China has finally entered the graphics card race. NVIDIA sells cards with the same specs for $10k. Is NVIDIA finished?
>>106436296Model size?
>>106436159>Is NVIDIA finished?what's the driver support for the Huawei card like?
>>106436159>Memory: LPDDR4Xsouds slow
imagine the difference a couple of decades will make in the gpu market
>>106436159Is this an official listing from Huawei or is this some third-party seller?
>>106436364Fine for LLMs.
>>106436159No active cooling? Probably just a toy.
>>106436296>>106436407amd 395+ has 196gb vram and is 2.2x than a 4090 while using 87% less power.Its over for Nvidia
It's going to take at least 10 years for the Chinese to develop their software. I wouldn't bother with it unless the price was like a fraction of Nvidia
I heard chinese ai companies tried huawei but ditched them for nvidia again. Big vram number isnt the whole story
The more competition the better, but westoids will seethe anyway
>>106436309you weren't suppose to mention that... westoids...
>>106436432a lot of people are vendor locked to cuda. using zluda gets you sued by nvidia. huawei and amd dont have anything as robust as cuda yet. their alternatives only work on a fraction of their products while cuda is designed to work on anything with a cuda core from laptop, to desktop, to datacenter.
>>106436500Imagine thinking chinks (and major chink corporations at that) are even remotely worried about being sued by Nvidia.
>>106436500>huawei and amd dont have anything as robust as cuda yet.AMD next design, UDNA, is supposedly coming to challenge that, unifying everything. I even abandoned getting a 9000 series to wait for it given my 6600 is still serviceable, AI workloads notwithstanding.
>>106436407LPDDR4X is trash compared to GDDR7 which is trash compared to HBM3E we're talking about scraps from the bottom of the barrel. they don't want you to have fast local ai
>>106436159I want a 10 usd GPU for LLM.
>>106436432>I heardnice source shlomo
>>106436159the future is chinese i told you
>>106436500>using zluda gets you sued by nvidiajust operate your business outside of globohomo controlled countries
Genuinely excited for affordable chinese gpus to hit the market. Might take a few years but even if they end up flawed they will hopefully force amd and nvidia to stop fucking around.
bing chilling
Probably using old as fuck ram because the NPUs are tiny 8W chips (it has 4)
>>106436374It's not time that's the issue, it's a monopoly. Asians love monopoloies.Just look at the way they got together to raise the prices of lcd tv's.
>NPU not GPU>so bad DeepSeek had to abandon even trying to use them>dumped onto Aliexpress for less than the VRAM chips costkys dumb wumao
>>106436419>10 yearsno it doesn't, all they need to do is put LLMs to work and verify it against automated test suite
>>106437991>NPU not GPUyou don't get it do you, doing NPU they don't give a shit about all the gpu compatibility legacy shit, one could argue intel could move much faster with their gpus if they didn't do a gpu just gone straight to npu
>>106437991>>so bad DeepSeek had to abandon even trying to use themisn't that just first gen? there are always some kinks to hammer out. those are certainly good enough for hobbyists
>the ram is more expensive~250 dollars in ram. (ie MT53E2G32D4DE-046 on AX)
>>106436432>I heard chinese ai companies tried huawei but ditched them for nvidia againi would soiquote you if i wasn't permabanned from posting imageshow's the weather in tel aviv?
>>106436500>a lot of people are vendor locked to cudaand this is why I've hated cuda from the beginning.
>>106436159>fake specs>fake driversbased!
