Donato Capitella
154 lines
7.4 KiB
HTML
154 lines
7.4 KiB
HTML
<!doctype html>
|
||
<html lang="en">
|
||
|
||
<head>
|
||
<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” — Benchmark 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>
|
||
<div class="legend">
|
||
<label>Legend</label>
|
||
<div class="legend-pills">
|
||
<button id="hipblas-modal-open" type="button" class="chip small legend-pill legend-pill-default">
|
||
hipBLASLt vs hblt0
|
||
</button>
|
||
<button id="rpc-modal-open" type="button" class="chip small legend-pill legend-pill-rpc">
|
||
RPC · dual server
|
||
</button>
|
||
<button id="rocwmma-modal-open" type="button" class="chip small legend-pill legend-pill-rocwmma">
|
||
rocWMMA
|
||
</button>
|
||
</div>
|
||
</div>
|
||
</header>
|
||
|
||
<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 class="control">
|
||
<label for="filter-quant">Quant</label>
|
||
<select id="filter-quant">
|
||
<option value="">Any</option>
|
||
</select>
|
||
</div>
|
||
<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="range-values">
|
||
<span id="sizeLoVal">0B</span> – <span id="sizeHiVal">0B</span>
|
||
</div>
|
||
</div>
|
||
</section>
|
||
|
||
<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>
|
||
</section>
|
||
|
||
<section class="panel compact" id="tables-panel">
|
||
<div id="tables"></div>
|
||
</section>
|
||
|
||
<div id="hipblas-modal" class="modal hidden" role="dialog" aria-modal="true" aria-labelledby="hipblas-title">
|
||
<div class="modal-content">
|
||
<button id="hipblas-modal-close" class="modal-close" aria-label="Close dialog">×</button>
|
||
<h2 id="hipblas-title">hipBLASLt & hblt0 explained</h2>
|
||
<p>The ROCm toolboxes ship with <code>ROCBLAS_USE_HIPBLASLT=1</code> by default. This forces rocBLAS to
|
||
prefer
|
||
the hipBLASLt kernel library, which historically delivered the best throughput on gfx1151 (Strix Halo).
|
||
</p>
|
||
<p>Rows tagged with <code>__hblt0</code> were re-run with <code>ROCBLAS_USE_HIPBLASLT=0</code>, letting
|
||
rocBLAS
|
||
auto-select between hipBLASLt, Tensile, or other kernel providers. These runs show how performance
|
||
shifts when
|
||
the tuned hipBLASLt path is disabled.</p>
|
||
<p>hipBLASLt is AMD's LT (low-level tuned) matmul backend, optimized for transformer workloads. Disabling it
|
||
can
|
||
expose regressions or improvements depending on driver versions, so both configurations are published
|
||
for
|
||
comparison.</p>
|
||
</div>
|
||
</div>
|
||
|
||
<div id="rpc-modal" class="modal hidden" role="dialog" aria-modal="true" aria-labelledby="rpc-title">
|
||
<div class="modal-content">
|
||
<button id="rpc-modal-close" class="modal-close" aria-label="Close dialog">×</button>
|
||
<h2 id="rpc-title">RPC · dual server</h2>
|
||
<p>These results were produced with two Strix Halo systems (Framework Desktop + HP G1a workstation, each
|
||
128 GB)
|
||
connected over 5 Gbps Ethernet. One runs <code>rpc-server</code> from llama.cpp; the other runs
|
||
<code>llama-bench --rpc</code>.
|
||
</p>
|
||
<p>This setup allows distributed inference, splitting large GGUF models across both machines. The metric
|
||
shows what
|
||
you can expect when latency is limited by the network and the workload is balanced between two RPC
|
||
participants.</p>
|
||
</div>
|
||
</div>
|
||
|
||
<div id="rocwmma-modal" class="modal hidden" role="dialog" aria-modal="true" aria-labelledby="rocwmma-title">
|
||
<div class="modal-content">
|
||
<button id="rocwmma-modal-close" class="modal-close" aria-label="Close dialog">×</button>
|
||
<h2 id="rocwmma-title">rocWMMA variants</h2>
|
||
<p>Backends labeled <code>-rocwmma</code> are rebuilt with AMD's rocWMMA library, which unlocks matrix
|
||
multiply
|
||
pipelines accelerated via wave matrix multiply-accumulate (WMMA) instructions.</p>
|
||
<p>rocWMMA kernels can significantly accelerate BF16/F16 workloads on RDNA3 but may trade stability or
|
||
memory
|
||
usage; comparing plain toolboxes against <code>-rocwmma</code> ones highlights the benefit or cost.</p>
|
||
</div>
|
||
</div>
|
||
|
||
<div id="rocwmma-impr-modal" class="modal hidden" role="dialog" aria-modal="true"
|
||
aria-labelledby="rocwmma-impr-title">
|
||
<div class="modal-content">
|
||
<button id="rocwmma-impr-modal-close" class="modal-close" aria-label="Close dialog">×</button>
|
||
<h2 id="rocwmma-impr-title">rocWMMA-improved builds</h2>
|
||
<p>Toolboxes tagged <code>-rocwmma-improved</code> bake in an experimental llama.cpp patch that retunes
|
||
rocWMMA
|
||
kernels for long-context throughput on Strix Halo.</p>
|
||
<p>Patch reference: <a
|
||
href="https://github.com/hjc4869/llama.cpp/commit/12bb5c371bd3c647ef75e8e13de9e311edba604d"
|
||
target="_blank" rel="noreferrer">12bb5c371bd3</a>. These builds often run faster for 32k+ contexts,
|
||
but
|
||
the changes are not upstream and may be unstable.</p>
|
||
</div>
|
||
</div>
|
||
|
||
<script src="assets/index2.js" type="module"></script>
|
||
</body>
|
||
|
||
</html> |