<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title><![CDATA[JP Reviews]]></title><description><![CDATA[Independent PC hardware and SSD reviews with real-world benchmarks, sustained write tests, and clear analysis that goes beyond advertised speeds.]]></description><link>https://jp-reviews.hashnode.dev</link><generator>RSS for Node</generator><lastBuildDate>Sat, 21 Nov 2026 17:05:03 GMT</lastBuildDate><atom:link href="https://jp-reviews.hashnode.dev/rss.xml" rel="self" type="application/rss+xml"/><language><![CDATA[en]]></language><ttl>60</ttl><item><title><![CDATA[KingSpec P3 512 GB SSD review: fast reads, slow sustained writes]]></title><description><![CDATA[KingSpec P3 512GB SSD review: real-world benchmarks reveal fast burst writes, steep sustained slowdowns, and how TRIM helps restore write performance.]]></description><link>https://jp-reviews.hashnode.dev/kingspec-p3-512-gb-ssd-review-fast-reads-slow-sustained-writes</link><guid isPermaLink="true">https://jp-reviews.hashnode.dev/kingspec-p3-512-gb-ssd-review-fast-reads-slow-sustained-writes</guid><category><![CDATA[kingspec]]></category><category><![CDATA[ssd]]></category><category><![CDATA[review]]></category><dc:creator><![CDATA[Jason Player]]></dc:creator><pubDate>Tue, 17 Nov 2026 00:08:06 GMT</pubDate><content:encoded><![CDATA[<p><strong>Tested 15–16 November 2026 · One sample</strong></p>
<p>This review tests one KingSpec P3 512 GB SATA SSD. We look beyond its advertised peak speeds to measure short transfers, prolonged writes, and performance after the initial fast burst ends. We also test whether idle time, TRIM, or emptying the drive restores write performance.</p>
<p>The tested KingSpec P3 reads large files at approximately <strong>404–430 MiB/s</strong> and starts sequential writes at roughly <strong>375–400 MiB/s</strong>. Sustained writing tells a different story. In two extended post-TRIM tests, speed fell sharply after <strong>35.79–36.75 GiB</strong>. The remaining write periods averaged <strong>26.2–27.5 MiB/s</strong>, with final-minute averages of <strong>19.8–23.4 MiB/s</strong>.</p>
<p>TRIM tells the SSD which filesystem blocks are unused. It made the fast burst larger in a matched comparison, but did not remove the slow tail. Deleting every file, trimming the empty filesystem and testing again produced almost the same result.</p>
<p><strong>Verdict:</strong> the measured performance favors reading existing files and occasional writes. This sample is a poor fit for sustained large transfers or applications that need consistently responsive writes. Its opening burst makes it look much faster than it is over a long write job.</p>
<h2>The Amazon listing and the tested sample</h2>
<p>The supplied <a href="https://www.amazon.de/dp/B0B2K3ZCHH">Amazon.de product page, ASIN B0B2K3ZCHH</a>, identifies the product as a <strong>KingSpec P3 512 GB internal 2.5-inch SATA SSD</strong>. The following details come from a saved listing captured on <strong>16 November 2026</strong>; the displayed price and wording may change.</p>
<table>
<thead>
<tr>
<th>Listing item</th>
<th>Advertised or displayed information</th>
</tr>
</thead>
<tbody><tr>
<td>Capacity and form factor</td>
<td>512 GB, internal 2.5-inch SATA SSD</td>
</tr>
<tr>
<td>Sequential speed</td>
<td>Up to <strong>550 MB/s read</strong> and <strong>500 MB/s write</strong></td>
</tr>
<tr>
<td>NAND</td>
<td><strong>3D TLC NAND</strong>, according to the listing</td>
</tr>
<tr>
<td>Features</td>
<td>TRIM, SMART, garbage collection, wear leveling and NCQ support claimed</td>
</tr>
<tr>
<td>Warranty</td>
<td><strong>Three-year limited warranty</strong> and lifetime technical support claimed</td>
</tr>
<tr>
<td>Price in the saved page</td>
<td><strong>€62.99</strong>; a listing snapshot, not a verified purchase price</td>
</tr>
</tbody></table>
<h3>What the advertised speeds mean for this review</h3>
<p>The listing uses decimal <strong>MB/s</strong>; fio reports binary <strong>MiB/s</strong>. Its 550 MB/s read and 500 MB/s write claims correspond to approximately <strong>524.5 MiB/s</strong> and <strong>476.8 MiB/s</strong>. The saved listing does not specify a sustained minimum speed or how much data can be written before a slowdown.</p>
<h2>Test setup and reading the results</h2>
