• Telecom & ICT Equipment Supplier
  • Multi-Vendor Sourcing
  • Worldwide Shipping
YT CIG Telecom & ICT Equipment
Get a Quote

Choosing spine and leaf switches: port counts, ASICs and optics

Updated on 2026-11-19

A 100G spine-leaf fabric needs leaf switches with four or more 40/100G QSFP28 uplinks and spines carrying enough 100G ports to reach every leaf. In the YT CIG catalogue that points to Juniper QFX5110-48S, QFX5120-48Y, QFX5210-64C, QFX5220-128C and QFX10002-60C, plus the Arista 7280CR3K-96.

Eight catalogue models have a documented 40G or 100G port map. Each figure comes from the document named in the last column; “not stated” marks a cell the vendor leaves out.

Model Best role Front-panel ports 100G-capable ports Throughput (documented) Buffer or latency Source
Juniper QFX5110-48S Leaf, 10G servers 48 SFP+ (1/10G) + 4 QSFP28 4 960 Gbps aggregate 550 ns family figure QFX5110 Hardware Guide
Juniper EX4650-48Y Leaf, 25G servers 48 SFP28 (1/10/25G) + 8 QSFP28 8 not stated not stated EX4650 Hardware Guide
Juniper QFX5120-48Y Leaf, 25G servers 48 SFP28 (1/10/25G) + 8 QSFP28 8 4 Tbps bidirectional not stated QFX5120 Hardware Guide
Juniper QFX5120-32C Leaf or small spine 32 QSFP28 32 6.4 Tbps bidirectional not stated QFX5120 Hardware Guide
Juniper QFX5210-64C Spine, 2U 64 QSFP28 + 2 SFP+ 64 not stated not stated QFX5210 Hardware Guide
Juniper QFX5220-128C Spine, 4U 128 QSFP28 + 2 SFP+ 128 12.8 Tbps not stated QFX5220 Hardware Guide
Juniper QFX10002-60C Spine, routing scale 60 QSFP28 60 12 Tbps, 4 Bpps Juniper Q5 ASIC QFX10002 Hardware Guide
Arista DCS-7280CR3K-96 Spine, deep buffer 96 QSFP100 96 (192 x 50G) 9.6 Tbps, 19.2 Tbps FDX, 4 Bpps 16 GB buffer, 3.8 us Arista 7280R3 datasheet

The QFX5110-48S-AFO is the direct route into a 100G mesh: 48 server ports at 1G or 10G, four QSFP28 uplinks at 40G or 100G, and each uplink can break out to four 10G ports with a DAC cable (Juniper QFX5110 Hardware Guide). A 48-server leaf with all four uplinks at 100G gives 400 G of fabric capacity against 480 G of server-facing capacity, a 1.2:1 ratio.

The EX4650-48Y-AFI and QFX5120-48Y-AFI2 move the server side to 25G SFP28 and keep eight 100G QSFP28 ports: 48 x 25G is 1.2 Tbps downlink against 800 G of uplink, a 1.5:1 ratio.

Spines are the same exercise at larger scale. The QFX5210-64C-AFI2 puts 64 QSFP28 sockets and two dedicated SFP+ ports in 2U, and the guide allows 4 x 25G breakout on ports 0 through 31 only, with 2 x 50G breakout across the full port range. The QFX10002-60C trades port count for routing scale — Juniper’s guide credits it with 12 Tbps and 4 Bpps from its Q5 ASICs.

At the top of the catalogue, the DCS-7280CR3K-96-F carries 96 QSFP100 cages, 16 GB of packet buffer and a 3.8 us latency floor in its datasheet.

Juniper’s SKU suffixes encode airflow. The QFX5110 ordering table lists AFO as ports-to-FRUs airflow and AFI as the reverse, and applies both suffixes to the same 48 SFP+ plus 4 QSFP28 port map (Juniper QFX5110 Hardware Guide). Arista and Huawei use their own suffixes, so a purchase order should quote the full SKU beside the port configuration.

Port arithmetic: leaf count, spine count, oversubscription

Let N be the server ports to connect, D the downlink ports per leaf, L the leaf count, S the spine count, U the links between one leaf and one spine, and d and u the downlink and uplink speeds.

L = ceil(N / D)

Uplinks consumed per leaf = S x U

Ports consumed per spine = L x U

Oversubscription = (D x d) / (S x U x u)

The ratio is measured at one leaf, so it describes the whole fabric only when every leaf carries the same port map. A leaf with 48 x 25G downlinks and eight 100G uplinks runs at (48 x 25) / (8 x 100) = 1.5:1. Add a second spine and split the same eight uplinks across two planes, and the ratio stays 1.5:1 while the failure domain halves.

