<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>gdbg — GPU Profiler for AI Agents</title>
<meta name="description" content="GPU profiling for AI agents. Three collection phases, one REPL, 30+ analysis commands. Your agent stops guessing at GPU performance and starts measuring it.">
<link rel="preconnect" href="https://fonts.googleapis.com">
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
<link href="https://fonts.googleapis.com/css2?family=DM+Mono:ital,wght@0,300;0,400;0,500;1,400&family=Instrument+Serif:ital@0;1&display=swap" rel="stylesheet">
<style>
*{margin:0;padding:0;box-sizing:border-box}
:root{
--void:#06080c;
--ink:#0b0e14;
--slab:#111620;
--rule:#1a2233;
--mute:#4a5568;
--read:#94a3b8;
--lit:#d1dce8;
--white:#f1f5f9;
--trace:#3dffa0;
--probe:#5eead4;
--signal:#38bdf8;
--fault:#ff5c5c;
--warn:#fbbf24;
--gpu:#c084fc;
--glow:rgba(61,255,160,.06);
--mono:"DM Mono",ui-monospace,"Cascadia Code","Fira Code",Menlo,Consolas,monospace;
--serif:"Instrument Serif",Georgia,"Times New Roman",serif;
}
html{scroll-behavior:smooth}
body{
background:var(--void);
color:var(--lit);
font-family:var(--mono);
font-weight:400;
font-size:15px;
line-height:1.65;
overflow-x:hidden;
-webkit-font-smoothing:antialiased;
}
body::after{
content:'';
position:fixed;
inset:0;
pointer-events:none;
z-index:9999;
background:repeating-linear-gradient(
0deg,
transparent,
transparent 2px,
rgba(0,0,0,.03) 2px,
rgba(0,0,0,.03) 4px
);
}
.grid-bg{
position:fixed;
inset:0;
z-index:-1;
background-image:
linear-gradient(var(--rule) 1px,transparent 1px),
linear-gradient(90deg,var(--rule) 1px,transparent 1px);
background-size:60px 60px;
opacity:.25;
mask-image:radial-gradient(ellipse 80% 70% at 50% 30%,black 20%,transparent 70%);
}
a{color:var(--probe);text-decoration:none;transition:color .2s}
a:hover{color:var(--trace)}
.top-nav{
display:flex;
justify-content:center;
gap:4px;
padding:16px 24px 0;
}
.top-nav a{
font-size:.75rem;
font-weight:500;
letter-spacing:.06em;
padding:6px 16px;
border-radius:4px;
color:var(--mute);
transition:all .2s;
}
.top-nav a:hover{color:var(--read)}
.top-nav a.active{
color:var(--gpu);
background:rgba(192,132,252,.06);
}
.hero{
position:relative;
display:flex;
flex-direction:column;
align-items:center;
text-align:center;
padding:48px 24px 40px;
max-width:800px;
margin:0 auto;
}
.hero-glow{
position:absolute;
top:-120px;left:50%;
transform:translateX(-50%);
width:600px;height:400px;
background:radial-gradient(ellipse,rgba(192,132,252,.1) 0%,transparent 70%);
pointer-events:none;
}
.hero .logo{width:120px;filter:drop-shadow(0 0 40px rgba(192,132,252,.2));margin-bottom:24px}
.hero .badge{
display:inline-flex;align-items:center;gap:8px;
font-size:.72rem;
color:var(--mute);
letter-spacing:.08em;
margin-bottom:24px;
}
.hero .badge a{color:var(--read)}
.hero h1{
font-family:var(--serif);
font-weight:400;
font-size:clamp(2.2rem,5vw,3.4rem);
line-height:1.15;
color:var(--white);
margin-bottom:20px;
letter-spacing:-.02em;
}
.hero h1 em{
font-style:italic;
color:var(--gpu);
text-shadow:0 0 30px rgba(192,132,252,.3);
}
.hero .sub{
font-size:.95rem;
color:var(--read);
max-width:500px;
margin-bottom:36px;
line-height:1.7;
}
.hero .actions{display:flex;gap:12px;flex-wrap:wrap;justify-content:center}
.btn{
display:inline-flex;align-items:center;gap:6px;
padding:10px 22px;
