Deepseek was big because not only did they publish the full model for everyone to use, but the MoE structure significantly brought down the hardware requirements in terms of processing power. As long as you have enough VRAM, you can run it on older hardware with no need for the latest Nvidia stuff.
Now they got v4 which many have found to be within a 10% margin of Claude and ChatGPT.
On top of that, China has cheapo VRAM GPUs available or soon to be released, like the MTT S80. Yeah it sucks as a Graphics card because the chip is behind, but you get 16Gb of GDDR6 for much cheaper than anything else.
But its not a conspiracy to fight China. The infinite scaling was just Nvidia solidifying themselves as the monopoly because they want all AI infrastructure to be dependent on them, which is why they still illegally export to China, despite an export ban attempting to reduce their potential competition.
Moore Threads (MTT) already has their own CUDA like system called MUSA, and I’m sure they’ll be happy to put in proper hardware support for new stuff like Bf16 and FP8/4. It’ll take a few years, but eventually China will catch up to the point where Nvidia gets shanked by cheaper hardware.
Wasn’t there development of a linux translation layer for CUDA workloads to run on AMD GPUs? I haven’t heard about it in a while, but I’d imagine that’d help the situation.
You mean ZLUDA? AMD gave up on it and released the project on github.
Personally I dream for CUDA API implementation in open source GPU drivers on Linux. That would absolutely level the playing field for the whole industry
MTT is just a pipe dream, last I checked. But Deepseek is actively being served, in mixed FP8/FP4, on racks of Huawei accelerators.
I believe Baidu trained a model on them, too. But most training (like Deepseek’s) is still done on CUDA.
…Also, be careful equating this stuff with any kind of “consumer friendly” hardware you or I could buy. That’s less likely. The Huawei accelerators (and other local Chinese hardware experiments) are geared towards huge servers serving requests in parallel.
In a way it has actually.
Deepseek was big because not only did they publish the full model for everyone to use, but the MoE structure significantly brought down the hardware requirements in terms of processing power. As long as you have enough VRAM, you can run it on older hardware with no need for the latest Nvidia stuff.
Now they got v4 which many have found to be within a 10% margin of Claude and ChatGPT.
On top of that, China has cheapo VRAM GPUs available or soon to be released, like the MTT S80. Yeah it sucks as a Graphics card because the chip is behind, but you get 16Gb of GDDR6 for much cheaper than anything else.
But its not a conspiracy to fight China. The infinite scaling was just Nvidia solidifying themselves as the monopoly because they want all AI infrastructure to be dependent on them, which is why they still illegally export to China, despite an export ban attempting to reduce their potential competition.
Moore Threads (MTT) already has their own CUDA like system called MUSA, and I’m sure they’ll be happy to put in proper hardware support for new stuff like Bf16 and FP8/4. It’ll take a few years, but eventually China will catch up to the point where Nvidia gets shanked by cheaper hardware.
Wasn’t there development of a linux translation layer for CUDA workloads to run on AMD GPUs? I haven’t heard about it in a while, but I’d imagine that’d help the situation.
You mean ZLUDA? AMD gave up on it and released the project on github.
Personally I dream for CUDA API implementation in open source GPU drivers on Linux. That would absolutely level the playing field for the whole industry
MTT is just a pipe dream, last I checked. But Deepseek is actively being served, in mixed FP8/FP4, on racks of Huawei accelerators.
I believe Baidu trained a model on them, too. But most training (like Deepseek’s) is still done on CUDA.
…Also, be careful equating this stuff with any kind of “consumer friendly” hardware you or I could buy. That’s less likely. The Huawei accelerators (and other local Chinese hardware experiments) are geared towards huge servers serving requests in parallel.