Hydra kernel evidence

Poison Green Packet

GPU-grouped public template for Hydra kernel packaging, benchmark, capacity, quality, and claim-boundary evidence.

Public framing: Hydra is a bounded-residency decode attention kernel. These rows support package coverage, selected capacity unlocks, tested quality parity, and exact-Qwen fit/usability evidence. They do not support universal speedup or universal FlashAttention replacement claims.

RTX 3060

Small-card capacity and 32k RULER parity.

package smoke0.2574 ms
HF mean0.3229 ms
capacityqualityboundary

RTX 3070

Kernel package and HF benchmark coverage.

package smoke0.1474 ms
HF mean0.2532 ms
packageHF benchmark

RTX 3080

Package coverage plus negative revolver boundary.

package smoke0.2051 ms
HF mean0.3157 ms
packageHF benchmarknegative diagnostic

RTX 3090

Capacity row and exact-Qwen FP8 evidence through 14k.

package smoke0.1492 ms
HF mean0.3107 ms
capacityexact-Qwenfit

RTX 4070 / 4070 Ti

4070 capacity/quality rows; 4070 Ti package rows.

package smoke0.1261 ms
HF mean0.2215 ms
capacityqualityhardware split

RTX 4090

Qwen3-4B 32k/65k RULER quality parity rows.

package smoke0.1132 ms
HF mean0.2245 ms
qualityRULERstress

A100 SXM4

80GB capacity row through bounded T=327,680.

package smoke0.1408 ms
HF mean0.2568 ms
capacitypackageHF benchmark

RTX A6000

HF benchmark and final kernel-builder gate.

builder smoke0.2166 ms
HF mean0.3230 ms
builderHF benchmarkno package row

RTX PRO 6000 Blackwell

Blackwell package/HF rows and exact-Qwen frontier data.

package smoke0.1158 ms
HF mean0.1371 ms
exact-Qwenfrontierheadroom