Updated benchmarks

This commit is contained in:
Donato Capitella
2025-11-15 08:36:25 +00:00
parent 1d945f2c21
commit 67fb3a002b
664 changed files with 32551 additions and 15807 deletions
+45 -752
View File
@@ -2,780 +2,73 @@
<html lang="en">
<head>
<meta charset="utf-8" />
<meta name="viewport" content="width=device-width,initial-scale=1" />
<title>AMD Ryzen AI MAX+ 395 "Strix Halo"Llama.cpp Backend Performance Comparison</title>
<style>
:root {
--bg: #ffffff;
--ink: #000000;
--muted: #555555;
--accent: #0645ad;
--card: #f9f9f9;
--border: #cccccc;
--sticky: #f4f4f4;
--pill: #e0e0e0;
--roc: #d3caff;
--roc-text: #000000;
--warn: #d88f00;
--bad: #d32f2f;
--model-w: 300px;
--winner-w: 220px;
}
* {
box-sizing: border-box;
}
body {
margin: 0;
background: var(--bg);
color: var(--ink);
font: 15px/1.5 system-ui, -apple-system, Segoe UI, Roboto, Inter;
}
a {
color: var(--accent);
text-decoration: none;
}
a:hover {
text-decoration: underline;
}
header {
padding: 14px 20px 0;
}
header h1 {
margin: 0 0 6px;
font-size: 22px;
}
header p {
margin: 4px 0 12px;
color: var(--muted);
font-size: 13px;
}
.controls {
display: grid;
grid-template-columns: 1fr 160px 260px 180px;
gap: 10px;
align-items: end;
padding: 10px 20px;
}
label {
display: block;
font-size: 12px;
color: var(--muted);
margin: 0 0 6px;
}
input[type="text"],
select,
input[type="number"] {
width: 100%;
background: #fff;
color: var(--ink);
border: 1px solid var(--border);
border-radius: 10px;
padding: 8px 10px;
outline: none;
}
.slider {
display: grid;
grid-template-columns: 1fr 90px;
gap: 8px;
align-items: center;
}
.range-wrap {
position: relative;
height: 30px;
}
.range-wrap input[type="range"] {
position: absolute;
inset: 0;
width: 100%;
margin: 0;
background: transparent;
pointer-events: none;
appearance: none;
}
.range-wrap input[type="range"]::-webkit-slider-thumb {
pointer-events: auto;
appearance: none;
height: 16px;
width: 16px;
border-radius: 50%;
background: var(--accent);
border: 2px solid #fff;
box-shadow: 0 0 2px rgba(0, 0, 0, 0.3);
cursor: pointer;
transition: background 0.2s, transform 0.2s;
}
.range-wrap input[type="range"]::-webkit-slider-thumb:hover {
background: #0451cc;
transform: scale(1.1);
}
.range-wrap input[type="range"]::-moz-range-thumb {
pointer-events: auto;
height: 16px;
width: 16px;
border-radius: 50%;
background: var(--accent);
border: 2px solid #fff;
box-shadow: 0 0 2px rgba(0, 0, 0, 0.3);
cursor: pointer;
transition: background 0.2s, transform 0.2s;
}
.range-wrap input[type="range"]::-moz-range-thumb:hover {
background: #0451cc;
transform: scale(1.1);
}
.range-track {
position: absolute;
top: 50%;
height: 6px;
width: 100%;
transform: translateY(-50%);
border-radius: 999px;
background: #e5e5e5;
}
.range-values {
margin-top: 4px;
font-size: 12px;
color: var(--muted);
text-align: center;
}
input[type="range"] {
width: 100%;
}
.ticks {
display: flex;
justify-content: space-between;
color: var(--muted);
