I bought a graphics dock.
The specific situation is like this,Last year, several domestic handheld manufacturers didn't all come out with AMD mobile graphics cards as the core of the graphics dock,At that time, the pricing of such products was generally around 5000 yuan,Although it is much cheaper than ROG's graphics card dock,But it is still sold at the price of a complete notebook/Windows handheld,How many people are willing to support it can be imagined。
This year, however, things have changed a little. The emergence of a large number of open-source graphics dock manufacturers has begun to squeeze out the already small market space, and the existence of MXM graphics docks has enhanced the market position of portable graphics docks. Affected by this, a large amount of inventory of the old GPD G1 graphics card dock appeared on Xianyu, and the price dropped to about 3500 yuan, which is already in the price range that I can accept.
(Source: GPD)
So I bought a graphics dock.
What I didn't expect was that after I bought it, the price of the old GPD G1 graphics dock was still steadily declining, and now the second-hand price went straight to 2900 yuan.
But anyway, since I've already bought this graphics card, I'm definitely going to experience it. So the question is, what kind of experience can this domestic graphics dock bring us? Is the so-called "portable" selling point meaningful? Let's take a look.
There is metal, no texture
From the perspective of appearance design,Compared with the streamlined design of the No. 1 graphics cardONEXGPU,GPD G1The appearance design can only be said to be square and concise,The front metal panel adopts the design of horizontal strips and oblique strip texture collision,There is a short horizontal strip-shaped heat dissipation hole in the middle of the fan,GPD logo is designed in the lower left corner,At least the metal materials are sufficient。
But if you want to talk about texture,Then GPD G1It's really not,Whether it's new or oldGPD G1It looks like a simple docking station or power bank or something,The dark gray metal coating will even make the rare metal panel with a hint of polycarbonate flavor,It can only be said that you can't really ask too much for small factories。
In the promotion of this product, one of the key words that GPD likes to advertise is "portable".
In fact, the GPD G1 should indeed be the most portable graphics dock out there. The measured size of this product is 225 x 111 x 29.8 mm, and the weight is about 0.92kg, which is similar to the volume of a 20000mAh power bank.
Of course, there are no products on the market that are smaller than the graphics dock alone, but GPD G1 is the only graphics dock with integrated built-in power supply in the same volume, with a built-in 240W gallium nitride power supply, although it may have a slight impact on heat dissipation, but it is a good level of integration of the whole machine.
(Source: Lei Technology)
The question that remains is, how many people will go out with a graphics dock?
I'm not good at making an evaluation.
In terms of interfaces, GPD G1 supports both USB 4 and Oculink (SFF-8612) dedicated data interfaces. Among them, USB 4 has a theoretical transmission speed of 40Gbps, of which about 16-22Gbps is for data, which can take into account charging, storage expansion and graphics card connection at the same time, but the performance loss is high, while Oculink can have a theoretical transmission speed of up to 63Gbps in the case of PCIE 4*4, with less performance loss and loss, and can give full play to the real strength of GPU.
(Source: Lei Technology)
Looking at the back of the fuselage, the back interface of GPD G1 is quite rich, it can be said that it is not just a graphics card dock, but also takes into account the functions of the docking station. From left to right, there are power jacks, 3 USB 3.2 Type-A, 1 SD 4.0 card reader, 2 DP 1.4a, and 1 HDMI 2.1 port.
(Source: Lei Technology)
However, in actual use,This HDMI 2.1 interface can only achieve the highest 4K30Hz/2K60Hz picture output due to internal interference,In the case of connecting multiple large screens, it can be said that it is very chicken,This problem can be solved on the new version of the graphics card dock。
120W performance unleashed, comparable to the RTX 3060
Next, let's see how the GPD G1 graphics dock performs.
The GPD G1 graphics dock has a built-in AMD RX7600M XT mobile graphics card, this graphics card is designed with AMD's latest RDNA 3 architecture, with 32 computing units, 64 raster units, and a total of 4096 stream processors, with a default frequency of 2300 MHz, a boost frequency of 2615 MHz, and a maximum power consumption of 120W, with 32MB InfinityCache and 8GB GDDR6 video memory, and a video memory bit width of 128bit.
