<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: 梁育水</title>
    <description>The latest articles on DEV Community by 梁育水 (@shayne914816).</description>
    <link>https://dev.to/shayne914816</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4249167%2Fb3f3e661-2328-48ea-b4f7-e39d0ac03f1c.png</url>
      <title>DEV Community: 梁育水</title>
      <link>https://dev.to/shayne914816</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/shayne914816"/>
    <language>en</language>
    <item>
      <title>QSFP28 vs QSFP56: Port, Cable and Cost Differences Explained</title>
      <dc:creator>梁育水</dc:creator>
      <pubDate>Thu, 19 Nov 2026 02:35:34 +0000</pubDate>
      <link>https://dev.to/shayne914816/qsfp28-vs-qsfp56-port-cable-and-cost-differences-explained-54lo</link>
      <guid>https://dev.to/shayne914816/qsfp28-vs-qsfp56-port-cable-and-cost-differences-explained-54lo</guid>
      <description>&lt;p&gt;QSFP28 and QSFP56 share the same physical cage, the same four-lane electrical architecture and the same I2C management interface — what separates them is modulation. QSFP28 delivers 100 Gb/s as 4 x 25 Gb/s NRZ; QSFP56 delivers 200 Gb/s as 4 x 50 Gb/s PAM4. Pick QSFP28 when your estate is 25G/100G, your servers are on PCIe 3.0, and your fiber plant was built for NRZ optics. Pick QSFP56 when you need 200G per port or a 2 x 100G breakout, your hosts have PCIe 4.0 x16 slots, and your cabling is clean enough to survive PAM4. Every other difference — optics count, cable length ceilings, switch density, total cost — falls out of that single choice, so it is worth ten minutes of attention before you buy either.&lt;/p&gt;

&lt;h2&gt;
  
  
  What Actually Differs Between QSFP28 and QSFP56
&lt;/h2&gt;

&lt;p&gt;Both form factors expose four transmit and four receive lanes in a 38-pin cage that measures the same in the rack. That physical identity is why the confusion exists: you can slide a QSFP28 module into a QSFP56 port and it will seat perfectly. Electrically the two are not interchangeable in the direction people assume. A QSFP56 port can generally host a QSFP28 module and run it at 100G, because the port's retimer or gearbox can clock down to 25G NRZ per lane. A QSFP28 port cannot make a QSFP56 module run at 200G, because the host silicon has no PAM4 encoder on those lanes and the switch ASIC was never specified for 50G per lane signalling. In practice: upgrades go one way only, from NRZ to PAM4, never back.&lt;/p&gt;

&lt;p&gt;The naming follows line rate per lane, not total throughput. The "28" in QSFP28 refers to roughly 28 Gb/s of raw line rate per lane, of which about 25 Gb/s is usable payload after 64b/66b encoding. The "56" in QSFP56 refers to roughly 56 Gb/s of raw line rate per lane, of which about 50 Gb/s is payload. Multiply by four lanes and you get the headline 100G and 200G figures that vendors quote.&lt;/p&gt;

&lt;h2&gt;
  
  
  Lane Encoding: Why PAM4 Is Not Just "Faster NRZ"
&lt;/h2&gt;

&lt;p&gt;NRZ encodes one bit per symbol by comparing signal voltage against a single threshold. PAM4 encodes two bits per symbol using four distinct voltage levels, which doubles the bit rate at the same symbol rate — and therefore at the same lane count and roughly the same Nyquist frequency. That is the entire commercial argument for QSFP56: double the bandwidth per port without doubling the fiber count, the connector count, or the switch port count.&lt;/p&gt;

&lt;p&gt;The price of that trick is margin. Four levels mean one third of the vertical eye height of NRZ, so a PAM4 link tolerates less insertion loss, less crosstalk and less connector contamination before errors appear. PAM4 links also depend on forward error correction at the physical layer (Reed-Solomon FEC, often referred to as KP4 at 50G per lane) to hold a usable bit error rate. Two practical consequences follow. First, both ends of a 200G link must agree on FEC settings; a mismatch is one of the most common reasons a 200G port stays down while the optics look healthy. Second, a fiber plant that ran 100G cleanly for years may need cleaning, re-termination or a patch-panel inspection before it will carry 200G — a cost item that never appears on the bill of materials.&lt;/p&gt;

&lt;h2&gt;
  
  
  Port Capacity and Density at the Switch Level
&lt;/h2&gt;

&lt;p&gt;The same 1U chassis can look dramatically different depending on which port type it carries. A 32-port QSFP28 switch is a 3.2 Tb/s box; a 32-port QSFP56 switch is a 6.4 Tb/s box in the same rack unit, with the same number of cages, the same cable count and a very similar power draw. That density jump is the main reason 200G platforms displaced 100G platforms in spine roles almost immediately, even where the leaf layer stayed at 100G.&lt;/p&gt;

&lt;p&gt;Port count also determines your uplink story. A 64-port QSFP28 leaf gives you 64 x 100G of capacity; a 64-port QSFP56 leaf gives you 64 x 200G. At the same oversubscription ratio, the QSFP56 box can support twice the downstream bandwidth, or half the spine uplinks for the same downstream load — which is where the real savings in optics and switch ports appear.&lt;/p&gt;

&lt;h2&gt;
  
  
  QSFP28 vs QSFP56 at a Glance
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Attribute&lt;/th&gt;
&lt;th&gt;QSFP28&lt;/th&gt;
&lt;th&gt;QSFP56&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Lane signalling&lt;/td&gt;
&lt;td&gt;25 Gb/s NRZ&lt;/td&gt;
&lt;td&gt;50 Gb/s PAM4&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Lanes used&lt;/td&gt;
&lt;td&gt;4&lt;/td&gt;
&lt;td&gt;4&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Speed per port&lt;/td&gt;
&lt;td&gt;100 Gb/s&lt;/td&gt;
&lt;td&gt;200 Gb/s&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Physical cage&lt;/td&gt;
&lt;td&gt;QSFP 38-pin, 4-lane&lt;/td&gt;
&lt;td&gt;QSFP 38-pin, 4-lane (same)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Management interface&lt;/td&gt;
&lt;td&gt;I2C EEPROM per SFF-8636&lt;/td&gt;
&lt;td&gt;I2C EEPROM per SFF-8636&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Typical breakout&lt;/td&gt;
&lt;td&gt;4 x 25G&lt;/td&gt;
&lt;td&gt;2 x 100G, or 4 x 50G where supported&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Passive DAC reach (NVIDIA LinkX family)&lt;/td&gt;
&lt;td&gt;Up to 5 m&lt;/td&gt;
&lt;td&gt;Up to 2.5 m&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Active copper reach&lt;/td&gt;
&lt;td&gt;3 m to 100 m&lt;/td&gt;
&lt;td&gt;3 m to 100 m&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Multimode fiber reach (SR4)&lt;/td&gt;
&lt;td&gt;100 m on OM4&lt;/td&gt;
&lt;td&gt;100 m on OM4&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Singlemode reach&lt;/td&gt;
&lt;td&gt;500 m, 2 km, 10 km, 40 km classes&lt;/td&gt;
&lt;td&gt;500 m, 2 km classes&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;FEC requirement&lt;/td&gt;
&lt;td&gt;Optional on short NRZ links&lt;/td&gt;
&lt;td&gt;Effectively mandatory at the PHY&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Host interface that feeds it comfortably&lt;/td&gt;
&lt;td&gt;PCIe 3.0 x16&lt;/td&gt;
&lt;td&gt;PCIe 4.0 x16&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Representative switch&lt;/td&gt;
&lt;td&gt;SN2700 (32 x QSFP28, 3.2 Tb/s)&lt;/td&gt;
&lt;td&gt;SN3700 (32 x QSFP56, 6.4 Tb/s)&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;h2&gt;
  