font-size: 11px;
margin-top: 2px;
}
.cols {
display: flex;
gap: 8px;
flex-wrap: wrap;
padding: 0 20px 8px;
}
.colbox {
border: 1px solid var(--border);
border-radius: 12px;
padding: 8px 10px;
min-width: 220px;
background: var(--card);
}
.colbox label {
display: flex;
gap: 8px;
align-items: center;
margin: 0;
}
.colbox input {
transform: translateY(1px);
}
.badge {
display: inline-block;
padding: 2px 8px;
border-radius: 999px;
background: var(--pill);
font-size: 11px;
margin-left: 6px;
}
.roc {
background: var(--roc);
color: var(--roc-text);
border: 1px solid #b19cff55;
}
.faall {
background: #cfe9ff;
/* light blue chip */
color: #000000;
border: 1px solid #9bc9ff55;
}
.meta {
padding: 0 20px 14px;
color: var(--muted);
font-size: 12px;
}
.section {
padding: 6px 20px 0;
}
.section h2 {
margin: 6px 0 8px;
font-size: 16px;
color: var(--ink);
font-weight: 600;
}
.tablewrap {
margin: 0 0 16px;
border-top: 1px solid var(--border);
}
.scroller {
overflow-x: auto;
overflow-y: hidden;
}
.scroller::-webkit-scrollbar {
height: 10px;
}
.scroller::-webkit-scrollbar-thumb {
background: var(--border);
border-radius: 8px;
}
.scroller-top {
overflow-x: auto;
overflow-y: hidden;
height: 12px;
/* slim */
margin: 0 0 6px;
/* a little gap above the table */
}
.scroller-top .scroller-spacer {
height: 1px;
/* tiny content so the bar renders */
}
table {
width: max-content;
min-width: 100vw;
border-collapse: separate;
border-spacing: 0;
}
thead th {
position: sticky;
top: 0;
background: var(--card);
z-index: 2;
}
th,
td {
padding: 10px 12px;
border-bottom: 1px solid var(--border);
vertical-align: top;
}
th {
white-space: nowrap;
}
td {
white-space: normal;
}
.col-model {
position: sticky;
left: 0;
background: var(--sticky);
z-index: 3;
max-width: var(--model-w);
width: var(--model-w);
}
.col-winner {
position: sticky;
left: calc(var(--model-w));
background: var(--sticky);
z-index: 3;
max-width: var(--winner-w);
width: var(--winner-w);
}
.model-title {
font-weight: 600;
word-wrap: break-word;
overflow-wrap: anywhere;
}
.model-meta {
color: var(--muted);
font-size: 12px;
margin-top: 2px;
}
.cell {
min-width: 120px;
text-align: center;
}
.tps {
display: block;
text-align: center;
}
.std {
color: var(--muted);
font-size: 12px;
text-align: center;
}
.warn {
color: var(--warn);
}
.err {
color: var(--bad);
}
.winner-pill {
display: inline-block;
background: var(--pill);
border-radius: 999px;
padding: 2px 8px;
margin: 4px 4px 4px 0;
font-size: 12px;
}
.winner-wrap {
display: flex;
flex-wrap: wrap;
}
.env-sub {
color: var(--muted);
font-size: 12px;
}
</style>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>AMD Strix Halo — Backend Benchmarks (Grid View)</title>
<link rel="stylesheet" href="assets/index2.css">
</head>
<body>
<header>
<h1>AMD Ryzen AI MAX+ 395 "Strix Halo"Llama.cpp Backend Performance Comparison</h1>
<p class="muted">
Compare model throughput across backends (pp512 & tg128).