border-radius:4px;
font-family:var(--mono);font-size:.85rem;font-weight:500;
text-decoration:none !important;
transition:all .2s;
border:1px solid transparent;
}
.btn-g{
background:var(--gpu);color:var(--void) !important;
box-shadow:0 0 20px rgba(192,132,252,.15);
}
.btn-g:hover{box-shadow:0 0 30px rgba(192,132,252,.3);transform:translateY(-1px)}
.btn-o{
background:transparent;color:var(--probe) !important;
border-color:var(--rule);
}
.btn-o:hover{border-color:var(--probe);background:rgba(94,234,212,.04)}
.wrap{max-width:820px;margin:0 auto;padding:0 24px}
section{padding:36px 0}
.sec-label{
font-size:.7rem;
font-weight:500;
letter-spacing:.15em;
text-transform:uppercase;
color:var(--mute);
margin-bottom:10px;
}
.sec-title{
font-family:var(--serif);
font-weight:400;
font-size:clamp(1.6rem,3vw,2.2rem);
color:var(--white);
margin-bottom:12px;
letter-spacing:-.01em;
}
.sec-desc{
color:var(--read);
font-size:.9rem;
margin-bottom:40px;
}
.versus{
display:grid;
grid-template-columns:1fr 1fr;
gap:2px;
background:var(--rule);
border-radius:6px;
overflow:hidden;
}
@media(max-width:640px){.versus{grid-template-columns:1fr}}
.versus-card{
background:var(--ink);
padding:32px 28px;
}
.versus-card .tag{
display:inline-block;
font-size:.65rem;
font-weight:500;
letter-spacing:.12em;
text-transform:uppercase;
padding:3px 8px;
border-radius:3px;
margin-bottom:14px;
}
.versus-card.before .tag{background:rgba(255,92,92,.12);color:var(--fault)}
.versus-card.after .tag{background:rgba(192,132,252,.12);color:var(--gpu)}
.versus-card p{font-size:.88rem;color:var(--read);line-height:1.7}
.versus-card p strong{color:var(--lit);font-weight:500}
.demo-grid{
display:grid;
grid-template-columns:repeat(2,1fr);
gap:16px;
margin-bottom:28px;
}
.demo-panel{
background:var(--ink);
border:1px solid var(--rule);
border-radius:6px;
padding:14px 16px;
overflow:hidden;
position:relative;
}
.demo-panel::before{
content:'';position:absolute;inset:0;
background:radial-gradient(ellipse at 50% 0%,rgba(192,132,252,.04),transparent 70%);
pointer-events:none;
}
.demo-cmd{
font-family:var(--mono);
font-size:.78rem;
color:var(--read);
padding-bottom:8px;
margin-bottom:8px;
border-bottom:1px solid var(--rule);
letter-spacing:.02em;
}
.demo-out{
font-family:var(--mono);
font-size:.72rem;
line-height:1.55;
color:#c9d1d9;
white-space:pre;
overflow-x:auto;
margin:0;
}
.demo-out::-webkit-scrollbar{height:4px}
.demo-out::-webkit-scrollbar-track{background:transparent}
.demo-out::-webkit-scrollbar-thumb{background:var(--rule);border-radius:2px}
.c-fg-green{color:#3dffa0}
.c-fg-yellow{color:#fbbf24}
.c-fg-red{color:#ff6b6b}
.c-fg-magenta{color:#c084fc}
.phase-grid{
display:grid;
grid-template-columns:repeat(3,1fr);
gap:2px;
background:var(--rule);
border-radius:6px;
overflow:hidden;
}
@media(max-width:640px){.phase-grid{grid-template-columns:1fr}}
.phase-card{
background:var(--ink);
padding:28px 24px;
}
.phase-card .num{
font-family:var(--serif);
font-size:1.3rem;
color:var(--gpu);
opacity:.7;
margin-bottom:10px;
}
.phase-card h3{
font-size:.85rem;
font-weight:500;
color:var(--white);
margin-bottom:8px;
}
.phase-card .tool-name{
display:inline-block;
font-size:.7rem;
color:var(--gpu);
background:rgba(192,132,252,.08);
padding:2px 8px;
border-radius:3px;
margin-bottom:10px;
}
.phase-card p{font-size:.82rem;color:var(--read);line-height:1.6}