  
  Cable Reach: Where the Two Diverges Most
&lt;/h2&gt;

&lt;p&gt;This is the difference that bites procurement teams, because engineers compare per-port speeds and assume the cabling plan carries over unchanged. It does not. In the NVIDIA LinkX cable and transceiver families, 100G QSFP28 passive copper is specified up to 5 m, while 200G QSFP56 passive copper is specified up to 2.5 m. Both use the same 4-lane architecture, but PAM4 burns the signal-integrity headroom that twinax relies on, so the passive copper ceiling drops by half.&lt;/p&gt;

&lt;p&gt;If your racks are wired with 3 m passive DACs at 100G — an extremely common in-rack standard — those exact cables will not be qualified for 200G. Active copper and active optical assemblies cover 3 m to 100 m for both generations, and multimode SR4 optics reach 100 m on OM4 at either speed. Singlemode is where the gap widens again: the 100G family spans 500 m, 2 km, 10 km and 40 km reaches, while the mainstream 200G family covers 500 m and 2 km classes. Long-haul campus or cross-building runs therefore stay on 100G optics unless you step to 400G platforms with a different port type entirely.&lt;/p&gt;

&lt;h2&gt;
  
  
  Breakout: 4 x 25G Versus 2 x 100G
&lt;/h2&gt;

&lt;p&gt;Breakout is often the deciding factor rather than raw speed, because most racks are still full of 25G hosts. A single 100G QSFP28 port breaks out four ways into 4 x 25G, which is why 100G leaves remain the natural home for 25G server fleets. A 200G QSFP56 port breaks out two ways into 2 x 100G — the classic HDR-to-2 x HDR100 splitter cable — or four ways into 4 x 50G where the platform supports it. Notice what is missing: there is no clean 200G-to-25G path. If your rack is 25G-dense, a QSFP56 leaf forces you either to aggregate through a 100G-capable port mode or to keep a QSFP28 layer in place.&lt;/p&gt;

&lt;p&gt;That asymmetry explains a pattern visible in most AI and HPC builds: 200G or 400G at the spine for east-west GPU traffic, 100G at the leaf for general compute and storage, and 25G breakout for the long tail of management and legacy hosts. Buying a 200G leaf before the hosts can use it usually converts a bandwidth upgrade into an expensive port-mode exercise.&lt;/p&gt;

&lt;h2&gt;
  
  
  Switch Platforms: Which Chassis Exposes Which Port
&lt;/h2&gt;

&lt;p&gt;The port type is a property of the chassis, not a preference you can configure. In the Mellanox Spectrum line most commonly found in secondary-market racks, Spectrum-1 platforms such as the &lt;a href="https://zhongxindata.hk/products/mellanox-sn2700-cs2ros-100gbe-32-port-qsfp28-switch-used/" rel="noopener noreferrer"&gt;SN2700 with 32 QSFP28 ports&lt;/a&gt; are 100G-only at 3.2 Tb/s. Spectrum-2 appears in both flavours: the 64-port QSFP28 SN3800 for dense 100G, and the &lt;a href="https://zhongxindata.hk/products/mellanox-sn3700-vs2r-200gbe-32-port-qsfp56-switch-used/" rel="noopener noreferrer"&gt;SN3700 with 32 QSFP56 ports&lt;/a&gt; for 6.4 Tb/s at 200G per port. Spectrum-3 scales the QSFP56 design up to 64 ports and 12.8 Tb/s in the &lt;a href="https://zhongxindata.hk/products/mellanox-sn4600-vs2r-200gbe-64-port-qsfp56-switch-2u-used/" rel="noopener noreferrer"&gt;SN4600&lt;/a&gt;, which is the usual pick when a spine has to absorb several hundred 200G leaf uplinks.&lt;/p&gt;

&lt;p&gt;Two practical notes when comparing these on a spec sheet. First, the QSFP56 chassis is generally backward compatible at the port level, so a 200G leaf can host 100G optics today and move to 200G later without a hardware swap — that forward path is worth paying for only if you actually intend to use it. Second, check the airflow SKU and the power budget, not just the port count; a 64-port 200G platform draws meaningfully more than a 32-port 100G unit, and connector-side airflow direction must match the aisle layout or the rack will run hot regardless of what the datasheet promises.&lt;/p&gt;

&lt;h2&gt;
  
  
  Adapter Side: Matching the NIC to the Port Type
&lt;/h2&gt;

&lt;p&gt;Host adapters are where most mismatches actually get discovered, after the switches have shipped. A ConnectX-5 class adapter such as the &lt;a href="https://zhongxindata.hk/products/100gbe-dual-port-network-card-used-for-mellanox-cx5-mcx516a/" rel="noopener noreferrer"&gt;dual-port 100G MCX516A&lt;/a&gt; is a QSFP28 device on PCIe 3.0 x16, which suits a 100G leaf perfectly and is the single most common 100G host pairing in existing estates. A ConnectX-6 class QSFP56 adapter, like the &lt;a href="https://zhongxindata.hk/products/100gbe-infiniband-hdr100-dual-port-network-card-used-for-mellanox-cx6-mcx653106a-ecat-qsfp56-pcie-4-0-nic/" rel="noopener noreferrer"&gt;MCX653106A in QSFP56 form&lt;/a&gt;, needs PCIe 4.0 x16 to feed 200G without bottlenecking the bus; on a PCIe 3.0 slot it will link, but you will not see the bandwidth you paid for. Older 100G single-port designs such as the &lt;a href="https://zhongxindata.hk/products/mellanox-mcx4621a-acab-100gbe-adapter/" rel="noopener noreferrer"&gt;MCX4621A&lt;/a&gt; remain useful for hosts that only need one high-speed uplink.&lt;/p&gt;

&lt;p&gt;Budget the adapter refresh with the switch refresh, or explicitly plan to run the new 200G switch at 100G for a phase. The second option is legitimate and common, but it only pays off if the host fleet refresh is genuinely funded — otherwise you are carrying the power and cost of 200G-capable silicon while moving 100G traffic.&lt;/p&gt;