Repo: <a href="https://github.com/kyuz0/amd-strix-halo-toolboxes" target="_blank"
rel="noreferrer">kyuz0/amd-strix-halo-toolboxes</a>
</p>
<p class="muted">Platform: Framework Desktop, 128GB Unified RAM (accelerator-performance tuned profile)</p>
<p id="meta-line">Loading meta…</p>
<h1>AMD Ryzen AI MAX+ 395 Strix HaloBenchmark Grid</h1>
<p>Framework Desktop · AMD Ryzen AI MAX 395+ · 128GB unified RAM</p>
<p>Fedora 42 · Linux 6.18.0-0.rc5.243.vanilla.fc42.x86_64 · llama.cpp build 1c398dc9e (7034)</p>
<p>Benchmarks captured 14 Nov 2025 · Repo: <a href="https://github.com/kyuz0/amd-strix-halo-toolboxes"
target="_blank" rel="noreferrer">kyuz0/amd-strix-halo-toolboxes</a></p>
</header>
<div class="controls">
<div>
<label for="search">Search model</label>
<input id="search" type="text" placeholder="e.g. llama, qwen, 30B, Q8_K…">
<section class="controls">
<div class="control">
<label for="filter-search">Search models</label>
<input id="filter-search" type="text" placeholder="e.g. llama, qwen, 30B…">
</div>
<div>
<label for="quant">Quant</label>
<select id="quant">
<div class="control">
<label for="filter-quant">Quant</label>
<select id="filter-quant">
<option value="">Any</option>
</select>
</div>
<div>
<label>Model params</label>
<div class="range-wrap" id="sizeRange">
<div class="control grow slider-block">
<label>Context windows</label>
<div id="context-chips" class="chip-row tight"></div>
</div>
<div class="control grow slider-block">
<label>Model params (B)</label>
<div class="range-wrap">
<input type="range" id="sizeLo" step="1">
<input type="range" id="sizeHi" step="1">
<div class="range-track" id="sizeTrack"></div>
</div>
<div class="ticks">
<span>4B</span><span>50B</span><span>96B</span><span>143B</span><span>189B</span><span>235B</span>
</div>
<div class="range-values">
<span id="sizeLoVal">4B</span> <span id="sizeHiVal">235B</span>
<span id="sizeLoVal">0B</span> <span id="sizeHiVal">0B</span>
</div>
</div>
<div>
<label for="faMode">Flash Attention</label>
<select id="faMode">
<option value="off">FA off</option>
<option value="on" selected>FA on</option> <!-- default ON -->
<option value="both">Both</option>
</select>
</div>
</div>
<div class="cols" id="columns"></div>
<div class="meta">Winner = every selected backend within the bests uncertainty range, combining ± errors from both
results.</div>
<div class="section">
<h2>Prompt Processing (pp512) — tokens/second</h2>
</div>
<section class="tablewrap">
<div class="scroller scroller-top" data-for="pp">
<div class="scroller-spacer"></div>
</div>
<div class="scroller" id="scroller-pp">
<table id="tbl-pp">
<thead>
<tr></tr>
</thead>
<tbody></tbody>
</table>
</div>
</section>
<div class="section">
<h2>Text Generation (tg128) — tokens/second</h2>
</div>
<section class="tablewrap">
<div class="scroller scroller-top" data-for="tg">
<div class="scroller-spacer"></div>
<section class="panel compact">
<div class="panel-split">
<div class="backend-header">
<div class="backend-label">
<label>Backends</label>
<div class="backend-actions">
<button type="button" id="backend-all" class="chip small">All</button>
<button type="button" id="backend-none" class="chip small">None</button>
</div>
</div>
<div id="backend-list" class="backend-list"></div>
</div>
<div class="stats-box">
<div class="stat-line" id="stats-line">Loading…</div>
<button id="reset-layout" type="button" class="chip small">Reset filters</button>
</div>
</div>
<div class="scroller" id="scroller-tg">
<table id="tbl-tg">
<thead>
<tr></tr>
