LX 7G100 proves hype trumps performance:
Lisuan Tech often markets the LX 7G100 as a competitor to the GeForce RTX 4060. However, reviews have revealed that the LX 7G100 didn't hit the performance goal. Instead, it's more along the lines of a GeForce RTX 3060, one generation behind the target, and two generations behind the latest GeForce RTX 5060. The issue was that Lisuan Tech priced the LX 7G100 like a GeForce RTX 5060 Ti 16GB; however, it seems the high price didn't impede its early adoption at all.
The LX 7G100 is only the beginning for Lisuan Tech, and June 18 will not only mark the highly anticipated restock of the Founders Edition but also serve as the official launch date for two new graphics cards: the LX Pro and the LX Ultra. The LX Pro specifically meets the demands of professional engineering applications, whereas the LX Ultra caters to cloud computing. Meanwhile, the LX Max, designed for creative professionals, has an uncertain launch date.
Lisuan Tech may be a startup in the graphics card market, but its leadership brings a wealth of experience and industry knowledge. Silicon Valley veterans Xuan Yifang, Kong Dehai, and Niu Yixin founded Lisuan Tech in 2021, all of whom had previously worked at the renowned but now-defunct S3 Graphics. It only took the company five years to put out a working graphics card that's competitive with models two generations behind. Everybody has to start from somewhere, and Lisuan Tech has a firm stepping stone.
(Score: 3, Insightful) by looorg on Friday May 29, @10:49PM (2 children)
Lets hope they can start to crank out gaming (or just normal desktop) graphics cards. Cause the current prices suck. It behind "two generations" behind isn't great for current aaa-gaming. But if they just make the card real quiet it could be great for a lot of other things. Lets just hope those current generation prices for previous generation performance doesn't stick. Still it's probably still cheaper then the second hand market ...
How is driver support?
(Score: 2) by Deep Blue on Friday May 29, @11:18PM
Probably tail light guaranteed. I don't remember how S3's driver support worked, probably ok, but that was 20 something years ago, so i wouldn't count on any of that now.
(Score: 2) by Bentonite on Monday June 01, @02:20AM
Driver support is terrible - all they offer is terrible proprietary drivers.
They could quite simply validly release their drivers as free software and then there would be a reason to purchase their hardware (after all, if it is possible for you or someone else to fix up the driver, then it can be ensured that the GPU is no longer e-waste, due to it having a possible life exceeding several decades - then the GPU is no longer such an environmental liability), but of course not.
(Score: 2) by Frosty Piss on Saturday May 30, @12:30AM (8 children)
Nvidia won't, can't hold the monopoly forever, there's every reason for competitors to catch up. And every reason for consumers of GOU's to look at alternatives. Especially with the "AI Psychosis" currently all the rage. Even slightly "slower" GPU's that are not Nvidia have the advantage that they are not Nvidia.
(Score: 0) by Anonymous Coward on Saturday May 30, @03:54AM (7 children)
You might not have noticed but AMD's GPUs are better than nVidia's now for everything except ray tracing
(Score: 0) by Anonymous Coward on Saturday May 30, @09:19AM (5 children)
(Score: 3, Informative) by Unixnut on Saturday May 30, @09:44AM (4 children)
Can AMD GPUs compile/run CUDA apps? Like it or not what keeps Nvidia on top is not only their HW, but the entire CUDA ecosystem, from the dev tools, first/third party libraries and support. They had first mover advantage and it has become an effective monopoly. I know AMD and "the rest" have OpenCL, but from what I remember it is not 1-1 compatible with CUDA (so you'd have to re-write your CUDA code, or develop two code-bases in parallel which few bother with). Pretty much all LLMs are written in CUDA that I have seen, which cements Nvidia in its position currently, even if other GPU card manufacturers provide better price/performance.
Although in a case of "your own success brings you down", if Nvidia powered LLMs do reach the point where they can correctly, efficiently and seamlessly refactor code, people may be able to bulk re-write all the CUDA software into OpenCL and thereby break Nvidia's hold on the market.
(Score: 2) by ls671 on Saturday May 30, @06:21PM (3 children)
If LLMs are written in CUDA, can you explain to me why I run them fine and efficiently on CPUs only with no GPU at all with pci 5 and fast memory (128 GB) and jlama with vector API? Maybe you mean inference engines like jlama is?
I suspect that as long as the inference engine supports AMD well, it could run those LLMs just as well as with Nvidia GPUs.
Everything I write is lies, including this sentence.
