corpus.blog
41554 blogs · [ { "id": "01a087ec-c20c-72c6-a237-4a6b7ba9b237", "title": "I Recommended the 96 GiB VRAM Carveout. Then a 21 GiB Model Hung on Load.", "url": "https://thefrontierlab.ai/i-recommended-the-96-gib-vram-carveout-then-a-21-gib-model-hung-on-load/", "published_at": "2026-08-27T11:09:10+00:00" }, { "id": "01a087ec-c20c-72c6-a237-4a6b7c2c0862", "title": "Speculative Decoding Across My Fleet: Gains From +200% to -80%, and No Single Setting Fits Every Model", "url": "https://thefrontierlab.ai/strix-halo-mtp-fleet-sweep/", "published_at": "2026-08-20T15:54:23+00:00" }, { "id": "01a087ec-c20c-72c6-a237-4a6b7cd3d29a", "title": "Quantized KV Cache Helps Six of My Nine Models and Actively Hurts the Two Newest Ones", "url": "https://thefrontierlab.ai/strix-halo-nine-models-kv-cache-split/", "published_at": "2026-08-13T11:28:23+00:00" }, { "id": "01a087ec-c20c-72c6-a237-4a6b7d9cc46a", "title": "I Rebuilt the Box, and the Cliff Was Gone. Then I Spent an Evening Proving Which Patch Did It", "url": "https://thefrontierlab.ai/strix-halo-dense-cliff-patch-attribution/", "published_at": "2026-08-06T15:03:31+00:00" }, { "id": "01a087ec-c20c-72c6-a237-4a6b7e0b24bb", "title": "A Reader Fixed Quantized KV Cache on Strix Halo. I Verified It at 262k Context, and His Theory About My Dense Cliff Survived Half Its Test", "url": "https://thefrontierlab.ai/strix-halo-quantized-kv-cache-verified/", "published_at": "2026-07-30T15:40:04+00:00" }, { "id": "01a087ec-c20c-72c6-a237-4a6b7e4f1c1b", "title": "Turning Off the IOMMU Made My Dense Model 37% Faster. Five More Strix Halo Tips, Measured, and One Crashed the GPU", "url": "https://thefrontierlab.ai/strix-halo-tuning-part-two-iommu/", "published_at": "2026-07-23T12:08:03+00:00" }, { "id": "01a087ec-c20c-72c6-a237-4a6b7f4d41e6", "title": "I Measured the Strix Halo Tuning Tips. One Flag Matters More Than All of Them", "url": "https://thefrontierlab.ai/strix-halo-tuning-tips-measured/", "published_at": "2026-07-16T10:20:15+00:00" }, { "id": "01a087ec-c20c-72c6-a237-4a6b7fcf52ca", "title": "One llama.cpp Update Made My MoE Model 38% Faster and My Dense Model 45% Slower", "url": "https://thefrontierlab.ai/one-update-38-faster-45-slower/", "published_at": "2026-07-09T11:05:23+00:00" }, { "id": "01a087ec-c20c-72c6-a237-4a6b7fe9a2be", "title": "The Strix Halo Reference Setup Pack", "url": "https://thefrontierlab.ai/setup-pack/", "published_at": "2026-07-06T23:32:29+00:00" }, { "id": "01a087ec-c20c-72c6-a237-4a6b8031b56f", "title": "Buy or Wait: Reading the Local LLM Hardware Question in a Memory Crunch", "url": "https://thefrontierlab.ai/buy-or-wait-local-llm-hardware/", "published_at": "2026-07-02T11:57:00+00:00" }, { "id": "01a087ec-c20c-72c6-a237-4a6b80f65480", "title": "What to Buy for Local LLM Inference: Strix Halo, Mac Studio, DGX Spark, or a GPU Rig", "url": "https://thefrontierlab.ai/strix-halo-buyers-guide/", "published_at": "2026-06-25T12:00:30+00:00" }, { "id": "01a087ec-c20c-72c6-a237-4a6b81234dc6", "title": "AMD Is Selling \"First-Class ROCm\" on Strix Halo. I've Run the Same Chip for Six Months.", "url": "https://thefrontierlab.ai/first-class-rocm-strix-halo-six-months/", "published_at": "2026-06-18T06:00:36+00:00" }, { "id": "01a087ec-c20c-72c6-a237-4a6b81c9dd64", "title": "A BIOS Update Won't Fix #6182 — I Tried the Newest One", "url": "https://thefrontierlab.ai/bios-update-wont-fix-6182/", "published_at": "2026-06-11T12:16:47+00:00" }, { "id": "01a087ec-c20c-72c6-a237-4a6b81e63270", "title": "Full Context on a Vulkan-Only Strix Halo: The Decode-Drop Reproduces, but the Sweet Spot Moves", "url": "https://thefrontierlab.ai/full-context-vulkan-only-strix-halo/", "published_at": "2026-06-04T06:00:38+00:00" }, { "id": "01a087ec-c20c-72c6-a237-4a6b8273092a", "title": "MTP Defaults Are a Trap: What 260 Runs Showed About Speculative Decoding on Qwen3.6", "url": "https://thefrontierlab.ai/mtp-defaults-are-a-trap/", "published_at": "2026-05-28T06:00:08+00:00" } ] posts Claim your blog
Back to thefrontierlab.ai
Blog · corpus.blog/blogs/thefrontierlab.ai/posts

thefrontierlab.ai

thefrontierlab.ai

2026

I Recommended the 96 GiB VRAM Carveout. Then a 21 GiB Model Hung on Load.original ↗
27 Aug 2026
Speculative Decoding Across My Fleet: Gains From +200% to -80%, and No Single Setting Fits Every Modeloriginal ↗
20 Aug 2026
Quantized KV Cache Helps Six of My Nine Models and Actively Hurts the Two Newest Onesoriginal ↗
13 Aug 2026
I Rebuilt the Box, and the Cliff Was Gone. Then I Spent an Evening Proving Which Patch Did Itoriginal ↗
6 Aug 2026
A Reader Fixed Quantized KV Cache on Strix Halo. I Verified It at 262k Context, and His Theory About My Dense Cliff Survived Half Its Testoriginal ↗
30 Jul 2026
Turning Off the IOMMU Made My Dense Model 37% Faster. Five More Strix Halo Tips, Measured, and One Crashed the GPUoriginal ↗
23 Jul 2026
I Measured the Strix Halo Tuning Tips. One Flag Matters More Than All of Themoriginal ↗
16 Jul 2026
One llama.cpp Update Made My MoE Model 38% Faster and My Dense Model 45% Sloweroriginal ↗
9 Jul 2026
The Strix Halo Reference Setup Packoriginal ↗
6 Jul 2026
Buy or Wait: Reading the Local LLM Hardware Question in a Memory Crunchoriginal ↗
2 Jul 2026
What to Buy for Local LLM Inference: Strix Halo, Mac Studio, DGX Spark, or a GPU Rigoriginal ↗
25 Jun 2026
AMD Is Selling "First-Class ROCm" on Strix Halo. I've Run the Same Chip for Six Months.original ↗
18 Jun 2026
A BIOS Update Won't Fix #6182 — I Tried the Newest Oneoriginal ↗
11 Jun 2026
Full Context on a Vulkan-Only Strix Halo: The Decode-Drop Reproduces, but the Sweet Spot Movesoriginal ↗
4 Jun 2026
MTP Defaults Are a Trap: What 260 Runs Showed About Speculative Decoding on Qwen3.6original ↗
28 May 2026