</thead>
<tbody></tbody>
</table>
</div>
</section>
<script>
(async function () {
const K_SIGMA = 1.0;
const MIN_TOL = 0.25;
const res = await fetch('results.json');
const data = await res.json();
const allRuns = data.runs || [];
const perfRuns = allRuns.filter(r => r.test === 'pp512' || r.test === 'tg128');
const metaLine = document.getElementById('meta-line');
const search = document.getElementById('search');
const quantSel = document.getElementById('quant');
const colPicker = document.getElementById('columns');
const faMode = document.getElementById('faMode');
// NEW dual slider elements
const sizeLo = document.getElementById('sizeLo');
const sizeHi = document.getElementById('sizeHi');
const sizeTrack = document.getElementById('sizeTrack');
const sizeLoVal = document.getElementById('sizeLoVal');
const sizeHiVal = document.getElementById('sizeHiVal');
metaLine.textContent = `${data.meta?.os_kernel || ''} — llama.cpp ${(data.meta?.llamacpp_builds || []).map(b => b.hash).join(', ') || 'unknown'} — generated ${data.meta?.generated_at || ''}`;
const models = [...new Set(perfRuns.map(r => r.model_clean || r.model))].sort((a, b) => a.localeCompare(b));
const byModel = {};
for (const name of models) { byModel[name] = { pp512: {}, tg128: {}, quant: null, sizeB: null }; }
for (const r of perfRuns) {
const name = r.model_clean || r.model;
const cfg = `${r.env}|fa=${r.fa ? 1 : 0}`;
byModel[name][r.test][cfg] = { mean: r.tps_mean, std: r.tps_std, err: r.error, et: r.error_type };
if (!byModel[name].quant && r.quant) byModel[name].quant = r.quant;
if (byModel[name].sizeB == null) byModel[name].sizeB = (r.name_params_b ?? r.params_b);
}
for (const r of allRuns) {
if (r.test || !r.error) continue;
const name = r.model_clean || r.model; if (!byModel[name]) continue;
const cfg = `${r.env}|fa=${r.fa ? 1 : 0}`;
for (const t of ['pp512', 'tg128']) {
if (!byModel[name][t][cfg]) byModel[name][t][cfg] = { mean: null, std: null, err: true, et: r.error_type || 'error' };
}
}
const envs = [...new Set(allRuns.map(r => r.env))].sort();
function envBox(env) {
const roc = env.includes('rocwmma');
const hipBLASTt_off = env.includes('hblt0');
const fa_all_quants = env.includes('fa_all_quants');
const id = `env_${env.replace(/[^a-z0-9_-]/gi, '_')}`;
return `
<div class="colbox">
<label for="${id}">
<input id="${id}" type="checkbox" data-env="${env}"
${/(vulkan_amdvlk|vulkan_radv|rocm6_4_4-rocwmma|rocm7_rc-rocwmma)(?![-\w])/.test(env.trim()) ? 'checked' : ''}>
<span>
<strong>${env}</strong>
${roc ? '<span class="badge roc">rocWMMA</span>' : ''}
${fa_all_quants ? '<span class="badge faall">FA all quants</span>' : ''}
${hipBLASTt_off ? '<span class="badge roc">hipBLASTt OFF</span>' : ''}
</span>
</label>
</div>`;
}
colPicker.innerHTML = envs.map(envBox).join('');
const selectedEnvs = () => [...colPicker.querySelectorAll('input[type="checkbox"]')].filter(i => i.checked).map(i => i.dataset.env);
const quants = [...new Set(Object.values(byModel).map(m => (m.quant || 'Unknown').toUpperCase()))].sort();
quants.forEach(q => { const o = document.createElement('option'); o.value = (q === 'UNKNOWN' ? '' : q); o.textContent = q; quantSel.appendChild(o); });
// --- Dual range slider setup ---
const sizes = Object.values(byModel)
.map(m => m.sizeB)
.filter(v => typeof v === 'number')
.sort((a, b) => a - b);
// force clean integer min/max
const MIN_B = Math.floor(sizes[0] ?? 0);
const MAX_B = Math.ceil(sizes[sizes.length - 1] ?? 100);
[sizeLo, sizeHi].forEach(inp => {
inp.min = MIN_B;