(Score: 2) by Unixnut on Saturday May 30, @10:55PM (2 children)
Because CUDA is a software abstraction layer, allowing to execute stuff on CPUs, (Nvidia) GPUs or a hybrid of the two. There is nothing technical stopping an implementation on say AMD GPUs. I suspect it is more a legal/IP type restriction.
(Score: 2) by ls671 on Saturday May 30, @11:29PM
Thanks! I checked on my side and you seem basically correct.
Everything I write is lies, including this sentence.
(Score: 2) by Bentonite on Monday June 01, @02:29AM
AMD has quietly funded a CUDA re-implementation - it of course is proprietary garbage, built on top of the ROCm proprietary garbage - but really only proprietary garbage is written in CUDA; https://www.phoronix.com/review/radeon-cuda-zluda [phoronix.com]
Although the software implementation of CUDA is fully proprietary and there are claims about muh imaginary property, the reason for not perfect support and porting being required the technical reason of how CUDA is mostly undocumented, with reverse engineering required to implement the undocumented parts.
(Score: 1, Flamebait) by Bentonite on Monday June 01, @03:14AM
Ironically, Nvidia GPUs are still better for the users freedom, as you can practically use old Nvidia GPUs with the free Nouveau driver, while the amdgpu driver does not work *at all* without proprietary software and for AMD GPUs, the best you can get is a native resolution 2D framebuffer with a hacked up radeon driver.
(Score: 2) by VLM on Saturday May 30, @02:36PM (1 child)
I searched around and to save ya all the time, it seems like at this time its a WinGPU, so windows desktops only.
Ironically, this might be "good" that its unusuable for linux servers because you can't run FOSS-ish AI loads if they keep the drivers non-FOSS so for "PC appliance gamers" this will work well. And siphoning off 30K windows gamers will make "real cards" cheaper for folks like us due to supply and demand.
Not like I need yet another experiment I don't have time for, or with summer arriving, yet more watts being dumped in my basement, but I've been toying with the idea of slapping some old graphics cards in my proxmox cluster and experimenting with GPU stuff. A VERY expensive hobby if I were using cutting edge cards but if I used old slop hardware from ebay it might be cheap and affordable. Never done this, looks interesting. I have done plenty of device passthru on proxmox with excellent results (mostly USB microcontroller and fpga dev boards) and I understand graphics card passthru is stepping up the game of complexity quite a bit... would be interesting if I had the time..
(Score: 2) by Bentonite on Monday June 01, @02:44AM
Ah yes, another case of "FOSS" being used to describe proprietary software.
Bonus points for considering only AMD or Nvidia GPUs that run proprietary software as "real cards".
After all, Linux is a proprietary kernel (it's not even 50% source available) and LLM & LIM loads are always proprietary stuff (LLM & LIM as a technology is inherently proprietary, as you don't get the source code of the important parts - only inscrutable binary blobs and the source code of the less important parts are often under a proprietary license too).
PCIe IOMMU is of similar complexity to usb passthrough - the PCIe device is best left unused by the host (so it doesn't need to be rebinded) and the hardware needs to support IOMMU and interrupt remapping without deciding to hang instead; https://pve.proxmox.com/wiki/Pci_passthrough [proxmox.com]
(Score: 2) by bzipitidoo on Saturday May 30, @06:13PM (1 child)
Ever since the start of the pandemic and all this cryptocurrency mining, GPUs have simply been too expensive for casual, home use. I don't care that much about superior gaming graphics. I don't want to spend $1000 on a dedicated GPU. $100, yes, okay. But not $1000+. No, not $500 either. I was flabbergasted to learn that the last GPU I got, waay back around 2010, an ATI Radeon HD5450, for a while was actually worth more than when I got it new! The machine it is in is terribly obsolete.
Instead of trying for dedicated graphics cards, I make sure the CPUs I get have integrated graphics. Seems to be the cost effective way to get at least a little 3D accelerated graphics.
(Score: 2) by Bentonite on Monday June 01, @03:12AM
Some online store sellers put ludicrous prices on items, because they figure that if they wait long enough, some idiot will actually pay that much.
The current ludicrous pricing of new dedicated GPUs, induces the purchase of fast enough and compatible used GPUs straight away in online stores - leaving only ludicrous priced GPUs.
The HD 5450 for example supports OpenGL 4.5 and will run any reasonable game that is the slightest bit optimized (it won't run completely unoptimised modern slop games, but would you want to run those?).
Any recent integrated graphics are just as fast or faster than old dedicated graphics, thus are all that is practically needed for a 3D accelerated screen or two or three (although motherboards with 3 display outputs are somewhat uncommon).