<p>The sample reports model <strong>P3-512</strong>, firmware <strong>HT5710A5</strong>, consistent with the supplied product listing.</p>
<table>
<thead>
<tr>
<th>Item</th>
<th>Test configuration</th>
</tr>
</thead>
<tbody><tr>
<td>Reported capacity</td>
<td>512,110,190,592 bytes: 512.11 decimal GB, approximately 476.94 GiB</td>
</tr>
<tr>
<td>Filesystem</td>
<td>ext4, with continuous discard disabled</td>
</tr>
<tr>
<td>Software</td>
<td>Linux 7.3.1, fio 3.42</td>
</tr>
<tr>
<td>I/O method</td>
<td>Direct I/O, one job, changing write buffers</td>
</tr>
<tr>
<td>Block-size overview</td>
<td>4 KiB, 16 KiB, 128 KiB and 1 MiB; QD1, with selected QD32 tests</td>
</tr>
<tr>
<td>Long-write tests</td>
<td>Sequential 1 MiB requests at QD16</td>
</tr>
</tbody></table>
<p><strong>MiB/s</strong> means 1,048,576 bytes per second; <strong>GiB</strong> means 1,073,741,824 bytes. Sequential tests access consecutive parts of the test file; random tests access scattered locations. <strong>IOPS</strong> counts completed I/O operations per second. Queue depth, or <strong>QD</strong>, is the number of outstanding requests. A higher QD lets the drive work on more requests at once, but can also leave each request waiting longer.</p>
<p>Direct I/O bypasses the operating system's page cache, but does not disable caches inside the drive. In the later burst-capacity tests, a fast interval means at least <strong>250 MiB/s</strong>, and a slowdown is confirmed after <strong>ten seconds below 100 MiB/s</strong>. The recorded byte count includes all data written before that slow stretch, not only data written above the fast threshold.</p>
<h2>Read performance is the strongest result</h2>
<p>Sequential 1 MiB reads reached <strong>403.95 MiB/s at QD1</strong> and <strong>429.66 MiB/s at QD32</strong>. Small random reads delivered <strong>4,794 IOPS at 4 KiB/QD1</strong>, increasing to <strong>29,310 IOPS at QD32</strong>.</p>
<img src="https://cdn.hashnode.com/uploads/covers/6afb086464756f0a8cb6d502/e9297977-07fa-4cd3-adb9-9f59cdff895b.png" alt="Sequential and random throughput by block size at QD1" style="display:block;margin:0 auto" />

<p>The block-size overview used an initialized <strong>8 GiB region</strong>, with <strong>eight seconds per test</strong>, after a five-minute sustained write. Reads ran before the write groups. These are measurements of a previously stressed device, not independently refreshed, best-case scores for every row.</p>
<table>
<thead>
<tr>
<th>Block size, QD1</th>
<th align="right">Sequential read</th>
<th align="right">Sequential write</th>
<th align="right">Random read</th>
<th align="right">Random write</th>
</tr>
</thead>
<tbody><tr>
<td>4 KiB</td>
<td align="right">41.36 MiB/s</td>
<td align="right">8.70 MiB/s</td>
<td align="right">18.73 MiB/s</td>
<td align="right">11.79 MiB/s</td>
</tr>
<tr>
<td>16 KiB</td>
<td align="right">116.45 MiB/s</td>
<td align="right">10.58 MiB/s</td>
<td align="right">45.91 MiB/s</td>
<td align="right">14.71 MiB/s</td>
</tr>
<tr>
<td>128 KiB</td>
<td align="right">307.42 MiB/s</td>
<td align="right">82.59 MiB/s</td>
<td align="right">205.37 MiB/s</td>
<td align="right">37.57 MiB/s</td>
</tr>
<tr>
<td>1 MiB</td>
<td align="right">403.95 MiB/s</td>
<td align="right">15.61 MiB/s</td>
<td align="right">356.46 MiB/s</td>
<td align="right">11.39 MiB/s</td>
</tr>
</tbody></table>
<p>Write performance varied considerably between short samples. The 128 KiB write result should not be read as evidence that this block size permanently fixes the slowdown: cache recovery and background activity could change the drive's state between tests.</p>
<h3>More queued requests increased read throughput in these samples</h3>
<table>
<thead>
<tr>
<th>QD32 test</th>
<th align="right">Throughput</th>
<th align="right">IOPS</th>
<th align="right">Mean completion latency</th>
<th align="right">p99 completion latency</th>
</tr>
</thead>
<tbody><tr>
<td>Sequential read, 1 MiB</td>
<td align="right">429.66 MiB/s</td>
<td align="right">430</td>
<td align="right">71.97 ms</td>
<td align="right">77.07 ms</td>
</tr>
<tr>
<td>Random read, 4 KiB</td>
<td align="right">114.49 MiB/s</td>
<td align="right">29,310</td>
<td align="right">1.086 ms</td>
<td align="right">2.277 ms</td>
</tr>
<tr>