.backend-grid{
display:grid;
grid-template-columns:repeat(3,1fr);
gap:2px;
background:var(--rule);
border-radius:6px;
overflow:hidden;
}
@media(max-width:640px){.backend-grid{grid-template-columns:1fr}}
.backend-card{
background:var(--ink);
padding:28px 24px;
}
.backend-card h3{
font-size:.95rem;
font-weight:500;
color:var(--white);
margin-bottom:6px;
}
.backend-card .tools{
font-size:.72rem;
color:var(--gpu);
letter-spacing:.03em;
margin-bottom:16px;
padding-bottom:12px;
border-bottom:1px solid var(--rule);
}
.backend-card p{font-size:.82rem;color:var(--read);line-height:1.6}
.backend-card ul{
list-style:none;
padding:0;margin:0;
display:grid;
gap:6px;
}
.backend-card ul li{
font-size:.82rem;
color:var(--read);
line-height:1.5;
padding-left:14px;
position:relative;
}
.backend-card ul li::before{
content:'\203A';
position:absolute;
left:0;
color:var(--gpu);
opacity:.6;
}
.cmd-group{margin-bottom:32px}
.cmd-group:last-child{margin-bottom:0}
.cmd-group-title{
font-size:.7rem;
font-weight:500;
letter-spacing:.12em;
text-transform:uppercase;
color:var(--gpu);
margin-bottom:12px;
padding-bottom:6px;
border-bottom:1px solid var(--rule);
}
.cmd-list{display:grid;gap:0}
.cmd-row{
display:grid;
grid-template-columns:200px 1fr;
gap:16px;
padding:8px 0;
border-bottom:1px solid rgba(26,34,51,.5);
align-items:baseline;
}
.cmd-row:last-child{border-bottom:none}
.cmd-name{
font-size:.82rem;
color:var(--trace);
white-space:nowrap;
}
.cmd-desc{
font-size:.82rem;
color:var(--read);
}
.workflow-list{display:grid;gap:0;counter-reset:wf}
.wf-item{
display:grid;
grid-template-columns:48px 1fr;
gap:0;
padding:20px 0;
border-bottom:1px solid var(--rule);
}
.wf-item:last-child{border-bottom:none}
.wf-item::before{
counter-increment:wf;
content:counter(wf,decimal-leading-zero);
font-family:var(--serif);
font-size:1.1rem;
color:var(--gpu);
opacity:.6;
padding-top:1px;
}
.wf-item strong{color:var(--white);font-weight:500;display:block;margin-bottom:2px}
.wf-item span{color:var(--read);font-size:.85rem}
.wf-item code{
background:var(--slab);
padding:1px 6px;
border-radius:3px;
font-size:.8em;
color:var(--probe);
}
.install-steps{display:grid;gap:2px;background:var(--rule);border-radius:6px;overflow:hidden}
.install-step{
background:var(--ink);
padding:24px 28px;
display:grid;
grid-template-columns:32px 1fr;
gap:16px;
align-items:start;
}
.install-step .num{
font-family:var(--serif);
font-size:1.4rem;
color:var(--gpu);
line-height:1;
padding-top:2px;
}
.install-step pre{
background:var(--slab);
border:1px solid var(--rule);
border-radius:4px;
padding:10px 14px;
font-family:var(--mono);
font-size:.82rem;
color:var(--gpu);
overflow-x:auto;
margin-top:8px;
}
.install-step p{font-size:.85rem;color:var(--read)}
.install-step p strong{color:var(--lit);font-weight:500}
.arch-list{display:grid;gap:0;counter-reset:arch}
.arch-item{
display:grid;
grid-template-columns:48px 1fr;
gap:0;
padding:20px 0;
border-bottom:1px solid var(--rule);
}
.arch-item:last-child{border-bottom:none}
.arch-item::before{
counter-increment:arch;
content:counter(arch,decimal-leading-zero);
font-family:var(--serif);
font-size:1.1rem;
color:var(--gpu);
opacity:.6;
padding-top:1px;
}
.arch-item strong{color:var(--white);font-weight:500;display:block;margin-bottom:2px}
.arch-item span{color:var(--read);font-size:.85rem}