Leaves Spines Links per leaf per spine Ports used per spine Catalogue spine that fits
4 1 4 16 QFX5120-32C (32 ports)
8 2 3 24 QFX5120-32C (32 ports)
16 2 4 64 QFX5210-64C (64 ports, fully consumed)
24 2 4 96 DCS-7280CR3K-96 (96 ports, fully consumed)

The bottom two rows are why port count alone is a poor selection criterion. A 64-port spine meshing 16 leaves at four links each leaves zero ports for WAN routers, storage arrays or a third spine, and the 96-port Arista at 24 leaves has the same problem. Both rows force a 96- or 128-port spine into the bill of materials.

Equal-cost multi-path spreads traffic across the parallel leaf-spine links, and RFC 2992 describes the hash-based algorithm most switches implement. Uniform link counts and uniform speeds across the fabric give that hash the best chance of an even split. A leaf with five 100G links to a spine where its neighbours have four creates a capacity mismatch the hash cannot correct, so keep U identical for every leaf in a plane.

The spine-leaf 100G port planning guide carries three fully worked designs with the same arithmetic, including a 40G leaf case.

Which datasheet numbers separate two switches with the same port count?

Port count is the least useful column on a datasheet. Arista’s 7280CR3-36S carries 4 GB of buffer across 36 ports, the 7280CR3-32P4 carries 8 GB across 36 ports, and the 7280CR3-96 carries 16 GB across 96 ports, all at 100G port speed (Arista 7280R3 datasheet). The ASIC sets those numbers, and only some vendors print its name.

Juniper does print it once. The QFX10002 Hardware Guide states that the switch runs on Q5 ASICs designed by Juniper Networks, and credits those ASICs with the 12 Tbps and 4 Bpps figures on the QFX10002-60C. No other Juniper guide for the QFX5110, 5120, 5210 or 5220 series names a silicon part in the hardware documentation; they describe ports, throughput and channelization instead.

Arista takes the opposite route in the 7280R3 datasheet. It never names an ASIC — it publishes the behaviours the silicon produces: virtual output queues, dynamic deep buffering, 16 GB of packet buffer on the 7280CR3-96 and 7280CR3K-96, 4 Bpps, and latency from 3.8 us. The product page adds 512-way Layer 3 and 256-way Layer 2 multi-pathing.

Raw datasheet quantities become comparable only as ratios.

  • Buffer per port: 16 GB across 96 ports is 170 MB per port if the pool divides evenly, while the 7280CR3-32P4 carries 8 GB across 36 ports, about 227 MB per port. Shared pools never divide evenly under real traffic, so treat either figure as a ceiling.
  • Packets per second: 4 Bpps appears on both the 96-port Arista and the 60-port QFX10002-60C, against 9.6 and 12 Tbps. Bpps depends on the frame size used in the test, and neither document states it, so the figure ranks forwarding engines only when the test condition matches.
  • Latency floor: 550 ns on the QFX5110 family against 3.8 us on the 7280R3. The first number suits a top-of-rack leaf feeding latency-sensitive query traffic; the second buys buffer for storage and AI flows that arrive in bursts.

Both figures are unloaded floors. A deep-buffer switch holds frames longer once a port is congested, so ask for the loaded latency curve when the fabric carries storage replication or checkpoint traffic.

Buffer sizing for incast and storage traffic

A 100G port running at line rate delivers 12.5 GB per second. A 10 ms microburst from eight such ports is 1 GB aimed at one egress, and the 16 GB buffer on a 7280CR3-96 holds at most sixteen of those bursts before the queue overflows (Arista 7280R3 datasheet).

The same datasheet attributes lossless forwarding to virtual output queues plus dynamic deep buffering. VOQ keeps a congested egress from blocking traffic destined to other ports, which is the head-of-line problem that makes a shallow-buffer spine drop innocent flows during incast. Storage replication, backup windows and AI checkpoint traffic all produce incast patterns: many leaves sending to one spine port that leads to an array.

A practical rule from the port arithmetic above: a 1:1 fabric with uniform link counts needs less spine buffer than a 2:1 or 3:1 fabric, because oversubscription converts every burst into a queue.

Every 100G fabric port in the tables above is a QSFP28-class cage, standardised in SFF-8665 as a four-lane pluggable running 28 Gb/s per lane, and the module that goes inside is chosen by reach. The 100G QSFP28 transceiver page groups the catalogue by part type, and the table below maps each class to the standard that defines it.