<td>Sequential write, 1 MiB</td>
<td align="right">12.67 MiB/s</td>
<td align="right">13</td>
<td align="right"><strong>2.16 s</strong></td>
<td align="right"><strong>3.00 s</strong></td>
</tr>
<tr>
<td>Random write, 4 KiB</td>
<td align="right">19.46 MiB/s</td>
<td align="right">4,983</td>
<td align="right">6.164 ms</td>
<td align="right">147.85 ms</td>
</tr>
</tbody></table>
<p>Raising queue depth did not restore fast large writes in these samples. The QD32 sequential-write result means that, with 32 requests outstanding, individual 1 MiB requests took an average of <strong>2.16 seconds</strong> to complete. It is a measure of delays under that load, not the time for a single isolated write. At QD1, 1 MiB sequential writes were also slow: <strong>15.61 MiB/s</strong>, with <strong>63.79 ms</strong> mean and <strong>149.95 ms</strong> p99 completion latency.</p>
<p>Mean latency is the average request completion time; <strong>p99</strong> is the completion time at or below which approximately 99% of sampled requests fell. These are fio completion latencies, excluding submission overhead. Jobs flushed at the end; this was not a per-write <code>fsync</code> durability-latency benchmark.</p>
<h2>TRIM restored a larger fast-write burst</h2>
<p>The TRIM experiment asked whether telling the SSD which deleted space it could reclaim would restore more fast-write capacity than waiting alone. It compared two runs: a control without TRIM, followed by a run with TRIM.</p>
<p>Both runs first wrote until the drive was slow, then deleted the temporary test file. Each then had a two-minute period: the control stayed idle, while the TRIM run issued <code>fstrim</code> and waited out the remaining time. Both recreated the test file, waited another two minutes, and began writing again. This gave them approximately <strong>four minutes of recovery in total</strong>—measured at <strong>241.37 and 241.80 seconds</strong>—with identical filesystem occupancy before the probes. Time spent executing TRIM was included in its two-minute period, preventing a longer wait from being mistaken for a TRIM benefit.</p>
<img src="https://cdn.hashnode.com/uploads/covers/6afb086464756f0a8cb6d502/5f66e223-376c-4270-98c3-c3bf2a7f1e07.png" alt="Historical idle-recovery results and the matched TRIM comparison" style="display:block;margin:0 auto" />

<p>The historical idle tests in this chart provide context. Only the two rows marked <strong>paired control</strong> and <strong>paired test</strong> belong to this matched comparison.</p>
<table>
<thead>
<tr>
<th>Matched experiment</th>
<th align="right">Initial write speed</th>
<th align="right">Written before sustained slowdown</th>
</tr>
</thead>
<tbody><tr>
<td>Control, no TRIM</td>
<td align="right">382.1 MiB/s</td>
<td align="right"><strong>4.53 GiB</strong></td>
</tr>
<tr>
<td>After TRIM</td>
<td align="right">391.7 MiB/s</td>
<td align="right"><strong>38.43 GiB</strong></td>
</tr>
</tbody></table>
<p>The post-TRIM burst was <strong>8.48 times as large</strong>, a difference of <strong>33.90 GiB</strong>. Peak speed changed little. The kernel recorded 2,799 completed discard requests covering the same <strong>340.4 GiB</strong> reported by <code>fstrim</code>, confirming successful discard requests in this setup.</p>
<p>This is one ordered comparison, not a guaranteed improvement factor. The control ran first, and TRIM cannot simply be undone to restore an identical internal flash state. Completed discard requests also do not reveal exactly when the SSD finished its internal cleanup.</p>
<h2>Six-minute writes reveal the real tradeoff</h2>
<p>The capacity probes above stopped once they confirmed a slowdown. The extended tests asked what happens <strong>if the transfer continues</strong>. They wrote for up to six minutes or 128 GiB, whichever came first, and continued after the speed drop. Each issued filesystem TRIM within a two-minute preparation period, allocated space for a 128 GiB test file, and allowed another two minutes of idle time before writing. Allocating the file only reserved filesystem space, it did not write 128 GiB of test data in advance.</p>
<img src="https://cdn.hashnode.com/uploads/covers/6afb086464756f0a8cb6d502/5d23b6cd-27c2-4729-b469-da0b6ffa11b8.png" alt="Two post-TRIM sustained-write traces, including the empty-filesystem retest" style="display:block;margin:0 auto" />

