AI drawing relies on the powerful computing power of the graphics card to process a large amount of data and run corresponding algorithms to generate the required images. At first, it was based on CUDA, which was once the exclusive domain of N cards. And with the unremitting efforts of the big guys, we can finally eat ready-made directly, let us thank the Buddha for feeding @秋葉aaaki、AI King Kong @Nuullll。 Thanks to their efforts, Intel graphics card users can get almost the same application experience as N-card users, and even run AI graphics with the latest Ultra processors.
There are three ways to deploy the corresponding environment: the first is to install WSL directly, the second is to install it through Docker, and the third is to directly copy jobs and use the integration package. The operation of the first two methods is too complicated: WSL needs to install the Linux subsystem after opening HyperV, and after installing the corresponding driver and Python, you can call Intel Arc graphics card runway through Intel Extension for Pytorch (IPEX), the setting here is to step on the pit, and Docker is also installed in the WSL system, but Docker has various problems. The easiest thing is to choose the modpack, so let's thank the Buddha for feeding the rice again @秋葉aaaki, AI King Kong @Nuullll.
1.A580与A770对比
When the price of the graphics card was still high, I bought the Arc A770 Photon OC of the blue halberd, and I have been happily using it for more than a year. The aforementioned environment deployment methods are all lessons of blood and tears, and the performance of the A770 cannot be fully invoked, and nothing can be satisfied. After the holiday, I needed a short card to plug the unit on the machine, so I directly made an Arc A580 Index with a blue halberd to play.
The A580 is a 253mm long dual-fan card, which is a pretty OK size. The exterior design style is basically the same as that of the A770 Photon, which is a little "rougher" than the naked eye, after all, the initial price is twice as bad. But to be honest, even if it's a little rougher, it's much better than the appearance quality of the beggar-version RTX 4070 from many manufacturers. Of course, this is a thousand-yuan graphics card, I didn't have much expectations at first, but I actually measured it and found that it was still worth the price.
Both the Arc A580 Index and the Arc A770 Photon are all-metal backplanes that provide support and protection for the graphics card PCB. The A770's metal backplate is no surprise, after all, the price and size are there. A580, a dual-fan thousand-yuan dessert card, is also a surprise to the metal backplate, and I won't criticize some graphics cards by name here. The workmanship and materials of Intel Arc series graphics cards are very low,
Like the A770, it is a 2*8Pin power supply, the base power of the A770 is 194W, and the base power of the A580 is 134W. Arc series graphics cards reserve nearly 30W or so of the official soft super space, and the power of the Arc A580 in the official driver can be pulled up to W, which is actually enough for a single 8Pin power supply.
The comparison chart of the parameters of Intel A-series desktop graphics cards is as follows, which comes from the product comparison on Intel's official website. It can be seen that the A580 has 8 fewer Xe cores than the A770, and the number of other unit components under it also conforms to this ratio. Of course, the frequency of the graphics card is more of a reference to the power consumption allocation, and it is not excluded that the A580 can exceed one by pulling up the power.
I won't say much about the appearance, just go straight to the test part and look at the watch. The corresponding test screenshots will not be released separately, and the score of the A770 is about 30% ahead of the A580, which is slightly lower than the difference between the two architecture levels. Judging from the test score of 3D Mark11 alone, the A580 should be the same as the RTX 4060, of course, the price gap is here, and the actual performance must be poor. Similar to the situation of the A770 counterpoint RTX3070 and the RTX 4060Ti, the price gap and the performance gap naturally exist.
XeSS is Intel's DLSS-like feature, which can effectively improve the frame rate performance of the picture. The test option is set to a performance file with a resolution of 1920*1080, and the frame rate can basically be increased by more than 40%. If you observe some scenes under the performance file, you can see the difference with the naked eye, and you can basically ignore the immersion in the game.
There is basically no big difference in the PCIe bandwidth between the two, only 200MB/s, which exceeds the 256GT/s communication rate corresponding to PCIe 4.0 x8 (that is, about 15.75GB/s effective bandwidth). In this way, both cards need to be under the PCIe 4.0 x16 specification to play full blood performance, but in fact, even if the specifications of the PCIe slot are not enough, it will lose 1-2% at most, which is basically negligible, and the old platform can also upgrade this card.
Through the benchmark test of Shadow of the Tomb Raider, four whole sets of tests were given to Arc A770 and Arc A580, and the API was DX12, which was tested at 1440p and 1080p resolutions, respectively. Each set of tests is divided into three sub-sections and records data: Disable XeSS, enable XeSS High Profile, and enable XeSS Performance Profile. It can be seen that except for a few cases, the A770 at 2K resolution is basically 120%-125% of the performance of the A580, and the A770 at 1080P resolution is basically 115%-120% of the performance of the A580
2.AI plot comparison
Based on the default settings of the Eishi Launcher 2.8.1 version, the first round of AI mapping test was carried out, with a single batch of single sheets, a resolution of 512*512, and an iteration of 20 steps. Since the time to display the overall average speed is too short to take a screenshot, it is necessary to take a screenshot when you are relatively late, so you can take a screenshot when you want to do it later to ensure that the data is relatively stable. The iteration speed of the Arc A770 reached 6.23 steps per second, and the 8G video memory was not completely full.
Using the exact same steps and prompts, the Arc A580 can generate images, and the iteration speed can reach 6.13 steps/second, and the same does not eat up to 8GB of video memory. The speed performance is slightly lower than that of the Arc A770, and it is not far away from the specifications of bit width, GPU frequency, and number of processing expectations, as well as in terms of power consumption at full load. Of course, under this load, the two graphics cards are not full, and if you need to quantitatively measure it, you still have to record the real-time power and the power consumption of a single plot.
So the next test has to be full of load to see how much the performance of the two can be stretched. After testing a single batch of 10 cards, I still couldn't eat up the video memory, so I directly set it at a single batch of 16 cards and iterated 20 times, and the video memory occupation was directly full. Going back to the test, the Arc A770 was able to run at a speed of 1.41 seconds per step, and the attentive reader will notice that the data units are directly flipped, so be careful to take the reciprocal when comparing the data.
Looking at the Arc A50 under the same setting, the iteration speed is 1.73 seconds per step, and the iteration speed of the Arc A770 is about 18% ahead of the Arc A580 after the weighted average data, which is basically in line with the performance gap between the two in other dimensions. In the future, I will borrow a 16G Arc A770 for comparison, and you can more accurately and quantitatively test the AI mapping performance.
3. Summary
This card is only ¥1299, and a brand new graphics card is only at this price, which really completely kills the RTX 3050 8G at ¥ 1700, the RX6600 at ¥1500 and the A750 at the same price. Mining cards have started to rise wildly, and the prices of new graphics cards in other grades are also rising, and this card is definitely the best choice for new accessories in the lower ¥3000 range.
It just so happens that my colleague also wants to install a machine to play games, and the daily life is CS2 and Naraka, and the budget in hand is about 3,000 yuan, and I don't want second-hand accessories, and I want a small chassis. All things considered, it's still 3I platform to save trouble,After that, it's a late 3A masterpiece,This card is also enough at 1080P resolution。
If you think this article is helpful to you, please like it and collect comments. This is the old pot, we'll see you next time!