Deschide articolul original într-un tab nou Schimbarea acestei opțiuni se aplică imediat și reîncarcă pagina.
LECTURA
Înapoi la articole →

The Frontier Lab

thefrontierlab.ai/rss/
Vizitează site-ul

Cel mai recent articol

Ultima verificare

Urmăritori 0 urmăritori
Articole: 15

  1. I Recommended the 96 GiB VRAM Carveout. Then a 21 GiB Model Hung on Load.

    (The Frontier Lab)

    In May I published a setup for this box that said, in so many words: give the iGPU the maximum, 96 GiB, because "the math is one-sided: more VRAM, more options", and run llama.cpp with --no-mmap, because "standard mmap behavior is incompatible with how St…

  2. Speculative Decoding Across My Fleet: Gains From +200% to -80%, and No Single Setting Fits Every Model

    (The Frontier Lab)

    Seven current models on the bench, five speculative paths that could actually be measured, one build, one evening.

  3. Quantized KV Cache Helps Six of My Nine Models and Actively Hurts the Two Newest Ones

    (The Frontier Lab)

    The usual advice is to quantize your KV cache because it is nearly free.

  4. I Rebuilt the Box, and the Cliff Was Gone. Then I Spent an Evening Proving Which Patch Did It

    (The Frontier Lab)

    Part four of the tuning series.

  5. 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

    (The Frontier Lab)

    Part three of the tuning series, and the most community-built one yet.

  6. Turning Off the IOMMU Made My Dense Model 37% Faster. Five More Strix Halo Tips, Measured, and One Crashed the GPU

    (The Frontier Lab)

    Part two of the tuning series.

  7. I Measured the Strix Halo Tuning Tips. One Flag Matters More Than All of Them

    (The Frontier Lab)

    Turning flash attention on at 32k context makes my MoE model 41% faster at prompt processing on Vulkan and my dense model 51% slower.

  8. One llama.cpp Update Made My MoE Model 38% Faster and My Dense Model 45% Slower

    (The Frontier Lab)

    Update, July 12: The numbers in this post are real measurements, but my interpretation of the first table didn't survive contact with the readers and a git bisect.

  9. The Strix Halo Reference Setup Pack

    (The Frontier Lab)

  10. Buy or Wait: Reading the Local LLM Hardware Question in a Memory Crunch

    (The Frontier Lab)

    On June 25, 2026, Apple raised the price of the 96GB Mac Studio from $3,999 to $5,299, and Tim Cook called the memory market a hundred-year flood.

  11. What to Buy for Local LLM Inference: Strix Halo, Mac Studio, DGX Spark, or a GPU Rig

    (The Frontier Lab)

    Ordered by the one number that actually decides it: memory bandwidth.

  12. AMD Is Selling "First-Class ROCm" on Strix Halo. I've Run the Same Chip for Six Months.

    (The Frontier Lab)

    A note on links: some hardware below is linked through affiliate programs, marked (affiliate).

  13. A BIOS Update Won't Fix #6182 — I Tried the Newest One

    (The Frontier Lab)

    A note on links: some hardware below is linked through affiliate programs, marked (affiliate).

  14. Full Context on a Vulkan-Only Strix Halo: The Decode-Drop Reproduces, but the Sweet Spot Moves

    (The Frontier Lab)

    A note on links: some hardware below is linked through affiliate programs, marked (affiliate).

  15. MTP Defaults Are a Trap: What 260 Runs Showed About Speculative Decoding on Qwen3.6

    (The Frontier Lab)

    Multi-Token Prediction (MTP) merged into llama.cpp main on May 16, 2026.