footer{
padding:60px 24px;
text-align:center;
border-top:1px solid var(--rule);
}
footer p{font-size:.78rem;color:var(--mute)}
footer a{color:var(--read)}
footer a:hover{color:var(--gpu)}
footer .sig{margin-top:12px;font-size:.7rem;color:var(--rule);letter-spacing:.05em}
.reveal{opacity:0;transform:translateY(16px);transition:opacity .6s ease,transform .6s ease}
.reveal.visible{opacity:1;transform:translateY(0)}
@media(max-width:640px){
body{font-size:14px}
.hero{padding:48px 20px 32px}
.hero .logo{width:80px;margin-bottom:16px}
.hero h1{font-size:1.7rem;margin-bottom:14px}
.hero .sub{font-size:.82rem;margin-bottom:24px;line-height:1.6}
.btn{padding:9px 16px;font-size:.78rem}
section{padding:20px 0}
.wrap{padding:0 16px}
.sec-title{font-size:1.3rem;margin-bottom:8px}
.sec-desc{font-size:.8rem;margin-bottom:28px}
.versus{grid-template-columns:1fr}
.versus-card{padding:20px 18px}
.versus-card p{font-size:.8rem;line-height:1.6}
.demo-grid{grid-template-columns:1fr;gap:12px}
.demo-panel{padding:12px 14px}
.demo-cmd{font-size:.72rem}
.demo-out{font-size:.66rem;line-height:1.5}
.phase-grid{grid-template-columns:1fr}
.phase-card{padding:20px 18px}
.backend-grid{grid-template-columns:1fr}
.backend-card{padding:20px 18px}
.cmd-row{grid-template-columns:1fr;gap:2px}
.cmd-name{font-size:.78rem}
.cmd-desc{font-size:.78rem}
.install-step{grid-template-columns:1fr;padding:18px 16px}
.install-step .num{display:none}
.install-step pre{font-size:.73rem;padding:8px 10px}
.wf-item{grid-template-columns:36px 1fr;padding:14px 0}
.arch-item{grid-template-columns:36px 1fr;padding:14px 0}
footer{padding:32px 16px}
}
@media(max-width:380px){
.hero{padding:40px 16px 28px}
.hero .logo{width:70px;margin-bottom:12px}
.hero h1{font-size:1.5rem}
.btn{padding:8px 14px;font-size:.75rem}
.sec-title{font-size:1.2rem}
}
</style>
</head>
<body>
<div class="grid-bg" aria-hidden="true"></div>
<nav class="top-nav">
<a href="index.html">dbg</a>
<a href="gdbg.html" class="active">gdbg</a>
</nav>
<header class="hero">
<div class="hero-glow" aria-hidden="true"></div>
<img class="logo" src="logo_t.png" alt="gdbg" width="120" height="120">
<h1>Multiple profilers. <em>One answer.</em></h1>
<p class="sub">GPU performance is opaque. Your AI agent stops guessing at bottlenecks and starts measuring them.</p>
<div class="actions">
<a class="btn btn-g" href="https://github.com/redknightlois/dbg">View source</a>
<a class="btn btn-o" href="#commands">See commands ↓</a>
</div>
</header>
<section>
<div class="wrap">
<div class="sec-title">Agents guess. Profilers measure.</div>
<p class="sec-desc">GPU kernels are opaque. Profiling tools are fragmented. Agents need structured data, not raw traces.</p>
<div class="versus reveal">
<div class="versus-card before">
<div class="tag">Without gdbg</div>
<p>The agent reads CUDA code, <strong>guesses at bottlenecks</strong>, suggests generic optimizations. Is the kernel compute-bound or memory-bound? <strong>No idea.</strong> Are there fusion opportunities? <strong>Can't tell without data.</strong></p>
</div>
<div class="versus-card after">
<div class="tag">With gdbg</div>
<p>The agent runs <code style="background:var(--slab);padding:1px 6px;border-radius:3px;font-size:.85em;color:var(--gpu)">gdbg train.py</code>, sees the roofline classification, finds <strong>the kernel eating 40% of GPU time is memory-bound</strong>, and targets the fix. <strong>Data, not hunches.</strong></p>
