Emory Lundberg
@emory
Context Engineer, Threat Model Architect, Photographer, Writer, Parent
Emory Lundberg
@emory
Context Engineer, Threat Model Architect, Photographer, Writer, Parent
A list of useful resources about Templater:
@GitMurf quick demo How to setup and run your first Templater JS script: #187
@shabegom How To Use Templater JS Scripts: https://shbgm.ca/blog/obsidian/how-to-use-templater-js-scripts
@chhoumann Templates showcase: https://github.com/chhoumann/Templater_Templates
@zachatoo Templates showcase: https://zachyoung.dev/posts/templater-snippets
@lguenth Templates showcase: https://github.com/lguenth/obsidian-templates
@tallguyjenks video: https://youtu.be/2234DXKbNgM?t=1944
@ProductivityGuru videos: https://www.youtube.com/watch?v=cSawi0tYPMM

Model performance from Gödel

sweet
New DreamMachine with multigig and a 10Gb SFP+ WAN uplink
Don't know if you're still looking. AMD recently released support for rocm for the HX 370 and it let me get ollama going with the GPU. I did a small benchmark using gpt-oss:20b and got 16.5 tokens/sec for a simple prompt. Once I reallocated 32G VRAM to the GPU and started the ollama server with:
HSA_OVERRIDE_GFX_VERSION=11.0.2 OLLAMA_DEBUG=1 ollama serve
ollama psshowed "100% GPU" and I got ... 20.5 tokens per second.So, not a huge uplift but it was only using ~100% cpu rather than ~1100% as before.
You can expect snappy, "ChatGPT-vibes" performance on models up to 20B, with user benchmarks reporting 15 to 27 tokens/second using the Radeon 890M iGPU. For larger models like DeepSeek 70B there aren't specific benchmarks, but you will absolutely need 96GB of RAM since memory bandwidth is the primary bottleneck and the iGPU shares system memory. Currently, the NPU isn't a major factor for these LLMs due to immature software support, so performance relies entirely on the iGPU.
Upvote1DownvoteReply AwardShare
OP•1mo ago
I know that DeepSeek 70B requires a lot of RAM. The problem is that if it generates, for example, 1 token/second, then it's not usable at all. And it makes no sense to take a version with a lot of RAM for it. Is there any data available on the performance of DeepSeek 32B?
Upvote1DownvoteReply AwardShare
Only the sparse MoE models will generate decent speeds on this machine if you go large. gpt-oss 20B, gpt-oss 120B, Qwen3 30B A3B for example. A dense 70B/72B model at Q4 will produce less than 2 tok/s, a 32B at Q4 will produce about 3-4 tok/s.