>>106438270I wouldn't expect Huawei to post fake numbers, although I would expect a Chinese reseller to do that.But after you mentioned it I visited their site, because vram is one thing but performance is another.I'm not sure if I'm surprised or not but it mentions nothing about performance such as flops or anything of the sort. https://support.huawei.com/enterprise/en/doc/EDOC1100285916/181ae99a/specifications
>>106436329I wouldn't be surprised if it had better loonix support
>brigade of jeets is here to tell us china can't make softwarepretending tiktok, genshin, and every other sorta impressive western face app/program doesn't exist only works if you're a 76 iq jeet who eats cow dung cakes. china has their own exclusive ecosystem of programs all natively developed and used by themselves which make western enshittifed trash look like a joke. the reason why nvidia/amd have software issues is cause they hire jeets. it's like how boeing has planes dropping out of the sky cause they hire jeets.this will be mass produced and sold to corporations in china by the end of 2026. by 2027 it will be hitting consumers in china. the only question is if they'll sell it internationally in any real supply which is difficult to say cause their own demand will be high and they don't want to give any to jew controlled nations
>>106438367>Ascend 320>16 TOPS@INT8 >8 TFLOPS@FP16 >8 W of power consumption.
>>106438420PS. the card has multiple Ascend 320, x2 (?) for this one and x4 for another card.It's a low power accelerator.
>>106436159>with the same specsBut are they really? Does this thing actually perform the same? Are its drivers as compatible as NVIDIA's? Show some benchmarks.
>>106436416>Its over for NvidiaNow they'll have to stop going overboard with the pricing. Finally.
>>106436440>but westoids will seethe anyway*brainrotted incels
>>106438489I, for once, believe the beatings will continue
If software support for it arrives then it will start to become a game changer
>>106436159as long as it doesn't have any software support its useless. Same problem as with AMD their software stack fucking sucks their docs are terrible and whatever tools they have are buggy or only work with loads of finetuning. Every last piece of chinese hardware has this issue idk when they will fix it. I would have thought that a commie government would force them to be more open but instead you get AI generated bullshit.
>>106436159>>106436364It's only 204 GB/sMight as well use system RAMToo slow for AI
>>106436419stealing code doesn't take that much
>>106436159>trusting Chinese components
>>106436382https://e.huawei.com/cn/products/computing/ascend/atlas-300i-duohttps://e.huawei.com/cn/products/computing/ascend/atlas-300v-pro
it's trashhttps://www.artificialintelligence-news.com/news/deepseek-reverts-nvidia-r2-model-huawei-ai-chip-fails/
>>106436159>>106441170are these GPUs? why are they marketed for IA?
>>106442442>are these GPUs?yes>why are they marketed for IA?I don't know
>>106436432>I heardYou... heard? Are you sure you didn't make all that up?
>>106436159What can it be used for? Anything beyond a chatbot?
>>106436419They can vibe-dogfood it.
>>106436416>AI optimized card better optimized for AI than gaming optimized card.More news at 11
>>106442442no one care about gpus outside of gamers (2% of nvidia revenues)
>>106436416>amdlmao. when the ai frenzy will be over, amd stock will halve again.
>>106442449>GPUNo, it's an NPU, it can't do FP32 and it isn't compatible with graphics APIs. It is for 'large' models inference, faster than a cpu but only as fast as a 5050 for AI, the advantage is the ram size.
>>106442797Ah, thanks for the clarification
>>106441128Surely China will stay behind forever, as they are obviously racial incapable. Oh wait, TSMC is Chinese.Same was said with Korean and Japan though.
>>106436159This is a threat to our democracy!! we need more tariffs, orange man
>>106442972Has anything really become more expensive lately because of this tariffs program?I dont live in US but it sounds like a shitlib whining
>>106442948Thankfully they slid into totalitarian dictatorship, which historically haven't performed that well in capital allocation.
>>106436412ah yes, the a100 is also a toy
>>106436412based retard
>>106436159>3000$ cardits half of an H100, let me guess it has 1/4 the performance
>>106444570Post where are selling H100s for $6000
>>106436416Shit metric because the 70B model can't fit inside the 4090 VRAM so it's partiall running on probably the slowest DDR4 RAM they could get their hands on.2.2x 2t/s is just 4.4t/s. Both are useless but the 4090 will run a 32B model at 50t/s and the AMD cpu shit will barely get 5.AMD will win once they put 12 memory controllers in consumer desktop cpus. They'll never do that though.
