 3 DisplayPort 20, which means that the ARC A750 can also support 8K resolution output on the interface.
The game driver continues to be optimized
Now let's get into the actual combat link, here the test platform uses Core i9-14900K, iGame Z790D5 Ultra, iGame DDR5 16GB 6800*2 Ultra W as a reference, and mainly focuses on 1080P highest image quality, as well as 3DMark benchmark for testing.
In the 3DMark benchmark test, 3DMark Time SPY, 3DMark Time Spy Extreme, 3DMark Fire Strike Extreme, 3DMark Fire Strike Ultra, and Port Royal were used as references, which slightly outperformed the GeForce RTX 3060 12GB at the base level.
The game session takes us one step further. At the beginning of the article, we mentioned that over time, the Intel ARC GPU driver has become more and more compatible with the game. In January 2024, the Arc graphics card driver ushered in a major update again, and the latest driver Game On supports a variety of new games, and also brings different performance improvements to more than 20 popular DX11 and DX12 games. Here we use 310.101.4972 drive with the latest 310.101.A comparison of the 5333 drive shows a significant improvement in just three months.
In the case of Just Cause 3, the new driver has increased by more than 160% at the highest image quality of 1080p, and the game has changed from basic smoothness to running at a high level of more than 170fps. Civilization 6 is a noticeable boost, with frame rates up more than 35%, and Dying Light 2: Stay Human is also impressive, allowing this parkour game to run easily at over 100fps. At the same time, Apex has a decent increase.
For example, in the newer triple-A masterpiece of "Cyberpunk 2077", the frame rate increase brought by the new driver has reached more than 40% under the premise of turning on high-end ray tracing at 1080p. At the same time, you can also see that Xess Super Sampling can be directly enabled in the settings interface.
XEss Super Sampling technology is similar to the fiery NVIDIA DLSS, AMD FSR, through a series of AI optimization algorithms, at the cost of lower computing resources, in exchange for higher performance and image quality. Similar to DLSS, it is a technology that combines space and time to improve AI images, that is, it uses a combination of spatial data (adjacent pixels) and temporal data (the vector of moving objects in the previous frame) to learn from a neural network.
In fact, Intel has been working on the ARC brand for a long time before announcing it, and has already optimized hundreds of games to make the ARC A750 run more and more smoothly on old and new titles.
For example, in Counter-Strike 2 and Atomic Heart, you can see a 15% increase, while Hunt: Showdown can see a boost of more than 35%.
It is also worth mentioning that the intuitive control panel of the Intel ARC driver not only provides a cool interface for game organization, but also integrates multiple functions such as broadcasting, capture, and wonderful time capture. You can also turn on the pinned performance panel to monitor the performance of your GPU while the game is running.
AI is dazzling, and new experts in creation
The powerful parallel processing power makes the GPU itself very suitable for content creation work, and the Intel ARC A750 also deliberately focused on content creation, AI acceleration, **1 codec and other aspects when designing the XE Core and XMX engine. To take the most intuitive example, Intel ARC's **1 encoding and decoding capabilities are very strong, even if you use D**Inci Resolve to encode a 2-minute 12GB 4K footage, the actual use time can even be faster than that of GeForce RTX 4090.
At the same time, we also use RTX 3060 Ti and RTX 4090 against H265 format output for comparison, you can also see that the Intel ARC A750 is really fierce.
The advantage of *1 can also be used directly in game streaming and streaming, because **1 encoder is better than hThe 264 is more efficient, and at the same bandwidth or volume, **1 can show more clear details. Here we use a live recording of **1 and **c of "Counter-Strike 2" for comparison, in the same scene, it can be clearly seen that the outline of the building and the gun body of **1 stream** is clearer.
And in the Procyon Benchmark, we can also see the comprehensive performance of Intel ARC A750 in **processing and**processing,It is quite good。
In the Blender Benchmark rendering output, the three output scenes of Moser, Junkshop, and Classroom are mainly detected, and the performance is as follows, which is on par with the RTX 3060.
Common specviewperf 2020 for engineering majors. This is a professional software graphics test in the fields of energy exploration, medicine, architectural design, mechanical design, automotive design, and aircraft design, including mainstream software such as 3dsmax, catia, creo, energy, maya, medical, snx, and solidworks. The Intel ARC A750 already runs smoothly in most professional software.
Finally, the Intel ARC A750 also has good AI performance. Here we take stable diffusion as an example. Stable Diffusion is an AICG tool for deep learning text-to-image conversion launched in 2022, which was developed by the startup Stability AI in collaboration with non-profit organizations and academic staff, so it is more open and extensible than Midjourney, which requires a fee, and provides a series of plugins to achieve more functions, such as AI** repair, text prompt guided images, and even image translation, etc. It is foreseeable that more powerful features will continue to be incorporated in the future.
The premise of stable diffusion is that at least 8GB of VRAM and a GPU with strong AI performance are required, otherwise the local experience is not as straightforward as purchasing cloud services. By directly obtaining one-click running resources at station B, Intel ARC A750 can easily run stable diffusion with a Chinese interface, even for novice players, configuration is no longer a problem.
Here we use a fixed text description to guide Stable Diffusion to create 20 architectural landscapes that match the description**. Set ARC A750 to calculate 2 sheets at a time** in the UI interface, and run it 10 times in total, that is, 20 sheets. The resolution of each ** sheet is 512x512 resolution, the number of sampling steps is set to 50, and the sampling method is selected as Euler A.
The text description is as follows:
beautiful render of a tudor style house near the water at sunset, fantasy forest. photorealistic, cinematic composition, cinematic high detail, ultra realistic, cinematic lighting, depth of field, hyper-detailed, beautifully color-coded, 8k, many details, chiaroscuro lighting, +dreamlike, vignette
In actual use, it can be seen that the efficiency and quality of the ARC A750 output ** are good, and it only takes 2 minutes and 14 seconds to complete the generation of 20 **, with an average of 67 seconds to generate a **, according to the algorithm of images per minute, the formula is 60 (totaltime (batchsize * batchcount)) = images per minute, and the final generation efficiency is 895 images per minute @ 512x512 is a very good performance, especially as a GPU at a price of 1649 yuan, the performance is very outstanding.
Written at the end: a cost-effective creative tool
The advantage of Intel ARC A750 is its powerful multi-processing power, especially the powerful **1 codec ability, even against the flagship GeForce RTX 4090. And with Intel's continuous driver optimization, the player's gaming experience is improving day by day, and the performance of Intel ARC A750 in mainstream games is also becoming more and more mature, and some game scenarios can get more than 2 times the performance improvement after updating the game driver, under the current positioning of 1649 yuan, it really makes people feel that they have made a lot of money.
If you want to take AI performance to the next level, I also recommend considering the Intel ARC A770 with 16GB of video memory, which is not only cost-effective, but also makes it more impressive in AI performance with larger video memory.
In short, we have seen Intel's sincerity in the GPU ecosystem, consumer applications, and games from the Intel ARC A750. Under the premise of limited funds, you can get the latest GPU technology, and there are many highlights in many application scenarios, coupled with the third-party design and good heat dissipation performance of the Blue Halberd, Intel ARC A750 can be written into the list when it is installed.