I am planning to build a new LINUX PC in September and am considering a build that not only updates my current Intel setup but also offers significantly faster performance. My heavy load primarily involves running Ollama with Llama 3.1 on a 4060 GPU. Although I don’t frequently play games, my current CPU struggles with Stellaris during the late game stages, so afaik cache is important there
I’ve noticed that many AMD processors come with an integrated GPU. Would opting for such a CPU be beneficial even if I intend to install a 4060 GPU in my new setup?
the iGPUs found in AMD CPUs are acceptable low end GPUs depending on task. But they are more or less to give you a display out but for most workloads I’d suggest a discreet GPU. To help put this in context. All 6 x86 computers in my house are AMD systems, 5 of which have an iGPU, and only the headless servers are using the iGPU as the PCI lanes in those machines are better used by NICs.
the correct question it would be, if open youtube, email, office stuff the IGPU will handle that but if i open stelarris everything will get moved to the 4060 and the benefit for me is saving power ? or the 4060 idle consumption and consumption increase for doing email, youtube, office stuff is negligible i am better picking a CPU that dont have a IGPU
If you plan to add a 4060 gpu, selecting an AMD CPU with an integrated gpu may not be necessary. While integrated gpus are useful as a backup or for simple tasks, your dedicated 4060 gpu will handle demanding workloads like running ollama with lama 3.1 and gaming in stellaris more efficiently. As a result, choose an AMD processor with a large cache and several cores, as these attributes will improve stellaris performance during demanding computational activities and late game situations.
Going off what @Caro_line44 said, the best cache availability is in the X3D chips. I’m not sure when AMD is planning on releasing their 9000X3D chips, but that gives you options. Whether you want to buy older 7000X3D stock that’ll be discounted but still performant or go right for the 9000X3D whenever it releases is up to you.
The question was asked about Llama 3.1, and there are several versions including and up to the 405b version. People don’t seem to be aware of this, but you can run the full 405b version on Vega GPUs. Vega has the unique ability to treat system ram as Vram. Although I would recommend using an MI50 or MI60 for this and not use Vega 56/64 as they don’t have PCIE gen 4 and thus have slower access to system memory. The Radeon VII is technically a Vega 20 card, but can’t use PCIE 4.0 either for some reason.
Note that you still need a minimum of 256GB of System ram to do this. Not as hard since a 128GB optane PMem is cheap as chips now days.
Other than that, the only Version a 4060 can handle is the 8b and that isn’t all that great imo. The IGPU in AMD CPUs is an option to extend VRAM to hold larger models, It’s limited and I don’t think it’ll work in tandom with your GPU. If you go that route, you need to beef up that system memory. No dimm slots unpopulated.