Should I say it or not,In the actual test of mobile graphics cards,AMD is really like a stranger now,There are basically no other brands in China except Xuanpai that will use AMD graphics cards to make game laptops,This also casts a veil of mystery over the performance of this RX7600M XT,RDNA 3How much can the architecture improve the mobile graphics card,Which Nvidia graphics card can the performance of this graphics card be benchmarked,This is a question I want to figure out。
(Source: Lei Technology)
In order to conduct a comprehensive test, this time I have prepared two products, Huawei Matebook E 2022 with Thunderbolt 3, Huawei Matebook E 2022 with Intel Iris Xe core graphics and USB 4, Oculink, and GPD Win mini 2023 with AMD 760M core graphics.
Take a look at 3DMark Timespy, and you can see that when using the Oculink+ external display graphics card to connect directly, this RX7600M XT can run nearly 10,000 graphics points, which is between RTX 3060M and RTX 4060M, using USB 4 external display and using Oculink internal display will result in 5% performance loss, and using Thunderbolt 3 performance loss of 25%.
(Source: Lei Technology)
Enter the game test,First of all, the test of "Call of Duty: Warzone",Under the 1920*1080 quality picture quality,The average frame rate of the game directly connected to the Oculink+ external screen graphics card can reach97fps,The average frame rate under other connection methods decreases in turn,But 760M This core display can have about 30fps performance is quite surprising。
(Source: Lei Technology)
Take a look at "Horizon: Dawn of Zero", under the 1920*1080 ultimate picture quality, the two core graphics cards are directly disarmed and surrendered, and after using Oculink+ external screen graphics card to connect directly, the average frame rate of the game can reach 55fps, so that as long as you open an FSR, you can play at 60 frames.
(Source: Lei Technology)
Looking at racing games, in F1 2023, we turned on 1920*1080 high image quality, and conducted a sunny day test on the Australian circuit, and finally the average frame rate of the Oculink+ external screen graphics card directly connected can reach an astonishing 179fps, this kind of high frame rate game can indeed better reflect the gap in graphics card performance.
(Source: Lei Technology)
The same racing game, "Forza Horizon 5" is going to eat more performance, we turn on the 1920*1080 highest image quality for performance testing, and finally the average frame rate of the direct connection of the Oculink+ external screen graphics card can reach 66fps, and the loss caused by different connection methods is relatively small in this game.
(Source: Lei Technology)
As a standard test game, Shadow of the Tomb Raider didn't miss out on this test. Without FSR turned on and the default is 1920*1080 maximum image quality, the average frame rate of the Oculink+ external screen graphics card can reach 126fps, and even in the case of maximum loss, it can still have a performance of more than 60fps.
(Source: Lei Technology)
Finally, there is the much-anticipated "Cyberpunk 2077", in the case of turning on the 1920*1080 non-ray tracing ultimate preset, the average frame rate of the direct connection of the Oculink+ external screen graphics card can reach 91fps, but the performance will be greatly reduced after turning on ray tracing, which is also an old problem with AMD graphics cards.
(Source: Lei Technology)
It can be seen that after connecting to the external GPD G1, whether it is an Intel 2-in-1 or an AMD handheld computer, it can greatly improve the smoothness of 3A games, and the performance of this RX7600M XT at 100W power consumption is enough to surpass RTX 3060M, although there is still a 15%-40% performance gap between the RTX 4060M with the highest 140W power consumption release (depending on the user's connection method), but it is already very capable of playing in similar products.
However, in the actual experience, there are still some game graphics cards with low occupancy, and occasional instantaneous frame drops occur, which need to be further improved by AMD's driver.
Thin and light laptops can also play with local large models
Next, the old rule, how to say it now, is also the application period of the AI large model, this has a graphics card that does not run the large model, just like buying a graphics card but not trying to mine when the mine tide, then it will definitely not work.
However, almost all of the large models on the market today are based on Nvidia's CUDA. Under normal circumstances, if you want to run a local large model, either you buy an Nvidia graphics card, or you use the general computing power of the CPU to run hard, since this is an AMD graphics card, then you should also know that it is not a simple thing to run a large model on it.
