A Kardashev scale for Network Interface Cards

Nikolai Kardashev was a Russian astrophysicist who devised a method to measure a civilization's status in technological evolution based on the amount of energy it is capable of harnessing and using. This is known as Kardashev scale. The more energy a civilization can harness, the more advanced it is. The levels proposed by Kardashev are:
Type I Civilization: Uses all the energy available on its home planet
Type II Civilization: Uses all the energy of its star
Type III Civilization: Uses all the energy of its galaxy
The Kardashev scale has been studied and reassessed by scientists, notably Carl Sagan, John D. Barrow, and Zoltán Galántai.

Just as civilizations harness energy, Network Interface Cards (NICs) harness bandwidth and intelligence.  What if we use this metaphor to measure the NICs by their technological advancements.  What if we measure NICs by the bandwidth harnessed and optimized, and the intelligent features such as programmability and autonomy.  So the scope would begin with local host traffic and expand to rack and then data center and then to intelligent fabrics.

Level	Analogy	NIC Capability	Examples
1	Planetary - Local	Designed to handle efficient communication within a single host or rack. Built for CPU–memory–disk traffic.	Basic Gigabit Ethernet, legacy 1GbE NICs
2	Stellar - Data Center	Harnesses full data center bandwidth, enabling low-latency, high-throughput communication across racks.	10–100 GbE, RDMA (RoCE/iWARP), SR-IOV
3	Galactic - Global	Operate seamlessly across distributed clusters and geographies, optimizing WAN + cloud interconnects.	NICs who support NVMe-over-Fabrics, SmartNICs, DPUs, programmable NICs with offload engines
4	Universal - AI/Exascale	Self-optimizing NICs that orchestrate massive parallel I/O across exascale supercomputers or AI fabrics.	CXL-integrated NICs, ultra-low latency optical NICs
5	Beyond Universal - Cognitive	NICs that not only move data but predict, adapt, and optimize traffic autonomously using AI. They become “data nervous systems.”	Hypothetical AI-driven NICs with autonomous congestion control, self-healing, and intent-based networking

These distinctions are based on the latency, programmability, and autonomy.  Level 2 became mainstream around 2010 to 2015 with the widespread adoption of 10GbE and RDMA.  Today, in year 2026, most production NICs are level 2 or level 3.  Level 4 is emerging.  Level 5 is currently speculative.

If I see where the Marvell FastLinQ Converged Network Adapters (CNAs) are (at which level), the Everest 3 adapter (QLE84XX-CNA) is at level 2 and the Everest 4 adapters (QL41X6X-CNA and QL45X6X-CNA) are at level 3.  If Marvell would have designed the next level of FastLinQ adapter, that would have been at level 4.

The Kardashev-inspired lens reveals a simple truth - network interfaces are evolving from wires to accelerators to distributed intelligence layers.  In the year 2026, we are midway through the journey.  400 Gbps NICs and photonic NICs are emerging now.  The next leap will come when NICs won't just execute instructions but start making decisions, thereby transforming the network from a transport medium into an active participant in computing itself.  Now that the age of AI has already begun, when will level 4 become commonplace and when would level 5 arrive?  I'd like to know your thoughts in the comments.

Is a DoS attack similar to Slowloris possible on the layer 3 network protocols? Or a Network interface controller (NIC)?

I have written this article in collaboration with Krishna Bhandarge  

Background

Slowloris is a type of denial of service (DoS) attack which exploits the application‑layer (HTTP) connection handling weaknesses.  The Slowloris is a technique of capitalizing on how web servers keep open partial HTTP requests.

A web server allocates a thread for each incoming request.  Slowloris opens a connection to the target web server and sends partial HTTP headers.  Slowloris holds the connection open as long as possible by sending subsequent HTTP headers, adding to, but never completing the request.  By holding many such connections, Slowloris attempts to exhaust the server's thread / connection pool.  If the affected server keeps the connections open, the maximum concurrent connection pool would get filled completely, eventually denying additional connection attempts from the web clients.

Can the Slowloris attack concept be used to exploit the layer 3 network protocols?  Can the Slowloris attack concept be used to exploit a Network interface controller (NIC)?

A Network interface controller (NIC) operates at Layer 1 (physical layer) and Layer 2 (data link layer).  Slowloris relies entirely on connection handling weaknesses.  Layer 1, layer 2 and layer 3 devices, including NICs, do not maintain a per‑connection state.  Hence, this attack pattern is impossible at Layer 2 and layer 3.

Layer 3 DoS attacks are possible, but they are different from the Slowloris concept.  They are :

- ICMP floods

- Fragmentation floods

- IP packet storms

- Malformed IP packet attacks

These attacks attempt to saturate:

- NIC Rx queues (ring buffers)

- Host CPU network stack processing

- PCIe bandwidth

- Device driver's interrupt-handling capacity

None of these attacks resemble the Slowloris pattern because:

- They do not use partial requests

- They do not require keeping connections open

- They do not exploit HTTP behavior

If a Slowloris attack is happening on a web server, the web server's NIC has no way to detect the attack.  The NIC can not provide any protection.  Here is why.

A Network interface controller (NIC) does not do layer 7 filtering.  A NIC typically provides :

- hardware offloads for TCP

- virtualization support (SR-IOV)

- traffic shaping / QoS

- RDMA (RoCE/iWARP) capabilities

But none of these features prevent Slowloris, because Slowloris is not a bandwidth or packet-rate attack. Slowloris is a protocol resource exhaustion attack.

However, SmartNICs and DPUs are the exception.  Some vendors (NVIDIA BlueField, Intel IPU, AMD Pensando) manufacture SmartNICs or Data Processing Units which can offload Layer 7 logic.

On a side note, here are some ways to prevent a Slowloris attack on a web server.  Layer 7 filtering is typically performed by firewalls, proxies, DPI engines, or service meshes.

Mitigations for Slowloris attack are :

- SYN cookies / TCP stack hardening

- connection rate-limiting via OS firewall

- load balancers or reverse proxies (nginx, Varnish)