&lt;h2&gt;
  
  
  Mixing Modules: Compatibility Rules Worth Knowing
&lt;/h2&gt;

&lt;p&gt;A QSFP56 port hosting a QSFP28 module at 100G is normally fine and is a sensible transition tactic. What is not fine is assuming automatic negotiation. Port speed, FEC mode and, on some platforms, the module's encoded speed rating must be configured explicitly; auto-negotiation between NRZ and PAM4 does not exist as a universal fallback. Verify three things before mixing: that the port is configurable to 100G, that the FEC setting on the port matches the module and the far end, and that the switch firmware release you are running supports the specific optic part number. Vendor compatibility matrices list part numbers, not just speeds, and the difference matters at 200G far more than it did at 100G.&lt;/p&gt;

&lt;p&gt;Keep a written module inventory per rack. In mixed estates, the fastest way to lose an afternoon is a 200G link that comes up at 100G because one end was populated from the wrong spares bin.&lt;/p&gt;

&lt;h2&gt;
  
  
  Cost: Where the Money Actually Goes
&lt;/h2&gt;

&lt;p&gt;Comparing sticker prices per port is the least useful way to compare these two generations, because the cost is distributed across five line items that move in opposite directions. Per port, a 200G QSFP56 port costs more than a 100G QSFP28 port. Per gigabit of switching capacity, it is usually cheaper, because you get twice the bandwidth from the same chassis, the same rack unit and the same port count. Optics follow the same pattern: a 200G module costs more than a 100G module, but you buy half as many for the same aggregate bandwidth. Cabling is roughly neutral per link and slightly cheaper per gigabit at 200G, since one 200G DAC replaces two 100G DACs in a two-rail design. The items that reliably go up are the host side — PCIe 4.0 servers and QSFP56 adapters — and the fiber remediation you may need before PAM4 will run clean.&lt;/p&gt;

&lt;p&gt;There is also a sparing cost that rarely gets modelled. Every distinct optic type in the estate needs spares, and a mixed 100G/200G estate carries two inventories instead of one. Projects that skip this line item tend to discover it during the first failure, at the worst possible moment, when the only spare on the shelf is the wrong generation.&lt;/p&gt;

&lt;h2&gt;
  
  
  When 200G Is Worth It, and When It Is Not
&lt;/h2&gt;

&lt;p&gt;Move to QSFP56 when at least two of these are true: the workload is east-west and bandwidth-bound (distributed training, storage replication, NVMe-oF, MPI across nodes); the host fleet already has PCIe 4.0 x16 slots free; the fiber runs are inside the row and under 100 m; and the switch is a spine or an AI-leaf where port count is the binding constraint. Stay on QSFP28 when the rack is 25G-dense and depends on 4 x 25G breakout, when the hosts are PCIe 3.0, when you have long singlemode runs beyond 2 km, or when the traffic profile is north-south and nowhere near saturating 100G.&lt;/p&gt;

&lt;p&gt;For most estates the correct answer is not one or the other but a deliberate two-tier design: 100G QSFP28 where the hosts live, 200G QSFP56 where the fabric aggregates. That split keeps the breakout flexibility where it is needed and takes the density benefit where it actually pays.&lt;/p&gt;

&lt;h2&gt;
  
  
  Frequently Asked Questions
&lt;/h2&gt;

&lt;h3&gt;
  
  
  Can I plug a QSFP28 module into a QSFP56 port?
&lt;/h3&gt;

&lt;p&gt;Physically yes — the cage and connector are identical. Electrically it depends on the platform: most QSFP56 ports can be configured to 100G and will host a QSFP28 module, but the port speed and FEC mode usually have to be set explicitly rather than negotiated. Confirm the combination in the switch vendor's compatibility matrix for your firmware release before standardising on it.&lt;/p&gt;

&lt;h3&gt;
  
  
  Is QSFP56 backward compatible with my existing 100G DAC cables?
&lt;/h3&gt;

&lt;p&gt;Not for 200G operation. Passive copper for 200G QSFP56 is specified to about 2.5 m in the NVIDIA LinkX family, while 100G QSFP28 passive copper reaches 5 m — so a 3 m 100G DAC that has worked for years is outside the qualified passive range at 200G. Those cables remain usable if the port is run at 100G, which is why many teams keep a 100G phase in the migration.&lt;/p&gt;

&lt;h3&gt;
  
  
  Why does a 200G link need FEC when my 100G links do not?
&lt;/h3&gt;

&lt;p&gt;Because 50G per lane PAM4 signalling has a much smaller margin than 25G NRZ. Forward error correction at the physical layer compensates for the reduced eye height introduced by four-level signalling, so it is effectively part of the standard rather than an option. Both ends must agree on the FEC mode; a mismatch is a common cause of a port that shows light but never comes up.&lt;/p&gt;

&lt;h3&gt;
  
  
  Does 200G double my fiber count?
&lt;/h3&gt;

&lt;p&gt;No. QSFP56 uses the same four lanes, in the same duplex pairs, as QSFP28 — the bandwidth doubling comes from modulation, not from additional strands. What changes is the quality requirement: PAM4 is less tolerant of insertion loss, contamination and poor terminations, so an existing multimode plant may need cleaning or re-termination before it will carry 200G reliably.&lt;/p&gt;

&lt;h3&gt;
  
  
  Should I buy a 200G switch if my servers are still on 100G?
&lt;/h3&gt;

&lt;p&gt;Only if you have a funded host refresh inside your planning horizon. A QSFP56 chassis will happily run 100G today and gives you a forward path without a hardware swap, but you pay for higher power draw and a second optics generation in the meantime. If the host fleet is staying on PCIe 3.0 for the next two years, a dense 100G platform usually delivers more usable bandwidth per unit of spend.&lt;/p&gt;

&lt;p&gt;If you are sizing a 100G-to-200G step and want a second opinion before committing, the useful starting point is a ports list, a host inventory with PCIe generations, and your longest cable run. Browse switch platforms, 100G and 200G adapters, and server hardware availability on &lt;a href="https://zhongxindata.hk" rel="noopener noreferrer"&gt;zhongxindata.hk&lt;/a&gt;, or send that list on WhatsApp to &lt;strong&gt;+86 13068616914&lt;/strong&gt; and we will map it against current stock and flag the compatibility traps before you order optics.&lt;/p&gt;

&lt;p&gt;Published by Zhongxin Data | November 2026&lt;/p&gt;