</div>
</div>
</div>
</section>
<section>
<div class="wrap">
<div class="sec-title">See what the agent sees.</div>
<p class="sec-desc">One command. Every angle on GPU performance.</p>
<div class="demo-grid reveal">
<div class="demo-panel">
<div class="demo-cmd"><span class="c-fg-magenta">gpu></span> stream-graph</div>
<pre class="demo-out"> Stream Graph (0.0ms → 20.9ms, span 20.9ms)
s13 │B AAA B│
s14 │ AA │
s15 │ AA │
└────────────────────────────────────────────────────────┘
Legend:
<span class="c-fg-green">A</span> burst_kernel(float *, int, int) 490.2us
<span class="c-fg-green">B</span> quiet_kernel(float *, int) 25.2us</pre>
</div>
<div class="demo-panel">
<div class="demo-cmd"><span class="c-fg-magenta">gpu></span> hotspot 5000</div>
<pre class="demo-out"> Hottest 5.0ms window:
Window: 574.3ms → 579.3ms
Busy time: <span class="c-fg-green">490.2us</span> (9.8% of window)
Launches: 12
Kernel Launches Time % busy
───────────────────────────── ──────── ──────── ──────
burst_kernel 12 490.2us <span class="c-fg-green">100.0%</span></pre>
</div>
<div class="demo-panel">
<div class="demo-cmd"><span class="c-fg-magenta">gpu></span> bandwidth</div>
<pre class="demo-out"> Per-kernel Memory Bandwidth:
# Kernel Achieved % peak Bound
── ──────────────── ────────── ────── ────────
1 coalesced_copy <span class="c-fg-green">612.4 GB/s</span> 82.3% memory
2 strided_copy <span class="c-fg-yellow">78.2 GB/s</span> 10.5% memory <span class="c-fg-red">←low</span>
1 kernel under 50% of peak — likely memory-access bound
(poor coalescing, low L2, uncoalesced strided loads)</pre>
</div>
<div class="demo-panel">
<div class="demo-cmd"><span class="c-fg-magenta">gpu></span> critical-path</div>
<pre class="demo-out"> Critical path chains (same stream, gap ≤ 100.0us):
Longest chain: stream 7 span 12.4ms active 12.2ms (<span class="c-fg-green">98%</span>)
24 kernel(s)
Top kernels on chain:
Kernel Launches Time % chain
───────────────────── ──────── ──────── ───────
trunk_step 24 12.2ms <span class="c-fg-green">100.0%</span></pre>
</div>
</div>
</div>
</section>
<section>
<div class="wrap">
<div class="sec-title">Collect everything. At once.</div>
<p class="sec-desc">One command triggers three independent collection phases. Each can fail without blocking the others. You always get the data that's available.</p>
<div class="phase-grid reveal">
<div class="phase-card">
<div class="num">1</div>
<div class="tool-name">Timeline</div>
<h3>GPU Timeline</h3>
<p>Kernel launches, memory transfers, stream activity, NVTX regions. The big picture of what the GPU actually did.</p>
</div>
<div class="phase-card">
<div class="num">2</div>
<div class="tool-name">Metrics</div>
<h3>Hardware Metrics</h3>
<p>Occupancy, throughput, registers, shared memory, L2 hit rate. Collected on the hottest kernels, where the detail matters.</p>
</div>
<div class="phase-card">
<div class="num">3</div>
<div class="tool-name">Mapping</div>
<h3>Op Mapping</h3>
<p>Which operator launched which kernel. Bridges the gap between your code and the hardware.</p>
</div>
</div>
</div>
</section>
<section>
<div class="wrap">
<div class="sec-title">Auto-detects your target.</div>
