3 ms·

AMD has a new chip for LLMs, right? Can that also be used for things like Blender rendering? I like the idea of having plenty of general purpose cores, but afte

by ilaksh 3y ago

AMD has a new chip for LLMs, right? Can that also be used for things like Blender rendering? I like the idea of having plenty of general purpose cores, but after around 16 it starts to feel like I need the AI stuff more now. It seems inefficient to have all of the parallel stuff like graphics/AI in what amounts to basically a second computer that plugs in. I am guessing that Apple has a lead in this stuff. Are Intel or AMD copying them at all?

brucethemoose2 3y ago
> I am guessing that Apple has a lead in this stuff. Are Intel or AMD copying them at all? Not exactly. The AMD chip you are thinking of is the MI300, and it was originally a giant server APU with 24 cores. It was marketed as a HPC supercomputer chip. Intel had something similar, Falcon Shores, a big APU. ...And as it turns out, almost no one cared. The ML industry doesn't want unified memory APUs, an inference chip, a training chip, a bigger training chip, and so on. They want Nvidia GPUs. Everything is already coded to work with this "second computer that plugs in," as you said, so thats what customers want. Even Nvidia themself seems to be spinning their wheels with grace-hopper. So AMD spun a MI300 into a variant with an even bigger GPU and no CPU. And they emulated CUDA (a work in progress). Intel canceled Falcon shores and redid it as a pure GPU, with their other AI IP (Gaudi) baked in like tensor cores. With a 2 year delay. And they are moving heaven and earth to support PyTorch and other frameworks out-of-the-box. I can't emphasize how disruptive this is, especially for Intel. Silicon takes years and 9+ figures to design. They would not have made these moves if there was even modest short-term interest in unified memory chips for AI.
brucethemoose2 3y ago
Also, forgot to mention, AMD does have an M3-Pro like chip in the pipe: https://hothardware.com/news/amd-strix-halo-cpu-rumors https://hothardware.com/news/amd-strix-halo-cpu-rumors Intel probably does too, but the rumors are even less solid.
washadjeffmad 3y ago
If you mean Ryzen AI, a product of AMD's Xilinx acquisition: https://www.amd.com/en/developer/resources/ryzen-ai-software-platform.html https://www.amd.com/en/developer/resources/ryzen-ai-software... "Select models of [AMD Ryzen™ 7040 series processors] offer the new XDNA architecture with built in AI Engines, the first dedicated AI processing silicon on a Windows x86 processor." It seems to be a more fully featured parallel to Intel's current GNA: https://github.com/intel/gna https://github.com/intel/gna Apple Silicon Macs only have a lead at a substantial markup over computers without modern dGPUs, and if Intel has a lead anywhere in hardware, they're keeping it well hidden. We're not saying that too loudly, though, because latest gen AMD prices could stand to be a little more competitive.

by ilaksh 3y ago

brucethemoose2 3y ago

> I am guessing that Apple has a lead in this stuff. Are Intel or AMD copying them at all? Not exactly. The AMD chip you are thinking of is the MI300, and it was originally a giant server APU with 24 cores. It was marketed as a HPC supercomputer chip. Intel had something similar, Falcon Shores, a big APU. ...And as it turns out, almost no one cared. The ML industry doesn't want unified memory APUs, an inference chip, a training chip, a bigger training chip, and so on. They want Nvidia GPUs. Everything is already coded to work with this "second computer that plugs in," as you said, so thats what customers want. Even Nvidia themself seems to be spinning their wheels with grace-hopper. So AMD spun a MI300 into a variant with an even bigger GPU and no CPU. And they emulated CUDA (a work in progress). Intel canceled Falcon shores and redid it as a pure GPU, with their other AI IP (Gaudi) baked in like tensor cores. With a 2 year delay. And they are moving heaven and earth to support PyTorch and other frameworks out-of-the-box. I can't emphasize how disruptive this is, especially for Intel. Silicon takes years and 9+ figures to design. They would not have made these moves if there was even modest short-term interest in unified memory chips for AI.

brucethemoose2 3y ago

Also, forgot to mention, AMD does have an M3-Pro like chip in the pipe: https://hothardware.com/news/amd-strix-halo-cpu-rumors https://hothardware.com/news/amd-strix-halo-cpu-rumors Intel probably does too, but the rumors are even less solid.

washadjeffmad 3y ago

If you mean Ryzen AI, a product of AMD's Xilinx acquisition: https://www.amd.com/en/developer/resources/ryzen-ai-software-platform.html https://www.amd.com/en/developer/resources/ryzen-ai-software... "Select models of [AMD Ryzen™ 7040 series processors] offer the new XDNA architecture with built in AI Engines, the first dedicated AI processing silicon on a Windows x86 processor." It seems to be a more fully featured parallel to Intel's current GNA: https://github.com/intel/gna https://github.com/intel/gna Apple Silicon Macs only have a lead at a substantial markup over computers without modern dGPUs, and if Intel has a lead anywhere in hardware, they're keeping it well hidden. We're not saying that too loudly, though, because latest gen AMD prices could stand to be a little more competitive.