<table>
<thead>
<tr>
<th>Measurement</th>
<th align="right">Post-TRIM test</th>
<th align="right">Empty-filesystem retest</th>
</tr>
</thead>
<tbody><tr>
<td>Initial write speed</td>
<td align="right">384.2 MiB/s</td>
<td align="right">387.1 MiB/s</td>
</tr>
<tr>
<td>Written before the slow stretch</td>
<td align="right"><strong>36.75 GiB</strong></td>
<td align="right"><strong>35.79 GiB</strong></td>
</tr>
<tr>
<td>Slowdown confirmed</td>
<td align="right">109.1 s</td>
<td align="right">107.8 s</td>
</tr>
<tr>
<td>Average after confirmation</td>
<td align="right"><strong>26.2 MiB/s</strong></td>
<td align="right"><strong>27.5 MiB/s</strong></td>
</tr>
<tr>
<td>Time measured after confirmation</td>
<td align="right">251.5 s</td>
<td align="right">252.4 s</td>
</tr>
<tr>
<td>Final-minute average</td>
<td align="right"><strong>19.8 MiB/s</strong></td>
<td align="right"><strong>23.4 MiB/s</strong></td>
</tr>
<tr>
<td>Total written</td>
<td align="right">43.30 GiB</td>
<td align="right">42.70 GiB</td>
</tr>
<tr>
<td>Whole-run average</td>
<td align="right">122.9 MiB/s</td>
<td align="right">121.4 MiB/s</td>
</tr>
</tbody></table>
<p>“Initial write speed” averages approximately the first five seconds. “Written before the slow stretch” counts bytes accepted before the start of the stretch that later triggered confirmation. The post-confirmation average measures the remainder of the job; the final-minute average measures its last 60 seconds. Both runs reached the time limit well before writing 128 GiB.</p>
<p>The speed drop begins before the confirmation timestamp: the detector requires ten seconds below 100 MiB/s. Afterward, writes remain variable, with occasional brief recoveries toward the original speed. Those recoveries are included in the post-drop averages.</p>
<p>The whole-run averages above 120 MiB/s conceal the long, slow remainder. For transfers that outlast the fast burst, the measured <strong>20–27 MiB/s tail</strong> is more relevant than the opening speed.</p>
<h3>Emptying the drive did not restore the original burst</h3>
<p>For the retest, all existing files were deleted and the filesystem was verified empty before TRIM. <code>fstrim</code> reported <strong>380.4 GiB</strong> submitted. The same test then produced a <strong>35.79 GiB</strong> burst—close to the preceding 36.75 GiB result.</p>
<p>That matters because the first long-write experiment had sustained approximately <strong>375–379 MiB/s for about 3 minutes 50 seconds</strong>, reaching an estimated <strong>84 GiB</strong> before its persistent slowdown. It was eventually cancelled after roughly <strong>143 GiB in 51.6 minutes</strong>, averaging <strong>47.4 MiB/s</strong>.</p>
<p>The initial 84 GiB estimate comes from an older bandwidth log and is less precise than the later cumulative-byte measurements. Nevertheless, the empty-filesystem retest shows that deleting files and trimming did <strong>not</strong> reproduce that favorable starting condition. The presence of existing files alone does not explain the difference. These measurements do not identify the drive's internal mechanism or establish permanent wear as the cause; deletion and TRIM do not guarantee a return to its original internal flash state.</p>
<h2>Does idle time restore the fast-write burst?</h2>
<p>The next practical question was: <strong>after a large write has slowed the drive down, how long must it sit unused before the next write can run fast again?</strong> The first long-write test had shown an unusually large burst of roughly 84 GiB. These recovery trials tested whether simply leaving the drive powered on could restore that behavior, without issuing TRIM.</p>
<p>Each trial followed the same basic sequence:</p>
<ol>
<li>Write until a slow-write condition was observed.</li>
<li>Stop writing and leave the drive powered on for the selected interval. Keep the test file allocated; do not delete it or issue TRIM during the wait.</li>
<li>Resume sequential writes and measure the initial speed and the amount written before the next confirmed slowdown.</li>
</ol>