Optic class Standard or MSA Reach Fibre and connector Lane layout
QSFP28 passive DAC SFF-8665 form factor 1 to 3 m Copper, QSFP28 plug 4 x 25G passive
100GBASE-SR4 IEEE 802.3bm-2015 70 m OM3, 100 m OM4 Multimode, MPO-12 4 x 25.78125 Gb/s
100GBASE-CWDM4 CWDM4 MSA, ITU-T G.694.2 grid 2 km SMF G.652, duplex LC 4 wavelengths, 1271 to 1331 nm
100GBASE-DR IEEE 802.3cd-2018 500 m SMF G.652, duplex LC 1 wavelength, 1310 nm
100GBASE-FR1 IEEE 802.3cu-2021 2 km SMF G.652, duplex LC 1 wavelength, 1310 nm
100GBASE-LR1 IEEE 802.3cu-2021 10 km SMF G.652, duplex LC 1 wavelength, 1310 nm
100GBASE-LR4 IEEE 802.3bm-2015 10 km SMF G.652, duplex LC 4 wavelengths near 1310 nm
100GBASE-ER4 IEEE 802.3bm-2015 40 km SMF G.652, duplex LC 4 wavelengths near 1310 nm
400GBASE-DR4 IEEE 802.3bs-2017 500 m SMF G.652, MPO-12 4 x 100G lanes
400GBASE-FR4 IEEE 802.3cu-2021 2 km SMF G.652, duplex LC 4 wavelengths
400GBASE-LR4-10 IEEE 802.3cu-2021 10 km SMF G.652, duplex LC 4 wavelengths

Inside one rack, a passive DAC removes both transceivers and their power draw. Across a data hall, 100GBASE-SR4 covers 70 m on OM3 and 100 m on OM4 over an MPO-12 connector, with four 25.78 Gb/s lanes per IEEE 802.3bm-2015. Beyond the building, the choice splits by fibre plant: 100GBASE-CWDM4 puts four lanes on the 20 nm CWDM grid that ITU-T G.694.2 defines and reaches 2 km over G.652, while 100GBASE-FR1 reaches the same 2 km on a single 1310 nm wavelength under IEEE 802.3cu-2021.

Those two 2 km options do not interoperate. CWDM4 multiplexes four wavelengths onto one duplex LC pair; FR1 sends one wavelength on one pair. A link with CWDM4 at one end and FR1 at the other has no common channel, even though both datasheets say 2 km and both use LC connectors.

Breakout changes the port budget and the standard at once. A 100G QSFP28 port configured for 4 x 25G feeds four SFP28 ports at 25GBASE-SR or 25GBASE-LR, PHYs defined in IEEE 802.3by-2016, so the leaf side needs SFP28 cages. Channelization depends on the platform: the QFX5210-64C guide allows 4 x 25G breakout on ports 0 through 31 only, and 2 x 50G breakout across all 64 ports (Juniper QFX5210 Hardware Guide).

For 400G, the QSFP-DD cage carries eight 50G PAM4 lanes under the QSFP-DD MSA and accepts QSFP28 modules at 100G. 400GBASE-DR4 runs 500 m over MPO-12, and 400GBASE-FR4 runs 2 km over duplex LC, so the duplex LC plant chosen for CWDM4 or FR1 also carries the first 400G step. The 25G and 100G optical modules guide covers the SFP28 side in more detail.

What breaks when 40G leaves join a 100G spine?

Physically, nothing. A QSFP+ 40G module fits a QSFP28 cage, and every spine in the table accepts 40G on at least some ports. The QFX10002-60C configures each of its 60 ports as 100G, 40G or 4 x 10G (Juniper QFX10002 Hardware Guide); the QFX5210-64C sets 40G or 100G automatically from the inserted transceiver; the QFX5220-128C ports run at 100G or 40G; and the Arista 7280CR3K-96 lists 96 ports usable at 40G with QSFP+ optics in its model comparison table.

What changes is the arithmetic. Swap a leaf’s four 100G uplinks for four 40G uplinks and its fabric capacity drops from 400 G to 160 G, taking a 48 x 10G leaf from 1.2:1 to 3:1 oversubscription. In a 16-leaf, two-spine fabric, each spine carries 64 ports: 6.4 Tbps of mesh at 100G, 2.56 Tbps at 40G. The port count is identical — a spine sized for 16 leaves at 100G fills the same way at 40G and delivers 40 percent of the bandwidth.

Mixed rates also break the uniform-path assumption behind ECMP. A 40G leaf and a 100G leaf on the same spine receive the same number of hash buckets, but one delivers 2.5 times the bandwidth per bucket. RFC 2992’s hash assigns flows without regard to link speed, so the fabric develops hot spots that only a rate-aware scheduler or per-leaf link-count adjustment can offset.

The 40G path makes sense in one case: reusing existing QFX5110-32Q stock where the spine has spare 40G-capable ports and the workload tolerates 3:1. Treat it as a transition and budget the 100G replacement when the 40G optics reach end of support.

Which model fits a pilot, a regional site and a 16-leaf fabric?

Three builds cover most telecom sites, each sized from the formulas above.