</description>
      <category>networking</category>
      <category>devops</category>
      <category>hardware</category>
      <category>ai</category>
    </item>
    <item>
      <title>Upgrading a ToR Layer from 10G/25G to 100G: A Practical Planning Guide</title>
      <dc:creator>梁育水</dc:creator>
      <pubDate>Wed, 18 Nov 2026 02:40:20 +0000</pubDate>
      <link>https://dev.to/shayne914816/upgrading-a-tor-layer-from-10g25g-to-100g-a-practical-planning-guide-28o0</link>
      <guid>https://dev.to/shayne914816/upgrading-a-tor-layer-from-10g25g-to-100g-a-practical-planning-guide-28o0</guid>
      <description>&lt;p&gt;Most ToR upgrades that go wrong fail before a single cable moves: the team picked a switch generation that does not match their optics plan, forgot that a 100G QSFP28 port breaks out only four ways, or ordered servers whose PCIe slots cannot feed a dual-port 100G adapter. The conclusion up front: a 25G-to-100G ToR refresh is a straightforward project if you plan three things in order — silicon generation first, breakout and oversubscription math second, optics and server NICs third — and if you rehearse the cutover on paper before touching production. This guide walks through each decision in the sequence an infrastructure engineer would actually face it.&lt;/p&gt;

&lt;h2&gt;
  
  
  What a ToR Upgrade Actually Changes
&lt;/h2&gt;

&lt;p&gt;Top-of-rack switches sit at the edge of your failure domain. When you move a ToR layer from 25G to 100G, every downstream dependency changes with it: the cabling plan, the optics inventory, the server adapters, the leaf-spine uplink bandwidth, and in many cases the power and airflow profile of the rack itself. The bandwidth per server connection goes up 4x (or 8x if you are coming from 10G), but the number of physical ports usually goes down, because a single 100G QSFP28 port replaces four SFP28 lanes. That consolidation is where most of the planning value lives: fewer ports per rack means you need a clear policy for which hosts get dedicated 100G and which ones stay on 25G via breakout.&lt;/p&gt;

&lt;p&gt;Treat the upgrade as a redesign of the rack edge, not a like-for-kind swap. Teams that simply replace a 48-port 25G switch with a 32-port 100G switch often discover mid-migration that they cannot fit the old host count at native speed and end up paying for breakout cables they never budgeted.&lt;/p&gt;

&lt;h2&gt;
  
  
  Baseline Inventory: What to Audit Before You Order Anything
&lt;/h2&gt;

&lt;p&gt;Before comparing switch models, produce a written inventory of what already exists. The minimum list:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Current ToR models, firmware versions, and installed optics (type, speed, and vendor of every transceiver).&lt;/li&gt;
&lt;li&gt;Host-side NIC inventory: which servers have 25G-capable adapters, which are still on 10G, and which have free PCIe slots able to sustain a 100G card.&lt;/li&gt;
&lt;li&gt;Cabling: length ranges, whether runs are DAC (twinax), AOC, or optical transceivers with fiber, because reach rules change at 100G.&lt;/li&gt;
&lt;li&gt;Uplink utilization on the existing leaf-spine fabric, measured at peak, not averaged.&lt;/li&gt;
&lt;li&gt;Power headroom per rack and whether your new switch's airflow direction matches the cold/hot aisle layout.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;This inventory determines your constraints. A shop where 80% of hosts are on 10G and idle will plan a phased rollout; a shop with GPU nodes saturating 25G links will prioritize native 100G for those hosts on day one. The switch shopping comes after, not before.&lt;/p&gt;

&lt;h2&gt;
  
  
  Switch Silicon Generations: Matching the Platform to Your Roadmap
&lt;/h2&gt;

&lt;p&gt;Mellanox/NVIDIA Spectrum switches are the most common ToR choice for 100G estates, and the generations matter because they define your future upgrade path. Spectrum-1 is a proven, widely deployed 100G platform; Spectrum-2 adds deeper buffers and richer telemetry for larger fabrics; Spectrum-3 carries the estate to 200G per port; Spectrum-4 moves to 800G with OSFP. The table below maps the common models you will encounter to their positions.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Model&lt;/th&gt;
&lt;th&gt;Silicon&lt;/th&gt;
&lt;th&gt;Ports&lt;/th&gt;
&lt;th&gt;Speed per Port&lt;/th&gt;
&lt;th&gt;Form Factor&lt;/th&gt;
&lt;th&gt;Typical Role&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;SN2700&lt;/td&gt;
&lt;td&gt;Spectrum-1&lt;/td&gt;
&lt;td&gt;32 x QSFP28&lt;/td&gt;
&lt;td&gt;100G (breakout 4x25G)&lt;/td&gt;
&lt;td&gt;1U&lt;/td&gt;
&lt;td&gt;Cost-effective ToR leaf for 25/100G estates&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;SN3800&lt;/td&gt;
&lt;td&gt;Spectrum-2&lt;/td&gt;
&lt;td&gt;64 x QSFP28&lt;/td&gt;
&lt;td&gt;100G&lt;/td&gt;
&lt;td&gt;2U&lt;/td&gt;
&lt;td&gt;High-density 100G leaf when port count outranks speed&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;SN4600&lt;/td&gt;
&lt;td&gt;Spectrum-3&lt;/td&gt;
&lt;td&gt;64 x QSFP56&lt;/td&gt;
&lt;td&gt;200G&lt;/td&gt;
&lt;td&gt;2U&lt;/td&gt;
&lt;td&gt;200G-ready spine or dense AI/storage leaf&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;SN5600&lt;/td&gt;
&lt;td&gt;Spectrum-4&lt;/td&gt;
&lt;td&gt;64 x OSFP&lt;/td&gt;
&lt;td&gt;800G&lt;/td&gt;
&lt;td&gt;2U&lt;/td&gt;
&lt;td&gt;AI cluster fabric and 800G spine layers&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The practical rule: if your roadmap stops at 100G for the next several years, a Spectrum-1 platform such as the &lt;a href="https://zhongxindata.hk/products/100gbe-qsfp28-ethernet-switch-used-for-mellanox-sn2700-msn27/" rel="noopener noreferrer"&gt;Mellanox SN2700 100GbE QSFP28 switch&lt;/a&gt; is the workhorse choice and the most cost-efficient per port. If you want headroom and denser 100G port counts in the same rack unit budget, the &lt;a href="https://zhongxindata.hk/products/mellanox-sn3800-cs2ro-100gbe-64-port-qsfp28-switch-2u-used/" rel="noopener noreferrer"&gt;SN3800 64-port 100GbE switch&lt;/a&gt; doubles the port count in 2U. If 200G or AI-scale east-west traffic is on the roadmap within 24 months, evaluate Spectrum-3 class hardware now so you do not rebuy optics twice.&lt;/p&gt;

&lt;h2&gt;
  
  
  Breakout Planning: 4x25G, Oversubscription, and Port Budgets
&lt;/h2&gt;

&lt;p&gt;A 100G QSFP28 port can break out to four 25G SFP28 lanes. That single fact drives most ToR layouts. Work through the math per rack:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Count hosts that genuinely need native 100G today: storage heads, hypervisor uplinks, GPU nodes, backup gateways.&lt;/li&gt;
&lt;li&gt;Break out remaining 100G ports to 4x25G for the general host population, usually with passive breakout DACs for in-rack runs.&lt;/li&gt;
&lt;li&gt;Reserve uplinks: a typical layout allocates 4-8 ports of the 32 to leaf-spine uplinks, and those uplinks should be native 100G, not breakout.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Oversubscription is the number most teams skip. If 12 hosts at 25G each share one 100G uplink, that is a 3:1 oversubscription ratio toward the spine — acceptable for general workloads, painful for storage replication or backup windows. Write the ratio down per rack before you finalize the port budget, because it tells you how many spine uplinks the new design truly needs.&lt;/p&gt;