inp.max = MAX_B;
inp.step = 1; // integers only
});
sizeLo.value = MIN_B;
sizeHi.value = MAX_B;
const filters = { sizeLo: MIN_B, sizeHi: MAX_B };
function fmtB(n) {
return `${Number(n).toFixed(0)}B`;
}
function clampRange() {
if (+sizeLo.value > +sizeHi.value) {
if (document.activeElement === sizeLo) {
sizeHi.value = sizeLo.value;
} else {
sizeLo.value = sizeHi.value;
}
}
}
function paintTrack() {
const a = (+sizeLo.value - MIN_B) / (MAX_B - MIN_B) * 100;
const b = (+sizeHi.value - MIN_B) / (MAX_B - MIN_B) * 100;
sizeTrack.style.background = `
linear-gradient(to right,
#e5e5e5 ${a}%,
var(--accent) ${a}%,
var(--accent) ${b}%,
#e5e5e5 ${b}%)`;
}
function updateSizeUI(pushRender = true) {
clampRange();
// Fix rounding: snap to min/max if slider is at extremes
if (Math.abs(sizeLo.value - MIN_B) < 0.0001) sizeLo.value = MIN_B;
if (Math.abs(sizeHi.value - MAX_B) < 0.0001) sizeHi.value = MAX_B;
sizeLoVal.textContent = fmtB(sizeLo.value);
sizeHiVal.textContent = fmtB(sizeHi.value);
filters.sizeLo = +sizeLo.value;
filters.sizeHi = +sizeHi.value;
paintTrack();
if (pushRender) render();
}
sizeLo.addEventListener('input', () => updateSizeUI(true));
sizeHi.addEventListener('input', () => updateSizeUI(true));
updateSizeUI(false);
function setupScrollSync(prefix) {
const top = document.querySelector(`.scroller-top[data-for="${prefix}"]`);
const bottom = document.getElementById(`scroller-${prefix}`);
const spacer = top.querySelector('.scroller-spacer');
if (!top || !bottom || !spacer) return () => { };
function syncWidth() {
// Match the tables full scrollable width
spacer.style.width = bottom.scrollWidth + 'px';
// Keep positions aligned after rerender
top.scrollLeft = bottom.scrollLeft;
}
// Two-way sync
let lock = false;
top.addEventListener('scroll', () => {
if (lock) return; lock = true;
bottom.scrollLeft = top.scrollLeft;
lock = false;
});
bottom.addEventListener('scroll', () => {
if (lock) return; lock = true;
top.scrollLeft = bottom.scrollLeft;
lock = false;
});
// Resize/refresh on content changes
const ro = new ResizeObserver(syncWidth);
ro.observe(bottom);
syncWidth();
// Return a refresher for manual calls after rerender
return syncWidth;
}
// --- Filters ---
function matchesFilters(name) {
const q = (search.value || '').toLowerCase();
if (q && !name.toLowerCase().includes(q)) return false;
const info = byModel[name];
if (quantSel.value && (info.quant || '').toUpperCase() !== quantSel.value.toUpperCase()) return false;
if (info.sizeB != null && (info.sizeB < filters.sizeLo - 1e-9 || info.sizeB > filters.sizeHi + 1e-9)) return false;
return true;
}
function cleanTitle(name, quant) {
const qTokens = ['BF16', 'F16', 'F32', 'FP16', 'MXFP2', 'MXFP4', 'MXFP8'];
if (quant) qTokens.push(quant.toUpperCase());
let t = name;
t = t.replace(/[-_](Q\d+(?:_[A-Z0-9]+)*)/gi, '');
t = t.replace(new RegExp(`[-_]?(${qTokens.join('|')})\\b`, 'gi'), '');
t = t.replace(/-UD\b/gi, '');
t = t.replace(/[-_]{2,}/g, '-')
.replace(/\s{2,}/g, ' ')
.replace(/[-_]+$/, '')
.trim();
return t;
}
function buildColDefs(envSel, mode) {
const cols = [];
for (const e of envSel) {
if (mode === 'both') {
cols.push({ env: e, key: `${e}|fa=0`, label: `${e}<div class="env-sub">FA off</div>` });
cols.push({ env: e, key: `${e}|fa=1`, label: `${e}<div class="env-sub">FA on</div>` });
} else {
const k = `${e}|fa=${mode === 'on' ? 1 : 0}`;
cols.push({ env: e, key: k, label: `${e}<div class="env-sub">FA ${mode === 'on' ? 'on' : 'off'}</div>` });
}
}
return cols;
}
function headerHTML(test, colDefs) {