Site Servers Leaves Spines Spine ports used Oversubscription
PoP pilot 192 x 10G 4 x QFX5110-48S 1 x QFX5120-32C 16 of 32 1.2:1
Regional OSS/BSS 384 x 25G 8 x QFX5120-48Y 2 x QFX5210-64C 32 of 64 per spine 1.5:1
16-leaf fabric 768 x 25G 16 x QFX5120-48Y 2 x QFX5220-128C 64 of 128 per spine 1.5:1

The pilot works because 192 servers divide into four 48-port leaves, and four leaves at four 100G links each consume 16 of the 32 ports on a QFX5120-32C-AFI. The remaining 16 ports take the WAN router, a storage array and a third spine later. Downlink is 4 x 480 G and uplink is 16 x 100 G, which gives 1.2:1.

At regional scale the leaf moves to 25G and the spine to 64 ports. Eight leaves at eight 100G uplinks each, split four per spine, put 32 ports on each QFX5210-64C-AFI2 and leave 32 free. That headroom is what absorbs the first storage expansion without a second spine purchase.

Sixteen leaves at four links per spine need 64 ports on each spine, which is the point where a 64-port model fills completely. The QFX5220-128C-AFO carries 128 QSFP28 ports in 4U at 12.8 Tbps, so the same design uses half its capacity and leaves room for a third spine or 64 WAN and storage ports. The alternative at this size is the DCS-7280CR3K-96-F with 96 ports and 16 GB of buffer, which trades 32 ports for deeper queues.

Above 24 leaves at four links per spine the arithmetic demands 96 ports on every spine, and above 32 leaves it demands 128.

FAQ

Can a 100G QSFP28 port run 4 x 25G?

Not on every model. The Arista 7280CR3K-96 datasheet lists a maximum of 192 interfaces from its 96 QSFP100 ports, two per port at 25G or 50G, and the QFX5210-64C guide allows 4 x 25G breakout on ports 0 through 31 only, with a system ceiling of 128 interfaces across 64 ports (Arista 7280R3 datasheet; Juniper QFX5210 Hardware Guide). The QFX5220-128C guide does list 4 x 25G and 4 x 10G channelization across its 128 ports. Verify the breakout map before promising 48 SFP28 server ports from twelve 100G uplinks.

Which optic reaches 2 km over duplex LC?

Two classes do, and they use different multiplexing. 100GBASE-CWDM4 carries four wavelengths on the ITU-T G.694.2 CWDM grid over G.652 fibre, and 100GBASE-FR1 carries one 1310 nm wavelength under IEEE 802.3cu-2021. Both terminate in duplex LC and both reach 2 km. A CWDM4 module at one end and an FR1 module at the other will not form a link, because the receiver at each end expects a different wavelength plan.

Is a 400G spine worth buying for a 100G fabric today?

The port arithmetic decides this. Arista’s 7280DR3-24 carries 24 QSFP-DD cages that run 400G, or up to 96 x 100G interfaces when populated with QSFP28 optics (Arista 7280R3 datasheet). One 1U chassis therefore covers a 24-leaf fabric at four 100G links per spine, with the 400G option held for the next generation.

How much buffer does a storage-heavy fabric need?

Arista publishes the range inside one family: 2 GB on the 7280TR3-40C6 and 7280SR3-40YC6, 4 GB on the 7280SR3-48YC8, 8 GB on the 7280CR3-32P4, and 16 GB on the 7280CR3-96 and 7280CR3K-96 (Arista 7280R3 datasheet). A 1 GB burst at 100 Gb/s drains in 80 ms at line rate, so the buffer figure matters most when the oversubscription ratio is above 1:1 and the traffic arrives in bursts.

Does the QFX5110-48S support 100G on all four QSFP28 ports?

Yes. Ports 48 through 51 each run at 100G or 40G, and the four ports can serve as access ports or uplinks (Juniper QFX5110 Hardware Guide). The guide gives the 48S an aggregate throughput of 960 Gbps, a vendor figure above the 880 Gbps sum of its 48 x 10G and four 100G ports. Each of the four ports also breaks out to four 10G interfaces with a DAC breakout cable, which covers a small 10G server group without a second leaf.

What is the difference between the Arista 7280CR3-96 and 7280CR3K-96?

The datasheet model comparison gives both the same 96 QSFP100 ports, 9.6 Tbps (19.2 Tbps full duplex), 4 Bpps, 3.8 us latency, 16 GB buffer, 2U chassis and 1003 W typical draw (Arista 7280R3 datasheet). The difference sits in software scale and licensing: the catalogue record for the CR3K lists support for more than 5 million IPv4 routes and Group 4 EOS licensing with a minimum of EOS 4.24.0. Route-table scale justifies the K suffix; the port count is identical.

Need a specific part?

Send the part number and quantity — we confirm condition, availability and price before you commit.