Bandwidth is a design decision, not a given
Elastic Network Adapter (ENA) driver choice, cluster placement groups, and instance-type bandwidth ceilings together determine real achievable throughput, and the wrong choice silently caps it well below what the account-level quota allows.
Speaker notes
- The Elastic Network Adapter (ENA) is AWS's high-performance network driver, required for an instance to reach its full documented network bandwidth; an older or generic driver on the same instance type can silently cap throughput well below what the instance is actually capable of, with no error, just a lower number than expected.
- A cluster placement group packs instances close together on the same underlying hardware specifically to minimize inter-instance network latency and maximize achievable bandwidth between them, the standard choice for tightly-coupled workloads like HPC or a distributed cache, at the explicit tradeoff of losing the availability benefit a spread placement group provides.
- Real bandwidth ceilings are instance-type-specific, not account-wide: a smaller instance type has a genuinely lower network bandwidth cap than a larger one in the same family, regardless of what your account-level service quota allows, sizing for network throughput needs its own check against the specific instance type's documented network performance tier, not just vCPU and memory.
Deck map
01
AWS Networking, end to end
02
The certification landscape: Associate to Professional to Specialty
03
The VPC is the network; everything else attaches to it
04
A subnet lives in exactly one AZ, and that's the whole point
05
The route table decides, the resource doesn't
06
An Internet Gateway makes a subnet public, nothing else does
07
NAT Gateway vs. NAT Instance: managed reliability vs. control
08
Security Groups are stateful, and that's the whole trick
09
Two firewalls, two layers, evaluated in a fixed order
10
Elastic IPs and ENIs: identity that outlives the instance
11
VPC Peering is direct, private, and never transitive
12
Gateway endpoints are free; interface endpoints are PrivateLink
13
ALB, NLB, and Gateway Load Balancer solve different problems
14
Six routing policies, one DNS service
15
CloudFront optimizes what's delivered, not how it travels
16
Shield, WAF, and the edge defense stack
17
VPC Flow Logs: the record of what actually happened
18
Transit Gateway: from N-squared peering to one hub
19
Direct Connect: a physical circuit, not a tunnel
20
Direct Connect resiliency is a model you choose, not a default
21
Direct Connect Gateway: one circuit, many regions
22
Site-to-Site VPN and Client VPN solve different connectivity problems
23
Choosing hybrid connectivity: bandwidth vs. lead time vs. cost
24
Cloud WAN: the network becomes a policy document
25
VPC sharing: one network, many accounts
26
PrivateLink: a service exposed as an ENI, nothing more
27
VPC Lattice moves service networking up to Layer 7
28
PrivateLink vs. VPC Lattice vs. Transit Gateway: the 2026 pattern
29
Route 53 Resolver: DNS that crosses the on-premises boundary
30
Five layers deep: NACL to Shield Advanced
31
Global Accelerator: two static IPs, one fast path
32
Three tools that answer 'why can't this reach that'
33
IPv6 in AWS is additive, not a migration
34
Bandwidth is a design decision, not a given
35
Scenario: a global SaaS platform's hub-and-spoke backbone
36
Scenario: a trading platform's latency and compliance walls
37
Scenario: merging two companies' networks without an outage
38
Scenario: a network that survives losing a region
39
Scenario: exiting a data center without a hard cutover
40
The Advanced Networking Specialty exam, domain by domain
41
The traps that recur across every certification level
42
Quick-fire recall
43
The one-page decision cheat sheet
44
Readiness checklist: is this network actually enterprise-grade?