&lt;h2&gt;
  
  
  Optics: DAC, AOC, and Fiber Reach Rules at 100G
&lt;/h2&gt;

&lt;p&gt;At 25G and below, teams lean heavily on passive DACs and rarely think about optics. At 100G the physics tighten: passive twinax reach drops, and the choices multiply. The working rules of thumb:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Same-rack and adjacent-rack (roughly under 3 meters): passive DAC remains the cheapest and most reliable option, including breakout DACs for 4x25G.&lt;/li&gt;
&lt;li&gt;Across a few racks (3-30 meters): AOCs give you lighter cables and better bend tolerance than twinax.&lt;/li&gt;
&lt;li&gt;Beyond about 30 meters, or between racks with conduit paths: optical transceivers with OM4/MMF or single-mode fiber become the standard.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Two cautions. First, mixing transceiver vendors across a link (transmitter from one brand, receiver from another) is a common source of link-flap tickets; keep each hop consistent. Second, breakout ports sometimes require configuration changes to split the port into 4 lanes; do this deliberately and label both ends, because a mislabeled breakout is the classic root cause of "the link was fine until we recabled."&lt;/p&gt;

&lt;h2&gt;
  
  
  Server-Side Readiness: NICs, PCIe Lanes, and Adapter Selection
&lt;/h2&gt;

&lt;p&gt;The switch side is half the project; the host side is where budgets silently double. A server needs a PCIe slot with enough lanes to sustain 100G. As a rule of thumb, a PCIe 3.0 x16 slot can carry a single-port 100G adapter; dual-port 100G cards generally want PCIe 4.0-class slots, or at minimum a x16 PCIe 3.0 slot with acceptance that both ports cannot simultaneously saturate.&lt;/p&gt;

&lt;p&gt;For single-port 100G over QSFP28, the Mellanox ConnectX-5 family is the standard fit — for example the &lt;a href="https://zhongxindata.hk/products/mellanox-mcx4621a-acab-100gbe-adapter/" rel="noopener noreferrer"&gt;ConnectX-5 MCX4621A-ACAT 100GbE adapter&lt;/a&gt; — and it pairs cleanly with Spectrum-1 ToR hardware. For dual-port 100G, ConnectX-5 variants such as the &lt;a href="https://zhongxindata.hk/products/100gbe-dual-port-network-card-used-for-mellanox-connectx-5-m/" rel="noopener noreferrer"&gt;dual-port 100GbE ConnectX-5 network card&lt;/a&gt; cover top-of-rack hosts that need two independent fabric legs. If you plan toward 200G or InfiniBand-adjacent workloads, the ConnectX-6 generation on PCIe 4.0 is the better landing spot; the &lt;a href="https://zhongxindata.hk/products/100gbe-infiniband-hdr100-dual-port-network-card-used-for-mellanox-cx6-mcx653106a-ecat-qsfp56-pcie-4-0-nic/" rel="noopener noreferrer"&gt;dual-port HDR100 ConnectX-6 PCIe 4.0 adapter&lt;/a&gt; keeps Ethernet and HDR InfiniBand flexibility on the same card class.&lt;/p&gt;

&lt;p&gt;Audit free slots before you buy adapters: a chassis that looks roomy in the spec sheet often has its x16 slots occupied by GPU or HBA cards. This is also the right moment to standardize NIC firmware and driver versions across the fleet, because mixed firmware is the second most common cause of flaky 100G links after optics mismatch.&lt;/p&gt;

&lt;h2&gt;
  
  
  Cabling, Airflow, and Rack Physical Constraints
&lt;/h2&gt;

&lt;p&gt;100G twinax cables are thicker and stiffer than their 25G counterparts, and cable management failures show up as bent connectors and marginal links rather than clean outages. Plan for it:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Verify switch airflow direction matches the rack's cold/hot aisle; mixing airflow directions in one rack creates hot spots that throttle hardware long before software does.&lt;/li&gt;
&lt;li&gt;Order cable lengths conservatively; a DAC that is 20 cm too short cannot be "stretched," and re-ordering mid-migration stalls the whole cut.&lt;/li&gt;
&lt;li&gt;Label both ends of every breakout cable with the parent port ID, since a 4x25G breakout looks identical to a native 25G DAC at a glance.&lt;/li&gt;
&lt;li&gt;Check power supply count and circuit capacity on the ToR; dual-powered switches in a rack that assumed single-supply draw can trip breakers during migration.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Migration Sequencing: Cutting Over Without an Outage
&lt;/h2&gt;

&lt;p&gt;The safest migration pattern for a multi-rack estate is parallel-run per rack, not big-bang. The sequence that works in practice:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Stage the new 100G ToR in the same rack, powered but not cabled to production uplinks.&lt;/li&gt;
&lt;li&gt;Move or dual-home the highest-value hosts first: storage, virtualization management, GPU nodes. Dual-homing — one leg on the old 25G fabric, one leg on the new 100G fabric — gives you an instant rollback path.&lt;/li&gt;
&lt;li&gt;Bring up new fabric uplinks to the spine on dedicated ports, and verify routes and LLDP adjacency before shifting any host traffic.&lt;/li&gt;
&lt;li&gt;Migrate general hosts in batches, one rack or one cabinet at a time, keeping a change window with a named rollback owner.&lt;/li&gt;
&lt;li&gt;Retire the old ToR only after the last breakout-dependent host has moved, not before; breakout ports are almost always the long tail.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Teams with a maintenance window and no dual-homing option can still do this safely — the difference is that each batch's rollback plan becomes "recable to the old switch," so keep the old hardware powered and racked until sign-off.&lt;/p&gt;

&lt;h2&gt;
  
  
  Validation Checklist Before You Sign Off
&lt;/h2&gt;

&lt;p&gt;Do not accept a migrated rack on "the link is up." The minimum validation set:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Link-level: every port negotiated at the intended speed, FEC (forward error correction) mode consistent per link class, no retrain counters climbing.&lt;/li&gt;
&lt;li&gt;Throughput: run iperf3 or an equivalent between a sample host and a known-good peer, both directions, and record the number. A 100G link that tops out at 40G usually means a lane or FEC issue on one hop.&lt;/li&gt;
&lt;li&gt;Latency-sensitive paths: storage and hypervisor heartbeat networks get an RTT baseline comparison before and after.&lt;/li&gt;
&lt;li&gt;Redundancy: pull one uplink per rack during the window and confirm failover behavior matches the design, not hope.&lt;/li&gt;
&lt;li&gt;Telemetry: confirm the new switches export counters and events to whatever monitoring you run, because a fabric you cannot see is a fabric you cannot troubleshoot.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Common Failure Modes in First-Generation Upgrades
&lt;/h2&gt;