>>106436419you know they became competitive with the west since deepseek r1, right? still lagging behind on hardware for like 3 or 4 years but that's a far cry from the 10+ years they were at 8 years ago >>106436432nvidia straight up cancelled a lot of production since chinese companies stopped almost all nvidia hw imports
AMD will win.MLID released a leak that AMD will release a consumer AI card that uses LPDDR5X. Which then due to the density of the modules will enable them to release consumer cards with 512gb of VRAM.
>>106447699CAMM or gtfo
>>106447127>you know they became competitive with the west since deepseek r1, right?Yes, using Nvidia GPUs and possibly data mined from American models. >still lagging behind on hardware for like 3 or 4 years but that's a far cry from the 10+ years they were at 8 years ago A lot was done, it's true, but take a look at how long it took for Nvidia to build its moat. To think that China, for all the money poured into these projects, will magically get something competitive in a few years is ludicrous. I'm not even bearish on them, the Chinese are proving to be some tough, smart bastards, but they simply don't have the culture of building software, especially one that must comply with international relations, which these GPUs will have to play their part if they want to truly become competitive (I say this because customer feedback is crucial). I'm sure they get there, though. 37B is no fucking joke.
>>106436159Does it support Cuda or ROCm? If not it's trash
>>106447699>Which then due to the density of the modules will enable them to release consumer cards with 512gb of VRAM.Yeah for a small price of $10000+ lmao I don't buy any of this bullshit. AMD won't even sell their current AI Workstation cards direct to consumer. You have to go through OEMs or used market to get your hands on them they are so anti-consumer.
>>106442678The thing is it's not optimized. It just has more RAM so it can fit the model in memory. 70B LLMs suck ass, too. There's no real way to make small LLMs perform better because they just contain less information. It's like trying to get a retard to do math
>>106436159Open source drivers? Only ML bullshit?
>>106448390>70B LLMs suck ass, too.idk man, i've had some good success with qwen3-coder 30B A3B. Especially for auto completion you don't need insane models.
>>106448390AI is more than LLM
>>106448515in particular than generalist LLMs*
>>106436159Are those any good for img or video gen?
>>106447699lmao. amd behaves like a fucking monopolist already for years, they don't give a fuck about cards 3 generations back, all amd is good for fucking universities in europe that don't do anything important because it's cheaper than serious nvidia hardware
>>1064476992 more weeks
nvidia will never be replaced. they know what they have which is why they're priced so high.
>>106451178Yeah, just like Intel. Except Nvidia can't even build their own stuff.
I kneel.In 10 years Chinks are going to be a major player in these markets, as their systems are going to be a fraction of the price and they have proper chip production going on.I bet they'll get plenty of state subsidies to keep their own prices low, simply to fuck with Nvidia's pricing.
>>106436159people like you don't seem to understand, that a lot of the valuation of nvidia comes from stuff like cuda (and many more).Nvidias APIs are so powerfull, france wanted to fee over it.Nvidia investef decades into this infrastructure and the frsmeworks.Nvidia could only scramble if nvidia decides to let "ai"-powered jeets take over the sources...
>>106443954china's capital allocation is massive investment into a field almost akin to feeding a river of piranhas with hundreds of small companies fighting viciously for the top. japan's is the softbank gook giving away whatever's left of japan's real wealth to america.I wish I could feel bad for the japs, but they're genetically gooks whom I despise so instead I'm only wishing for their demise. don't give me the jomon cope btw, those are the natives wiped out by gook colonization and then ruled by an elite class from china, which is why the grammar is korean and the vocab is chinese (and american, kek)
>>106448390>70B llms suck assyou sound like you tried them like 2 years ago>>106451285kek, one of the biggest reasons intel is almost going bankrupt exactly because they are using their own fab
>>106436416Can it run Crysis?
>>106453488It can do better than that: it can generate new Crysis.
>>106436159Explain why should I NOT buy it over Nvidia or AMD?
>>106454021Very unpatriotic.