return `<th class="col-model">Model</th><th class="col-winner">Winner (${test})</th>` +
colDefs.map(c => `<th>${c.label}</th>`).join('');
}
function cellHTML(name, test, key) {
const v = byModel[name][test][key];
if (!v) return '<div class="cell"><span class="muted">—</span></div>';
if (v.err || v.mean == null) return `<div class="cell"><span class="warn">⚠ ${v.et || 'error'}</span></div>`;
return `<div class="cell"><span class="tps">${v.mean.toFixed(2)}</span><span class="std">± ${v.std?.toFixed(2) ?? '—'}</span></div>`;
}
function winnersHTML(name, test, colDefs) {
const vals = [];
for (const c of colDefs) {
const v = byModel[name][test][c.key];
if (v && !v.err && typeof v.mean === 'number') {
vals.push({
env: c.env,
fa: c.key.endsWith('|fa=1') ? 'on' : 'off',
mean: v.mean,
std: (typeof v.std === 'number' ? v.std : 0)
});
}
}
if (!vals.length) return '<span class="muted">—</span>';
let best = vals[0];
for (const v of vals) if (v.mean > best.mean) best = v;
const winners = [];
for (const v of vals) {
const pooled = Math.sqrt((best.std || 0) ** 2 + (v.std || 0) ** 2);
const tol = Math.max(MIN_TOL, K_SIGMA * pooled);
if ((best.mean - v.mean) <= tol) winners.push(v);
}
if (faMode.value !== 'both') {
const envs = [...new Set(winners.map(w => w.env))];
return `<div class="winner-wrap">${envs.map(env =>
`<span class="winner-pill">🏆 ${env}</span>`).join('')}</div>`;
} else {
const byEnv = {};
for (const w of winners) (byEnv[w.env] ||= new Set()).add(w.fa);
return `<div class="winner-wrap">${Object.entries(byEnv).map(([env, fas]) => {
const list = [...fas].sort().map(x => `FA ${x}`).join(' & ');
return `<span class="winner-pill">🏆 ${env} (${list})</span>`;
}).join('')
}</div>`;
}
}
function renderTable(prefix, test) {
const envSel = selectedEnvs(); const mode = faMode.value;
const colDefs = buildColDefs(envSel, mode);
const h = document.querySelector(`#tbl-${prefix} thead tr`) || document.getElementById(`${prefix}-h`);
if (!h) return; // nothing to render into; bail safely
h.innerHTML = headerHTML(test, colDefs);
const tbody = document.querySelector(`#tbl-${prefix} tbody`);
tbody.innerHTML = '';
for (const name of models.filter(matchesFilters)) {
const tr = document.createElement('tr');
const tdM = document.createElement('td');
tdM.className = 'col-model';
const info = byModel[name];
tdM.innerHTML = `<div class="model-title">${cleanTitle(name, info.quant)}</div><div class="model-meta">${info.quant || '—'} · ${(info.sizeB ?? '—')}B</div>`;
tr.appendChild(tdM);
const tdW = document.createElement('td');
tdW.className = 'col-winner';
tdW.innerHTML = winnersHTML(name, test, colDefs);
tr.appendChild(tdW);
for (const c of colDefs) {
const td = document.createElement('td');
td.innerHTML = cellHTML(name, test, c.key);
tr.appendChild(td);
}
tbody.appendChild(tr);
}
}
let refreshScrollPP = null, refreshScrollTG = null;
function render() {
renderTable('pp', 'pp512');
renderTable('tg', 'tg128');
// Update top scrollbar widths after content changes
if (typeof refreshScrollPP === 'function') refreshScrollPP();
if (typeof refreshScrollTG === 'function') refreshScrollTG();
}
[search, quantSel, faMode].forEach(el => el.addEventListener('input', render));
colPicker.addEventListener('change', render);
render();
refreshScrollPP = setupScrollSync('pp');
refreshScrollTG = setupScrollSync('tg');
})();
</script>
<section class="panel compact" id="tables-panel">
<div id="tables"></div>
</section>
<script src="assets/index2.js" type="module"></script>
</body>
</html>
</html>