Singtel Fiber vs StarHub Business for SG Processing

1. What “SG Processing” Actually Requires from a Line

Processing in Singapore is not a generic “internet for the office” use case. The workloads that break your operation are the ones pinned to real-time constraints: SGX order-entry sessions, credit-card authorization, NETS payment settlement, or cron-driven reconciliation between a UAT cluster and an Oracle database in Keppel DC.

Those need three concrete numbers, not a “1Gbps” sticker:

– RTT to the processing node below 5ms. If the packet crosses more than four L3 hops from your router to the DC spine, the round trip climbs past 6–7ms and starts killing timeout budgets.

– Core packet loss under 0.01%. A shared contention ratio of 1:48 will drop frames at 8pm even on uncongested fiber. For a payment callback, a dropped SYN means a re-try — which means double latency.

– Jitter under 1ms. Trading gateways and API aggregators fail on bursty latency, not average latency. A contended line delivers perfect pings at noon and 8ms swings at 6:30pm.

Many companies get a “business fiber” plan and never check whether the ISP’s last-mile path is actually on-net to the processing host. That is the single mistake this comparison exists to prevent.

2. Singtel Enterprise Fibre: STiX Backbone and Static IP Registry

Singtel sells two different products under the fiber banner, and the second one is the one worth buying for SG processing.

The consumer-grade “Business Fibre” — 300Mbps to 1Gbps symmetric, shared aggregation — starts around SGD $130–$180 per month. It comes with a dynamic IPv4 address by default, and the SLA is effectively network-level only. Do not cable that into a production payment server.

The relevant product is Singtel Enterprise Fibre (DIA). A 1Gbps symmetrical dedicated line contracts in the $1,200–$1,800/month range on a 24-month term. What you pay for:

– Dedicated aggregation port — no contention ratio on the upstream, verified by the fact that Singtel issues a 99.99% uptime SLA with real service credits, liquidated damages that scale with outage duration.

– Static IPv4 blocks — from a single IP up to a /29, with reverse DNS control. Critical because SGX gateways and bank SFTP endpoints whitelist your source IP.

– BGP session on request — the DIA product lets you announce your own ASN and prefix into the STiX exchange (Singapore Internet Exchange). That puts your traffic on Singtel’s own Tier-1 backbone, which terminates inside Equinix SG1, SG2, and the three main Singtel DCs without hair-pinning through a third party.

The concrete result: an ICMP probe from a Singtel DIA router to an SGX trading node typically returns 1.8–2.4ms RTT with zero packet loss across a 72-hour test window. That is the number that matters for a 500ms POS timeout.

3. StarHub Business: Cheaper Bandwidth, Contended Edge

StarHub’s equivalent is BizFibre, sold at 500Mbps, 1Gbps, 2Gbps, and 10Gbps, all symmetric. The 1Gbps shared plan runs roughly $150–$200/month, and the 10Gbps contended product drops to about $2,800/month. Static IPv4 is an add-on, usually a /29 block for an extra $30–$50 per month.

What is different at the network level:

– Aggregation is shared. StarHub’s 1Gbps and 2Gbps plans use PON or active ethernet with oversubscription ratios that are not published but behave like 1:24 to 1:48 during 6–10pm. Your bandwidth test shows line-rate at 3am and drops to 60–70% at peak, which is fine for backups but fatal for synchronized replication.

– Peering is smaller. StarHub has an AS (AS55430) with a presence at SGIX, but its international and data-center footprint is not Singtel’s. For traffic destined to an Equinix fabric, StarHub often interconnects via transit or buys a patch through a third-party exchange — adding one or two extra hops that show up as +1.5–2ms extra RTT.

– SLA credits are shallow. StarHub’s business SLA for the access network is 99.9%, and the credit formula typically maxes out at one month’s fee for a total outage. Contrast that with Singtel DIA’s per-incident credits and you’ll see why nobody runs a clearing system on a $200/month line.

StarHub’s legitimate strength is bandwidth-per-dollar. For bulk ETL — copying 8TB nightly from an M1 colo rack to a Synology NAS in a branch office — StarHub 10Gbps is a superior buy because the transfer is time-shifted to off-peak anyway and one retry costs nothing.

4. Real Route Behaviour for Processing Nodes

Run this before you sign either contract. Configure a probe router, execute a 24-hour MTR to your actual processing endpoint, and compare.

Singtel DIA path, from a Raffles Place office to a server in Equinix SG1:

“`

hop 1 router.local 0.4ms 0.4ms 0.3ms

hop 2 sng-agg-01.stix.sg 1.1ms 1.0ms 1.1ms

hop 3 eqx-sg1-01.stix.sg 2.1ms 1.9ms 2.0ms

hop 4 target-server 2.3ms 2.2ms 2.3ms

“`

Four hops, under 2.5ms, zero loss. That is a line that can carry a real-time reconciliation loop without a trip to the drawing board.

StarHub BizFibre path, same target:

“`

hop 1 router.local 0.5ms 0.4ms 0.4ms

hop 2 sh-agg-01.starhub.net.sg 1.9ms 1.8ms 1.8ms

hop 3 sh-core-02.starhub.net.sg 3.4ms 3.6ms 4.0ms

hop 4 transit-peer.equinix.net.sg 4.8ms 4.9ms 5.6ms

hop 5 target-server 5.2ms 5.1ms 6.0ms

“`

Five hops, 5.2ms, with jitter spikes up to 6ms during peak hours. The RTT is still ‘green’ to an office LAN, but the jitter is what bites: a payment API with a 500ms timeout and three retries will pass the test 95% of the time and fail the remaining 5%, randomly, at peak. That is the failure mode of a contended line.

5. Verdict: Pair the Line with the Workload Type

There is no universal winner. There are only correct line-to-workload mappings.

The decision rule: if the processing pipeline has a timeout or a money settlement, pay for Singtel DIA — the 3ms RTT and zero-loss core are the difference between transactions that clear and transactions that silently void. If the job is a scheduled bulk mover that can restart on failure, StarHub’s 10Gbps at under $3,000 is the rational cost choice.

The worst error you can make is sizing the link by download speed. Processing in Singapore is a latency story. The line that routes on-net to your host is the one you build on; the line that routes through a transit peer is the one you debug at 2am.