&lt;p&gt;Four patterns account for most first-attempt pain. One, optics mismatch across a hop — keep transmitter and receiver consistent and buy spares of the same type. Two, forgotten breakout configuration — a port that worked as native 100G in the lab needs an explicit split command to serve 4x25G in production. Three, PCIe slot starvation discovered after adapters arrive — audit chassis interiors first. Four, spine uplink undersizing — hosts get 4x bandwidth while uplinks stay the same, and the fabric becomes congested in ways that look like application problems. Every one of these is caught by the audit and validation steps above, which is the real argument for doing them.&lt;/p&gt;

&lt;h2&gt;
  
  
  Frequently Asked Questions
&lt;/h2&gt;

&lt;h3&gt;
  
  
  Can I mix native 100G ports and 4x25G breakout ports on the same ToR switch?
&lt;/h3&gt;

&lt;p&gt;Yes. Spectrum-based ToR switches are designed for mixed-speed operation, and this is the standard layout: a few native 100G ports for uplinks and high-demand hosts, with breakout DACs serving 25G hosts on the remaining ports. Configure breakout ports explicitly and label both cable ends, since a breakout port appears as a single parent port to the switch OS.&lt;/p&gt;

&lt;h3&gt;
  
  
  Do I need to replace all server NICs when moving the ToR to 100G?
&lt;/h3&gt;

&lt;p&gt;No, and you usually should not. Keep 25G or 10G hosts on breakout lanes and upgrade adapters only for hosts whose workloads demand native 100G. A ConnectX-5 class adapter provides single- or dual-port 100G for those hosts; the rest of the fleet can migrate to native speed later as part of routine hardware refresh.&lt;/p&gt;

&lt;h3&gt;
  
  
  How far can a passive DAC run at 100G?
&lt;/h3&gt;

&lt;p&gt;Passive twinax at 100G is reliable at short reach, typically up to around 3 meters depending on cable gauge and platform support. For runs beyond that, move to AOCs (roughly 3 to 30 meters) or optical transceivers with fiber. When reach is uncertain, choose the AOC or optical option; marginal DAC runs produce intermittent link retrains that are expensive to diagnose.&lt;/p&gt;

&lt;h3&gt;
  
  
  Is Spectrum-1 hardware still a sensible choice for a new 100G ToR layer?
&lt;/h3&gt;

&lt;p&gt;For estates whose roadmap stays at 100G, Spectrum-1 platforms such as the SN2700 remain the most cost-efficient per-port choice and are deployed at scale across the industry. Choose Spectrum-2 or 3 class hardware when you need higher port density, deeper buffering, or a 200G upgrade path within your planning horizon. Match the silicon to the roadmap, not to the newest catalog page.&lt;/p&gt;

&lt;h3&gt;
  
  
  What oversubscription ratio is acceptable for a 100G ToR design?
&lt;/h3&gt;

&lt;p&gt;It depends on workload. General purpose compute tolerates 3:1 or 4:1 toward the spine without users noticing; storage replication, backup bursts, and east-west AI traffic want 1:1 or close to it. Compute the ratio per rack from measured peak traffic, not vendor defaults, and budget spine uplink ports accordingly — this single number determines whether the fabric feels fast after the upgrade.&lt;/p&gt;

&lt;p&gt;If you are scoping a 25G-to-100G refresh and want a second opinion on switch generation, breakout layout, or adapter matching, the fastest path is a short conversation with a parts list attached. Browse current 100G ToR switches, adapters, and server hardware availability on &lt;a href="https://zhongxindata.hk" rel="noopener noreferrer"&gt;zhongxindata.hk&lt;/a&gt;, or message us on WhatsApp at &lt;strong&gt;+86 13068616914&lt;/strong&gt; with your rack count and host list, and we will help you pressure-test the plan before you spend on optics.&lt;/p&gt;

&lt;p&gt;Published by Zhongxin Data | November 2026&lt;/p&gt;

</description>
      <category>networking</category>
      <category>devops</category>
      <category>hardware</category>
      <category>ai</category>
    </item>
    <item>
      <title>Mellanox SN2700 vs SN3700: Choosing the Right Spectrum Switch for 100G and 200G Fabrics</title>
      <dc:creator>梁育水</dc:creator>
      <pubDate>Tue, 17 Nov 2026 08:01:44 +0000</pubDate>
      <link>https://dev.to/shayne914816/mellanox-sn2700-vs-sn3700-choosing-the-right-spectrum-switch-for-100g-and-200g-fabrics-2ilb</link>
      <guid>https://dev.to/shayne914816/mellanox-sn2700-vs-sn3700-choosing-the-right-spectrum-switch-for-100g-and-200g-fabrics-2ilb</guid>
      <description>&lt;p&gt;Both switches are 1U, both present 32 ports in the front panel, and both are sold under names that look interchangeable. They are not. The &lt;strong&gt;Mellanox SN2700&lt;/strong&gt; is a Spectrum-1 platform with 32 QSFP28 cages and a hard ceiling of 100GbE per port. The &lt;strong&gt;Mellanox SN3700&lt;/strong&gt; is a Spectrum-2 platform with 32 QSFP56 cages and 200GbE per port. If every link in your fabric terminates at 100G, the SN2700 does the same job for less power and less money. If you need 200G uplinks, a 42MB buffer, or a platform that still has a support life ahead of it, the SN3700 is the correct answer. Everything below is the detail behind that split.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the SN2700 Actually Is
&lt;/h2&gt;

&lt;p&gt;The SN2700 is built on the Spectrum-1 ASIC and exposes 32 QSFP28 ports running 10, 25, 40, 50 or 100GbE. Switching capacity is 6.4Tb/s bidirectional, forwarding rate 4.76Bpps, and port-to-port latency is quoted at 300ns in cut-through mode. The packet buffer is 16MB, fully shared rather than carved into per-port slices — which is the reason this switch absorbs many-to-one incast and storage microbursts better than segmented-buffer boxes in the same class. Control plane is an Intel x86 dual-core at 1.40GHz with 8GB of DDR3, four hot-swap fans in an N+1 arrangement, and two hot-swap power supplies. Typical draw is around 130W with passive copper, up to roughly 180W loaded with optics. Chassis depth is 686mm.&lt;/p&gt;

&lt;p&gt;One practical consequence of the design: the SN2700 family is end-of-life as a product line. That does not make it a bad switch — it makes it a stable, well-documented switch with no remaining roadmap. Firmware still installs, the feature set is frozen and predictable, and the secondary market is deep. If your fabric already runs SN2700 units, adding more of the same model is usually cheaper and less risky than introducing a second silicon generation into the same leaf layer. A current unit is listed &lt;a href="https://zhongxindata.hk/products/mellanox-sn2700-cs2ros-100gbe-32-port-qsfp28-switch-used/" rel="noopener noreferrer"&gt;here for the SN2700 32-port QSFP28 build&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the SN3700 Actually Is
&lt;/h2&gt;

