ajc9988

For stix halo, if you are doing local ai, it's insane. Gpt-oss 120b on windows and still getting 30t/s! You are still dealing with no Cuda, windows has no rocm for a couple more months, Linux is still spotty on npu support (but can be done). Wendell at level 1 tech was getting 33t/s on that same model same hardware. Thinking of pulling the board and putting it in a custom 3d printed case with a 280 aio. I'm excited for Medusa halo with the at3. Talking 384 bus with up to 512GB lpddr6, 26 cores, 48 CUs of rdna5. Still considering adding an nvme to occulink to pcie just to add an nvidia gpu to the mix. Developing an advanced RAG system with multiple models at the moment. And doing context coding. AI is awesome and the biggest pain in my butt all at the same time. It's like an intern that doesn't listen, but that you can't get the job done without the assistant or a much larger investment.

Thank you. I figured doing curve optimizer with the per core offset, followed by the curve shaper for different loads, then thinking of getting SOC stabilize the infinity fabric clocks (try to get 2200 fully stable, 2166 is without manually touching the voltage), all before doing ram overclocking. But it is his PC, not mine, so I am trying to step him through the overclock, test, validate process. He's excited for it. (might create my own profile and submit some benches, but yeah). Has anyone seen Prema recently? I have a locked down cheap Strix Halo mini pc ($1,700 for a 128GB variant; considering an NVMe to oculink to x4 PCIe for external graphics card). The development eco sucks, with Linux being a bit to set up to use the NPU, while Ryzen AI 1.5.1 is on Python 3.10, but the ROCm 7.0 preview is on 3.12, creating complications on Windows ATM for development, and issues (which can be overcome) on using the NPU (but you can develop for it) on Linux. But I can confirm it can run the new 120B parameter openAI model. Around 30T/s on windows. But I want more options in the BIOS, so wanted to talk more. Also may 3D print a SFF case for it and throw a better cooler on it.

BTW, I just helped my friend build a new PC, 9800x3d, MSI pro x870e-p, 32gb g.skill 6000 (the bundle from microcenter), and a 9070 XT. Lianli 207 case and arctic freezer 3 360 AIO. Is it better to do multiplier or try optimizing with the curve optimizer + curve shaper + bclk OC? What are good voltages for SOC, etc.?

Morning all. Lot's has happened over the past 4 years. To put this in perspective, the 9000 AMD cards are the polaris cards. That year, they did not bring the high end. And Moore's law is dead already mentioned and commented on leaks of a giant card in the next round of RDNA cards. Which is sorely needed. The real deal is we need MORE RAM. But I'm getting into local AI, so I am biased. Yes, part of the enemy. lol.

Boo! Ghost from the dead saying hi.