<p class="sec-desc">Point gdbg at a file. It reads the imports and picks the right collection strategy.</p>
<div class="backend-grid reveal">
<div class="backend-card">
<h3>CUDA</h3>
<ul>
<li>.cu source files</li>
<li>Compiled CUDA binaries</li>
<li>PyCUDA / CuPy scripts</li>
</ul>
</div>
<div class="backend-card">
<h3>PyTorch</h3>
<ul>
<li>Training scripts</li>
<li>Inference pipelines</li>
<li>Custom autograd ops</li>
</ul>
</div>
<div class="backend-card">
<h3>Triton</h3>
<ul>
<li>Triton kernels</li>
<li>Flash Attention variants</li>
<li>Custom fused ops</li>
</ul>
</div>
</div>
</div>
</section>
<section id="commands">
<div class="wrap">
<div class="sec-title">30+ commands.</div>
<p class="sec-desc">After collection, the agent queries everything from a single interface.</p>
<div class="cmd-group reveal">
<div class="cmd-group-title">Hotspots</div>
<div class="cmd-list">
<div class="cmd-row"><div class="cmd-name">kernels [N] [pattern]</div><div class="cmd-desc">Top kernels by total GPU time</div></div>
<div class="cmd-row"><div class="cmd-name">ops [N] [pattern]</div><div class="cmd-desc">Top operators by GPU time (needs op mapping data)</div></div>
<div class="cmd-row"><div class="cmd-name">stats</div><div class="cmd-desc">Overall session summary</div></div>
<div class="cmd-row"><div class="cmd-name">top-ops [N] [pattern]</div><div class="cmd-desc">Operators ranked by GPU time contribution</div></div>
</div>
</div>
<div class="cmd-group reveal">
<div class="cmd-group-title">Analysis</div>
<div class="cmd-list">
<div class="cmd-row"><div class="cmd-name">roofline [pattern]</div><div class="cmd-desc">Classify compute-bound vs memory-bound</div></div>
<div class="cmd-row"><div class="cmd-name">bound <kernel></div><div class="cmd-desc">Detailed boundedness diagnosis for a kernel</div></div>
<div class="cmd-row"><div class="cmd-name">occupancy [N]</div><div class="cmd-desc">SM occupancy ranking</div></div>
<div class="cmd-row"><div class="cmd-name">variance <kernel></div><div class="cmd-desc">Launch-to-launch timing variance</div></div>
<div class="cmd-row"><div class="cmd-name">warmup</div><div class="cmd-desc">Detect warmup launches before steady state</div></div>
<div class="cmd-row"><div class="cmd-name">small [N]</div><div class="cmd-desc">Kernels where launch overhead exceeds compute</div></div>
<div class="cmd-row"><div class="cmd-name">fuse [N]</div><div class="cmd-desc">Sequential kernels that could be fused</div></div>
<div class="cmd-row"><div class="cmd-name">concurrency</div><div class="cmd-desc">Stream utilization and parallelism gaps</div></div>
<div class="cmd-row"><div class="cmd-name">hotpath</div><div class="cmd-desc">Critical path through ops (CPU vs GPU bound)</div></div>
<div class="cmd-row"><div class="cmd-name">compare-ops [N]</div><div class="cmd-desc">CPU vs GPU time ratio per operator</div></div>
<div class="cmd-row"><div class="cmd-name">breakdown <op></div><div class="cmd-desc">Which kernels an operator expands into</div></div>
<div class="cmd-row"><div class="cmd-name">idle-between <a> <b></div><div class="cmd-desc">GPU idle gap between two operators</div></div>
</div>
</div>
<div class="cmd-group reveal">
<div class="cmd-group-title">Timeline</div>
<div class="cmd-list">
<div class="cmd-row"><div class="cmd-name">transfers [N]</div><div class="cmd-desc">Memory copies ranked by cost</div></div>