&lt;p&gt;The SN3700 moves to Spectrum-2 and QSFP56 cages, which is what unlocks 200GbE per port. Port speeds span 10, 25, 40, 50, 100 and 200GbE. Throughput stays at 6.4Tb/s on the 200G model because 32 ports at 200G is the same aggregate as 32 ports at 100G doubled — the difference is where the bandwidth sits, not how much of it exists. Forwarding rate rises to 8.33Bpps. Latency is 425ns, higher than the SN2700's 300ns because the pipeline is deeper and the feature set is larger. The packet buffer is 42MB, more than two and a half times the SN2700's, and the CPU is a quad-core x86 at 2.20GHz.&lt;/p&gt;

&lt;p&gt;Two details matter more than the headline numbers. First, chassis depth drops to 559mm — 127mm shallower than the SN2700, which is the difference between fitting and not fitting in a shallow rack or a two-post telecom frame. Second, global power consumption rises to about 250W, nearly double. In a 40-rack deployment that power delta becomes a real number on the PDU budget, and it is the single most under-estimated cost when teams move from 100G to 200G switching. A 200GbE build with the reverse-airflow option is listed &lt;a href="https://zhongxindata.hk/products/mellanox-sn3700-vs2r-200gbe-32-port-qsfp56-switch-used/" rel="noopener noreferrer"&gt;as SN3700-VS2R 32-port QSFP56&lt;/a&gt;, and the 100GbE variant sits &lt;a href="https://zhongxindata.hk/products/mellanox-sn3700-cs2f-100gbe-32-port-qsfp28-switch-used/" rel="noopener noreferrer"&gt;here as SN3700-CS2F 32-port QSFP28&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  SN2700 vs SN3700 Side by Side
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Specification&lt;/th&gt;
&lt;th&gt;SN2700&lt;/th&gt;
&lt;th&gt;SN3700 (200GbE)&lt;/th&gt;
&lt;th&gt;SN3700C (100GbE)&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;ASIC&lt;/td&gt;
&lt;td&gt;Spectrum-1&lt;/td&gt;
&lt;td&gt;Spectrum-2&lt;/td&gt;
&lt;td&gt;Spectrum-2&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Connector&lt;/td&gt;
&lt;td&gt;32 × QSFP28&lt;/td&gt;
&lt;td&gt;32 × QSFP56&lt;/td&gt;
&lt;td&gt;32 × QSFP28&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Max speed per port&lt;/td&gt;
&lt;td&gt;100GbE&lt;/td&gt;
&lt;td&gt;200GbE&lt;/td&gt;
&lt;td&gt;100GbE&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Throughput&lt;/td&gt;
&lt;td&gt;6.4Tb/s&lt;/td&gt;
&lt;td&gt;6.4Tb/s&lt;/td&gt;
&lt;td&gt;3.2Tb/s&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Forwarding rate&lt;/td&gt;
&lt;td&gt;4.76Bpps&lt;/td&gt;
&lt;td&gt;8.33Bpps&lt;/td&gt;
&lt;td&gt;4.76Bpps&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Latency (cut-through)&lt;/td&gt;
&lt;td&gt;300ns&lt;/td&gt;
&lt;td&gt;425ns&lt;/td&gt;
&lt;td&gt;425ns&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Packet buffer&lt;/td&gt;
&lt;td&gt;16MB shared&lt;/td&gt;
&lt;td&gt;42MB shared&lt;/td&gt;
&lt;td&gt;42MB shared&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Chassis depth&lt;/td&gt;
&lt;td&gt;686mm&lt;/td&gt;
&lt;td&gt;559mm&lt;/td&gt;
&lt;td&gt;559mm&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Typical power&lt;/td&gt;
&lt;td&gt;~130W&lt;/td&gt;
&lt;td&gt;~250W&lt;/td&gt;
&lt;td&gt;~242W&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Fans&lt;/td&gt;
&lt;td&gt;4, N+1&lt;/td&gt;
&lt;td&gt;6, N+1&lt;/td&gt;
&lt;td&gt;4, N+1&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Product status&lt;/td&gt;
&lt;td&gt;End of life&lt;/td&gt;
&lt;td&gt;In lifecycle&lt;/td&gt;
&lt;td&gt;In lifecycle&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;h2&gt;
  
  
  SN3700C Is Not an SN3700 — Read the Letter Before the Number
&lt;/h2&gt;

&lt;p&gt;This is where most sourcing mistakes happen. The SN3700C is a Spectrum-2 switch with 32 QSFP28 cages, 3.2Tb/s throughput and a 100GbE ceiling. It is a different product with the same family name. It is not a 200G switch that has been limited by firmware, and you cannot unlock 200G on it later.&lt;/p&gt;

&lt;p&gt;NVIDIA's ordering codes carry the speed class in the letter before the &lt;em&gt;S&lt;/em&gt;. In &lt;code&gt;MSN2700-CS2F&lt;/code&gt;, the &lt;em&gt;C&lt;/em&gt; means 100GbE. In &lt;code&gt;MSN3700-VS2R&lt;/code&gt;, the &lt;em&gt;V&lt;/em&gt; means 200GbE. A &lt;em&gt;B&lt;/em&gt; in that position is the 40GbE build, and a &lt;em&gt;W&lt;/em&gt; is 400GbE on later generations. The trailing letter is airflow, not performance: &lt;em&gt;F&lt;/em&gt; for front-to-back, &lt;em&gt;R&lt;/em&gt; for reverse. So &lt;code&gt;SN3700-CS2F&lt;/code&gt; and &lt;code&gt;SN3700-VS2R&lt;/code&gt; are not near-equivalents — one is a 100G box with forward airflow, the other is a 200G box with reverse airflow. Two independent variables in one string.&lt;/p&gt;

&lt;h2&gt;
  
  
  Airflow Decides Whether the Switch Belongs in Your Rack
&lt;/h2&gt;

&lt;p&gt;Airflow is set by the fan and power-supply modules, not by the chassis, and it cannot be changed in software. A power-to-connector (P2C, forward) unit pulls cold air in at the PSU side and exhausts at the port side. A connector-to-power (C2P, reverse) unit does the opposite. Put the wrong direction in a hot aisle and you recirculate exhaust into the intake, which shows up months later as thermal throttling and fan failures rather than as an obvious fault.&lt;/p&gt;

&lt;p&gt;Both platforms offer both directions, which is why the same model appears under several part numbers. When you specify a switch by model name only, you are leaving the airflow to whoever picks the unit. Specify the full OPN, or state the direction explicitly and ask for the fan and PSU colour coding to be confirmed against it. This single check eliminates the most common reason a delivered switch has to be returned.&lt;/p&gt;

&lt;h2&gt;
  