The good news is that although CUDA is exclusive to N cards, this does not mean that A cards cannot be accelerated for AI inference, and through DirectML, GGML, or OpenCL, A cards can bypass this restriction and use GPUs for AI inference.
Here, I directly choose LM Studio, a third-party solution recommended by agricultural enterprises, which is a closed-source, cross-platform and cross-vendor solution, and you can use AMD graphics graphics power by downloading the "LM Studio for AMD ROCm" version.
(Source: Lei Technology)
After the LM Studio software is installed, the next step is to choose the right model, there are currently two mainstream models, namely Mistral 7b and LLAMA v2 7b, the former has a good experience in novel long text generation, and the latter is a more general-purpose local large model.
Next, we switch to the chat tab, and then, scroll down in the right panel, find the corresponding model file, and click Download and install. Then switch to the chat tab, select the model from the drop-down menu in the middle of the top, and wait patiently for it to finish loading.
(Source: Lei Technology)
If you have installed the AMD graphics card version of LM Studio, please select the "GPU Settings" on the right panel, move the slider all the way to "Max", if the GPU type detected by LM Studio shows "AMD ROCm", the configuration is successful and you can chat happily.
(Source: Lei Technology)
From the actual experience, the experience of LM Studio is quite good, except that the local model is not compatible with Chinese, which is a bit worrying, basically the dialogue can be generated in one second, whether it is answering questions, translating or writing is very fast, and the content of the reply is basically free of logical errors, can be self-consistent.
(Source: Lei Technology)
It is worth mentioning that LM Studio has many adjustable options, which can realize a series of settings such as conversation temperature, context range, and reply repetition, so that as the user's understanding continues to deepen, the experience of communicating with large voice models will become more layered.
Let's talk about AI painting again, at present, the Chinese integrated Webui launcher has been launched with ZLUDA support, which supports all kernels including SD-WebUI and ComfyuUI, without manually replacing files or setting environment variables, which can be said to be a one-stop simple experience.
(Source: Lei Technology)
Unlike the current Stable Diffusion OpenVINO integration package for GPUs and CPUs, this version of SD-WebUI has complete functions, all kinds of Lora, plug-ins, and phrase presets can be used RX7600M XT acceleration can generate four 20-step rendering, 512*512 images in about two minutes, which is quite efficient.
The actual picture looks like this:
(Source: Lei Technology)
Summary:
In practice, the overall experience of the GPD G1 graphics dock is somewhat contradictory.
If your computer has a USB 4/Oculink port and needs this kind of high performance when working from a fixed location, then the GPD G1 graphics dock is really worth considering, as its compact design doesn't take up much desktop space and provides decent performance with a stable connection.
From a performance point of view, although the performance of the RX 7600M XT after connecting through the USB 4/Oculink interface has been lost, it still performs well, and it is not much of a problem for thin and light laptops with integrated graphics cards and Windows handhelds, whether it is 3A games or local large models, and it is not a big problem to run various picture editing, video rendering, and even 3D modeling software.
(Source: GPD)
If your thin and light laptop doesn't have a USB 4/Oculink port, then I don't recommend trying it. It can be seen that in the case of connecting the Thunderbolt 3 interface, the GPD G1 performance loss of the graphics dock is very significant, considering the retail price of this product, the loss of 25% of the performance will make its practicality close to the RTX 3050 8GB priced at only 1499 yuan, and the RTX 3050 8GB itself is the notorious IQ tax.
As a final reminder, as a graphics dock, the GPD G1 is quite limited in terms of use. If you have the capital to own both thin and light laptops, you don't need to consider this kind of product at all, but if you only have a Windows handheld, 2-in-1 device, or thin and light laptop, it can still help these devices effectively improve performance, which is an ingenious idea to make the most of it.
From April 25th to May 4th, the Beijing International Automobile Exhibition (Beijing Auto Show) is in full swing. The theme of this year's auto show is "New Era and New Vehicles", which is the vane of "automobiles from electrification to intelligence".
Lei Technology's "pay attention to electric vehicles, understand more intelligent" account of Tram Tong has sent a luxury reporting team to the Beijing Auto Show site to conduct first-line professional reports, welcome to pay attention to "Tram Pass" to get first-hand professional on-site reports.