<div class="cmd-row"><div class="cmd-name">gaps [N]</div><div class="cmd-desc">GPU idle periods</div></div>
<div class="cmd-row"><div class="cmd-name">overlap</div><div class="cmd-desc">Compute/transfer concurrency</div></div>
<div class="cmd-row"><div class="cmd-name">streams</div><div class="cmd-desc">Per-stream utilization breakdown</div></div>
<div class="cmd-row"><div class="cmd-name">timeline [N]</div><div class="cmd-desc">Chronological kernel launches</div></div>
</div>
</div>
<div class="cmd-group reveal">
<div class="cmd-group-title">Drill-down</div>
<div class="cmd-list">
<div class="cmd-row"><div class="cmd-name">inspect <kernel></div><div class="cmd-desc">Full detail from all data layers</div></div>
<div class="cmd-row"><div class="cmd-name">trace <op></div><div class="cmd-desc">Operator to kernel mapping</div></div>
<div class="cmd-row"><div class="cmd-name">callers <kernel></div><div class="cmd-desc">Which operator launched this kernel</div></div>
</div>
</div>
<div class="cmd-group reveal">
<div class="cmd-group-title">Filtering</div>
<div class="cmd-list">
<div class="cmd-row"><div class="cmd-name">focus <pattern></div><div class="cmd-desc">Show only matching kernels</div></div>
<div class="cmd-row"><div class="cmd-name">ignore <pattern></div><div class="cmd-desc">Hide matching kernels</div></div>
<div class="cmd-row"><div class="cmd-name">region <name></div><div class="cmd-desc">Focus on NVTX / profiler step</div></div>
<div class="cmd-row"><div class="cmd-name">reset</div><div class="cmd-desc">Clear all filters</div></div>
</div>
</div>
<div class="cmd-group reveal">
<div class="cmd-group-title">Sessions</div>
<div class="cmd-list">
<div class="cmd-row"><div class="cmd-name">save <name></div><div class="cmd-desc">Save session to .dbg/gpu/</div></div>
<div class="cmd-row"><div class="cmd-name">list</div><div class="cmd-desc">List saved sessions</div></div>
<div class="cmd-row"><div class="cmd-name">diff <name></div><div class="cmd-desc">Compare current session against a saved one</div></div>
<div class="cmd-row"><div class="cmd-name">layers</div><div class="cmd-desc">Show loaded data layers</div></div>
<div class="cmd-row"><div class="cmd-name">suggest</div><div class="cmd-desc">Suggest what data to collect next</div></div>
</div>
</div>
</div>
</section>
<section>
<div class="wrap">
<div class="sec-title">How the agent thinks.</div>
<p class="sec-desc">The agent doesn't follow a script. It reasons about GPU performance the way an engineer would.</p>
<div class="workflow-list reveal">
<div class="wf-item"><div><strong>“What's actually slow?”</strong><span>The agent runs <code>gdbg train.py</code>, then <code>stats</code> and <code>kernels</code>. Now it knows where GPU time goes — not where it assumed it went.</span></div></div>
<div class="wf-item"><div><strong>“Why is it slow?”</strong><span><code>roofline</code> answers the question that matters: is this kernel starved for compute or starved for memory? The fix depends on the answer.</span></div></div>
<div class="wf-item"><div><strong>“What else is wrong?”</strong><span><code>fuse</code> finds sequential kernels that should be one. <code>small</code> finds kernels where launch overhead dominates. <code>gaps</code> finds idle time the GPU wasted.</span></div></div>