  
  Breakout, Cabling and the 200G Transition
&lt;/h2&gt;

&lt;p&gt;On the SN2700, each QSFP28 port can be split into two 50GbE or four 25GbE links, which is how a 32-port switch serves a full rack of 25G servers without a second tier. On the SN3700, QSFP56 ports split to 2×100GbE as well as the lower rates, so a single 200G port can feed two 100G leaf-facing links. Passive copper DAC reaches roughly three metres at these rates; beyond that you move to active optical cables or optics, and the optics bill is frequently larger than the switch bill in a 200G refresh.&lt;/p&gt;

&lt;p&gt;400GbE is not available on either platform. Teams planning a 400G spine should be looking at Spectrum-3 hardware, and should treat an SN3700 purchase as a 100G/200G decision rather than a stepping stone to 400G. Where the SN3700 genuinely earns its place is as a 200G spine over 100G leaves: 32 ports of 200G gives 16 leaf connections with no oversubscription, using &lt;a href="https://zhongxindata.hk/products/100gbe-dual-port-network-card-used-for-mellanox-cx5-mcx516a/" rel="noopener noreferrer"&gt;dual-port 100GbE ConnectX-5 adapters&lt;/a&gt; at the server end and QSFP56 to 2×QSFP28 fan-out at the spine.&lt;/p&gt;

&lt;h2&gt;
  
  
  Network OS: The Decision That Outlives the Hardware
&lt;/h2&gt;

&lt;p&gt;Both switches run ONIE, which means the operating system is a separate decision from the box. NVIDIA Onyx (MLNX-OS) gives you the vendor CLI and the vendor feature track. Cumulus Linux gives you a Linux-based stack with standard routing daemons. SONiC gives you an open, container-based stack that many large operators standardise on. The hardware does not care; your team's operational muscle memory does.&lt;/p&gt;

&lt;p&gt;Two things to confirm before committing to a batch. First, which NOS the units were originally ordered with — the OPN records it, and an ONIE unit with no licence arrives with no OS at all. Second, whether the software subscription attached to those units is transferable. Vendor NOS subscriptions are frequently tied to the original purchaser and do not follow the hardware, which turns a cheap switch into an expensive one after the first support renewal. Ask the question before the invoice, not after.&lt;/p&gt;

&lt;h2&gt;
  
  
  How to Decide in Four Questions
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Is any link in this fabric faster than 100G?&lt;/strong&gt; If no, the SN2700 covers it at lower power and lower cost. If yes, the SN3700 is the only option of the two.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;How deep is the rack?&lt;/strong&gt; Below 600mm usable depth, the SN2700's 686mm chassis will not close the door. The SN3700's 559mm will.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;What does the buffer have to absorb?&lt;/strong&gt; Storage backends, NVMe-over-Fabrics and AI east-west traffic produce incast that a 16MB buffer handles less gracefully than a 42MB one. Pure web-tier traffic does not care.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;How long does this fabric need to stay in service?&lt;/strong&gt; A five-year horizon argues for the platform still in lifecycle. A two-year lab or a spare-parts pool argues for matching what is already installed.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What to Verify Before You Commit
&lt;/h2&gt;

&lt;p&gt;Ask for five facts about the exact unit, not the model: the full OPN, the airflow direction, the installed NOS and firmware level, the port count actually passing traffic, and the error counters on the uplinks. A switch that links but accumulates CRC errors on two ports is a switch with a bad cage or a bad cable, and it will pass a power-on test. Front-panel photos of the specific serial number, plus a console capture of the firmware version, settle most of this in one exchange.&lt;/p&gt;

&lt;p&gt;For fabrics that need more than 32 ports at the spine, the same silicon family continues in the &lt;a href="https://zhongxindata.hk/products/mellanox-sn3800-cs2ro-100gbe-64-port-qsfp28-switch-2u-used/" rel="noopener noreferrer"&gt;SN3800 64-port 100GbE 2U build&lt;/a&gt; and, at 200G density, the &lt;a href="https://zhongxindata.hk/products/mellanox-sn4600-vs2f-200gbe-64-port-qsfp56-switch-2u-used/" rel="noopener noreferrer"&gt;SN4600 64-port 200GbE 2U build&lt;/a&gt;. Matching generation across leaf and spine keeps the feature set consistent and the spare-parts inventory small.&lt;/p&gt;

&lt;h2&gt;
  
  
  Frequently Asked Questions
&lt;/h2&gt;

&lt;h3&gt;
  
  
  Can an SN2700 run at 200GbE?
&lt;/h3&gt;

&lt;p&gt;No. The QSFP28 cages and the Spectrum-1 ASIC cap the port at 100GbE. No firmware release changes this. 200GbE requires QSFP56 cages and Spectrum-2 silicon, which means the SN3700 or a later generation.&lt;/p&gt;

&lt;h3&gt;
  
  
  Is the SN3700C a 200G switch?
&lt;/h3&gt;

&lt;p&gt;No. The SN3700C is the 100GbE member of the SN3700 family: 32 QSFP28 ports, 3.2Tb/s throughput, 100GbE ceiling. It shares the Spectrum-2 silicon and the 42MB buffer but not the port speed. Check the OPN letter — C for 100G, V for 200G.&lt;/p&gt;

&lt;h3&gt;
  
  
  Can I mix SN2700 and SN3700 in the same fabric?
&lt;/h3&gt;

&lt;p&gt;Yes, at 100GbE. Both run standard Ethernet, VXLAN and RoCE, and they interoperate at 100G over QSFP28. The SN3700 will not negotiate 200G with an SN2700 port, and the two platforms have different latency and buffer characteristics, so keep a given tier homogeneous where you can.&lt;/p&gt;

&lt;h3&gt;
  
  
  Do these switches ship with transceivers or cables?
&lt;/h3&gt;

&lt;p&gt;No. Optics, DAC and AOC are separate line items and are coded to the port type — QSFP28 for the SN2700 and SN3700C, QSFP56 for the 200G SN3700. Confirm the optic or cable part number against the switch port type before ordering; this is the second most common mismatch after airflow.&lt;/p&gt;

&lt;h3&gt;
  
  
  Which one holds its value better?
&lt;/h3&gt;

&lt;p&gt;The SN3700, because it is still in lifecycle and covers both 100G and 200G roles. The SN2700's value is now driven entirely by installed base — buyers who need spares for an existing SN2700 fabric pay for compatibility, not for capability.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Need a specific build?&lt;/strong&gt; Send the model list — full OPN, airflow direction and quantity — and we confirm availability per model at &lt;a href="https://zhongxindata.hk/" rel="noopener noreferrer"&gt;zhongxindata.hk&lt;/a&gt; or on WhatsApp &lt;strong&gt;+86 13068616914&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;Published by Zhongxin Data | November 2026&lt;/p&gt;

</description>
      <category>networking</category>
      <category>devops</category>
      <category>linux</category>
      <category>hardware</category>
    </item>
  </channel>
</rss>