<p>The drive was written back into a slow condition <strong>before every wait</strong>. The two-minute and five-minute entries therefore represent separate trials, not successive checkpoints during one uninterrupted recovery period. Small amounts of host I/O during the waits were logged, as discussed below.</p>
<p>Here, <strong>recovery</strong> means the return of useful fast-write capacity for the next transfer. The idle period itself has no write throughput to measure. The result is the size and speed of the burst that follows it.</p>
<table>
<thead>
<tr>
<th>Intentional idle time</th>
<th align="right">Initial speed when writes resumed</th>
<th align="right">Written before the next slow stretch</th>
</tr>
</thead>
<tbody><tr>
<td>None</td>
<td align="right">79.3 MiB/s</td>
<td align="right">No qualifying fast burst</td>
</tr>
<tr>
<td>30 seconds</td>
<td align="right">13.7 MiB/s</td>
<td align="right">0.75 GiB, including a later brief fast interval</td>
</tr>
<tr>
<td>2 minutes</td>
<td align="right">396.8 MiB/s</td>
<td align="right"><strong>18.28 GiB</strong></td>
</tr>
<tr>
<td>5 minutes</td>
<td align="right">394.4 MiB/s</td>
<td align="right"><strong>22.34 GiB</strong></td>
</tr>
<tr>
<td>15 minutes, separate run</td>
<td align="right">387.6 MiB/s</td>
<td align="right"><strong>18.47 GiB</strong></td>
</tr>
</tbody></table>
<p>Two minutes of idle time restored a useful burst in that trial. Five minutes restored somewhat more, but the separate fifteen-minute trial did not improve on it. The 30-second trial started slowly, accelerated briefly, and slowed again; its 0.75 GiB figure does not mean that all those bytes were written continuously at high speed.</p>
<p>These were single trials with different prior write histories, and the slowdown detector evolved during testing. The observed slow starting conditions do not guarantee identical internal flash states. Consequently, the results do not establish an optimum wait or show that a longer wait makes recovery worse. They show that <strong>idle time alone did not reliably restore the original roughly 84 GiB burst</strong>, even after fifteen minutes.</p>
<h2>Temperature and health checks</h2>
<p>SMART is the drive's built-in health reporting. The saved paired and extended TRIM tests showed <strong>no disk I/O errors</strong> in their kernel logs.</p>
<p>SMART returned PASSED, and reported reallocated, pending, offline-uncorrectable and CRC error counters remained zero. Temperature snapshots in the two extended post-TRIM tests ranged from <strong>35°C to 39°C</strong>.</p>
<p>Idle host I/O was recorded. The paired experiment had approximately <strong>4.1 MiB of writes in each arm</strong> during the post-allocation wait, with no additional discards; the extended tests recorded similar activity. Its source was not attributed, but the closely matched volumes are small relative to the observed differences in write capacity.</p>
<h2>Verdict</h2>
<p>The strongest result is reading large files at approximately <strong>400 MiB/s</strong>. Short writes can also be fast, especially after TRIM, but burst length depends strongly on prior activity. Across the paired TRIM test and two extended post-TRIM tests, later bursts clustered around <strong>36–38 GiB</strong>, followed by low throughput and occasional recovery spikes.</p>
<p>For sustained capture, scratch workloads or repeated large writes, the slowdown and long completion delays are substantial disadvantages. Reading existing content is a better match for the measured performance.</p>
<h2>How the results were calculated</h2>
<p>The overview used fio <code>libaio</code>, direct I/O, one job and changing buffers (<code>refill_buffers=1</code>). Recovery preparation accepted a trailing 30-second window with an average below 100 MiB/s, at least 80% of its time below that threshold, and a latest sample below it. Capacity probes confirmed a slow stretch after ten seconds below 100 MiB/s. Some earlier idle-recovery trials used a stricter consecutive-slow criterion, they are separate historical observations rather than points on a controlled recovery curve.</p>
<p>Charts use linear axes and include brief speed recoveries rather than filtering them out. The post-drop means are time-weighted and begin after slowdown confirmation. Values in the article are rounded.</p>
]]></content:encoded></item></channel></rss>