<div class="wf-item"><div><strong>“Show me everything about this kernel.”</strong><span><code>inspect</code> pulls hardware counters, occupancy, timing, and the operator that launched it — all in one view.</span></div></div>
<div class="wf-item"><div><strong>“Did the fix actually help?”</strong><span>The agent saves a baseline, makes changes, re-profiles, and runs <code>diff</code>. No guessing. Numbers go up or they don't.</span></div></div>
</div>
</div>
</section>
<section>
<div class="wrap">
<div class="sec-title">Get started.</div>
<p class="sec-desc">gdbg ships with dbg. One install gives you both.</p>
<div class="install-steps reveal">
<div class="install-step">
<div class="num">1</div>
<div>
<p><strong>Install</strong></p>
<pre>cargo install dbg-cli</pre>
</div>
</div>
<div class="install-step">
<div class="num">2</div>
<div>
<p><strong>Check dependencies</strong></p>
<pre>gdbg check</pre>
<p style="margin-top:10px">Verifies all required tools are available. Tells you exactly what's missing and how to install it.</p>
</div>
</div>
<div class="install-step">
<div class="num">3</div>
<div>
<p><strong>Profile</strong></p>
<pre>gdbg train.py</pre>
<p style="margin-top:10px">Auto-detects the target type, collects data, and drops you into the REPL.</p>
</div>
</div>
</div>
</div>
</section>
<section>
<div class="wrap">
<div class="sec-title">How it works.</div>
<div class="arch-list reveal">
<div class="arch-item"><div><strong>Three independent phases.</strong><span>Timeline, hardware metrics, and op mapping collect separately. A failure in one doesn't block the others. You always get the data that's available.</span></div></div>
<div class="arch-item"><div><strong>One session, every layer.</strong><span>Timeline, hardware counters, and op mapping merge into a single queryable session. No jumping between tools or parsing different output formats.</span></div></div>
<div class="arch-item"><div><strong>Top-5 targeting.</strong><span>Hardware metrics collection is expensive. gdbg runs it only on the hottest kernels from the timeline, not everything. Fast enough to use interactively.</span></div></div>
<div class="arch-item"><div><strong>Cross-layer correlation.</strong><span>Op mapping connects Python operators to GPU kernels. <code style="background:var(--slab);padding:1px 6px;border-radius:3px;font-size:.8em;color:var(--probe)">trace matmul</code> shows every kernel that a matmul op launched.</span></div></div>
<div class="arch-item"><div><strong>Session diffing.</strong><span>Save a baseline, optimize, re-profile, diff. The agent sees exactly what changed — faster kernels, fewer launches, better occupancy.</span></div></div>
</div>
</div>
</section>
<footer>
<p><a href="index.html">dbg</a>  ·  <a href="gdbg.html">gdbg</a>  ·  <a href="https://github.com/redknightlois/dbg">GitHub</a>  ·  MIT License  ·  Built by <a href="https://github.com/redknightlois">Federico Lois</a></p>
<div class="sig">gdbg — gpu profiler for ai agents</div>
</footer>
<script>
var rev=new IntersectionObserver(function(entries){
entries.forEach(function(e){if(e.isIntersecting){e.target.classList.add('visible');rev.unobserve(e.target)}})
},{threshold:.15,rootMargin:'0px 0px -40px 0px'});
document.querySelectorAll('.reveal').forEach(function(el){rev.observe(el)});
</script>
</body>
</html>
