Footer

Your go-to destination for discovering trending open-source projects and uncovering the insights that matter

Legal

  • Terms and conditions
  • Privacy policy

More

  • Pricing
  • Advertise
  • onedump
  • gosnakego
  • Contact us

Social

  • X (Twitter)
  • GitHub
Trendshift owl
What they build speaks louder than words
© 2026 Trendshift, created by Julian Li with ❤️
SignalAdvertise
Trendshift
SignalAdvertise
Sign in
Trendshift owl

Discover trends that matter

SignalAdvertise
DailyWeeklyMonthlyYearly
Live mentionsTopics
RepositoriesDevelopers
InsightsStats

OfinTech Desk
OfinTech Desk
Busbar
Busbar
It's a plan
It's a plan
Omnigraph
Omnigraph
Bifrost
Bifrost

Lurker17/fast-kernels

Putting down fast kernels here, planning for ampere, hopper, blackwell++

Visit GitHub
Python
2 contributors
FeaturedAdvertise to 180K+ monthly visitors
OfinTech Desk

An open-source AI workspace for finance and accounting teams, in which every figure is traced to its source

Busbar

Control what AI can do before it acts

It's a plan

Open-source Linear alternative

An issue tracker where AI agents take the work you assign them

Omnigraph

Versioned graph database on object storage. Git-style workflows and multimodal retrieval

Bifrost

Fastest Enterprise AI Gateway for Routing, Governing, and Securing all your AI Traffic

Momentum

How fast Lurker17/fast-kernels is gaining stars, and where it lands on Trendshift and GitHub Trending

Activity and trending

Stars, forks and merged pull requests each day or month, lined up with the repository's rank on Trendshift Trending

Nov 8 – Nov 19, 2026

Stars +0 in 12 days

01

Forks +0 in 12 days

01

Merged PRs 0 in 12 days

01

Trendshift Trendingwhere the repository ranked each day it made the top 25

Not in the Trendshift top 25 in this period

Nov 8Nov 10Nov 12Nov 14Nov 16Nov 18
The left and right arrow keys step from day to day; Home and End go to the first and the last

Stars, forks, merged pull requests and Trendshift Trending rank, by day
DayStars gainedForks gainedMerged PRs
Sun, Nov 8, 2026+0+00
Mon, Nov 9, 2026+0+00
Tue, Nov 10, 2026+0+00
Wed, Nov 11, 2026+0+00
Thu, Nov 12, 2026+0+00
Fri, Nov 13, 2026+0+00
Sat, Nov 14, 2026+0+00
Sun, Nov 15, 2026+0+00
Mon, Nov 16, 2026+0+00
Tue, Nov 17, 2026+0+00
Wed, Nov 18, 2026+0+00
Thu, Nov 19, 2026+0+00
Total+0+00
  • 1#1 on the list
  • 7Rank that day
  • Not trending

Social mentions

Recent discussions about this repository across the web
  1. Nov 191 post
    • Each kernel on its own is actually decent. I'm using one block per row, with 128 threads. All loads and stores are 128-bit, and each thread issues all the loads for its row before doing any math. The…

      Read full on X, posted by @lurker670291

      @lurker670291 · x.com

  2. Nov 181 post
    • Day 10 of writing a Hopper kernel: FP8 block-scaled GEMM, a TMA-store epilogue. 1155 TFLOP/s, 91.5% of DeepGEMM. The epilogue was a normal bf16 store: each consumer thread converts its accumulator to…

      Read full on X, posted by @lurker670291

      @lurker670291 · x.com

  3. Nov 171 post
    • There was another bottleneck in the consumer. After prefetching, address calculation and loop bookkeeping accounted for 38% of consumer samples. I also noticed that shared-memory bank conflicts were…

      Read full on X, posted by @lurker670291

      @lurker670291 · x.com

  4. Nov 151 post
    • Day 7 of writing a Hopper kernel: FP8 block-scaled GEMM(8192^3), first pipeline. The overall structure is still the same as the BF16 kernel: one producer warpgroup doing TMA loads into a 4-stage…

      Read full on X, posted by @lurker670291

      @lurker670291 · x.com

  5. Nov 131 post
    • With multicast, rank 0's TMA load also writes into rank 1's shared memory. So rank 0 cannot reuse a stage until consumers on both SMs have finished reading it. This is why empty[s] needs to be a…

      Read full on X, posted by @lurker670291

      @lurker670291 · x.com

  6. Nov 121 post
    • Day 4 of writing a Hopper kernel: GEMM (8192³), persistent kernel (650 TFLOPS -> 690 TFLOPS). Our occupancy is bounded by 1 block per SM after the two consumer changes, which increased the tile size.…

      Read full on X, posted by @lurker670291

      @lurker670291 · x.com

  7. Nov 111 post
    • Looking at what the consumer does on every K-step, there's quite a bit of work that doesn't involve Tensor Cores: Wait for the producer (full[s]). Build WGMMA descriptors and fence. Issue and commit…

      Read full on X, posted by @lurker670291

      @lurker670291 · x.com

  8. Nov 101 post
    • How do we choose the size of this arrangement? Let's say we design a swizzle consisting of p rows and 132/p columns of thread blocks. Since the H100 has 132 SMs, this gives us a group of 132 CTAs,…

      Read full on X, posted by @lurker670291

      @lurker670291 · x.com

  9. Nov 91 post
    • WGMMA is Hopper's Tensor Core instruction. All 128 threads in a warpgroup participate in issuing it. The A and B matrices are read directly from shared memory, while the 128×128 FP32 accumulator is…

      Read full on X, posted by @lurker670291

      @lurker670291 · x.com