Skip to content
← All news

News1 min read

71 first places on SOL-ExecBench

KDA-generated kernels on the Databricks entry hold 71 of SOL-ExecBench's 235 per-kernel first places, more than any other entrant. It started in June with one 8×B200 node running unattended for a week.

On a single 8×B200 node, over about a week in June, KDA generated kernels that took 53 first places on SOL-ExecBench, NVIDIA's leaderboard that scores each kernel against its speed-of-light bound.The June count of 53 is Humanfia-reported; the board does not keep history. The counts below are read from the live board on 5 October. The run never really stopped. On the board today, the entry it submits under, Databricks, holds more per-kernel first places than anyone.

Per-kernel first places · v1.1 · 5 October 2026
  1. 1Databricks · KDA71
  2. 2SF Tensor47
  3. 3Infinigence AI39
  4. 4Geometric24
  5. 5RIAC18
Of 235 kernel leaderboards. Each kernel has one first place.

The firsts are not spread evenly. KDA is strongest on fused operations, where it holds 36 of 82, and weakest on the quantization and FlashInfer kernels, where specialists lead. Switch the baseline to set our count against any rival's.

First places by collection · v1.1 · 5 October 2026
L1 · single operations94 kernels
22
9
39
7
L2 · fused operations82 kernels
36
18
0
4
Quantization33 kernels
7
10
0
10
FlashInfer-Bench26 kernels
6
10
0
3
All four235 kernels
71
47
39
24

A collection is ranked by one SOL score over all of its kernels, so the most first places do not make the top spot. On those boards, the same entry is second or third everywhere, within 0.007 of the leader:

Collection leaderboards · SOL score (1 = speed of light, 0.5 = the scoring baseline) · 5 October 2026
L1 · single operations#30.7963Infinigence AI0.8032
L2 · fused operations#20.7921SF Tensor0.7923
Quantization#30.8861SF Tensor0.8916
FlashInfer-Bench#30.7962SF Tensor0.7992

One node, a week, nobody watching ​

The number to notice is not the count, it is the cost. Kernel optimisation has historically priced in a scarce human: someone who knows the architecture, reads the profiler and has the patience for the twentieth variant. This run cost a machine that was already there, and nobody watching it.

An unattended queue
queue8 × B200board

Swipe the diagram sideways to see all of it →

1/0 A queue of kernels to beat.

That changes which problems are worth attacking. A kernel that would take an engineer three days and probably yield nothing is not worth an engineer's three days. It is trivially worth a slot in a queue.

KDA · the board · #1 on L1 in July

Read next

News6 KDA PRs merged into SGLangSince August, SGLang has merged six pull requests carrying kernels our agents wrote, the best adding 8.7% end-to-end throughput. The MLSys contest kernels are public with a benchmark anyone can rerun. One merge elsewhere, Flash-KMeans, was reverted the next day.Read it BlogKDA²: Kernel Design Agents (KDA) Optimize Kimi Delta Attention (KDA)Our agents wrote Kimi Delta Attention kernels that run up to 2.96× faster than FlashKDA on B300 with a tenth of its state error. Here is how, and how the agents tried to cheat along the way.Read it News6.5× faster MSA prefill indexerOn B300, KDA found the hardware the MSA prefill and decode indexers were leaving idle: a 6.5× geometric mean on prefill, up to 14× on chunk tails, up to 3.3× on decode, and bitwise-identical output throughout. Humanfia-reported.Read it