# Tag

rdna3

Posts with this tag: 14

← Alle berichten
Generative media 7 min

The slow part was a memory layout, not a kernel

A strided tensor layout made AOTriton attention run 6 times slower than on the same data made contiguous. One small custom node that makes the data contiguous fixed it, and two hand-built kernel rewrites bought nothing.

Our ComfyUI Work · part 3
Generative media 6 min

Eight ways a ComfyUI job succeeds and is wrong

Noise images from a LoRA at the wrong strength, a cache that times itself, a silently dropped graph branch, a node pack that vanished, and a LoRA trained inside ComfyUI in 4.5 minutes. Each has a cheap guard.

Our ComfyUI Work · part 6
Measurement 6 min

The Bottleneck Was the PCIe Link, Not the Code

GPU idle was 60 percent of decode wall time, and the largest share was host-to-device copies running at 3.1 to 3.4 GB/s against 28.5 GB/s on a healthy boot. The root port negotiated 2.5 GT/s. Several timed A/Bs from that boot are void, one fusion ratio survives, and the machine is still on the slow link.

GPU kernels 8 min

Same Bits on the CPU and the GPU

Std.math exp disagreed between the CPU and the GPU on 19,437 of 40,017 inputs. A shared exp written with explicit fma matched on all 40,017, at 3.6 times the cost in an exp-bound kernel. One attention layer then matched bit for bit, and a float Freivalds certificate caught every planted fault.

Measurement 7 min

A Gate You Have Never Seen Fail Is Not a Gate Yet

The fused q4x kernels cut 1244 dispatches per decode token to 992, behind a gate that had never been red. One nudge of about 5e-7 per kernel turned it red three times, each time naming its own kernel. The identity check beside it has still never been seen to fail.

Gates that lie · part 5
Simulation & neuro 7 min

The fly engine waits on memory, not math

40.0 GB read and 15.9 GB written per launch is 59 to 87 percent of the card's bandwidth, depending on which clock mode you trust. By the review's estimate the biggest items are repeated reads of factors and vectors, not the Jacobian we blamed.

The Fly · part 8
Measurement 6 min

The profiler crashed because the CPU has a GPU

Four days of a wrong verdict ("unusable with our Mojo kernels") ended when we unbound the integrated GPU. A plain kernel then read SQ_WAVES 32,768 exactly, and a second fault, counters reading zero, came out from behind the first.

Practice 7 min

The GPU Is a Queue

A day with 142 GPU jobs and 34 failures, and a week of incidents that never needed the GPU to find. The fix is not a faster GPU, it's treating the card as a shared, budgeted resource with a preflight step before anything touches it.

GPU kernels 6 min

The Shape That Actually Matters

Square GEMM is the benchmark everyone runs; decode is M=1 streaming weights out of HBM. Four frozen rounds, three wrong predictions, and a one-percent win that came from occupancy relief rather than fewer bytes.

Kernels on a gaming GPU · part 2