800G Optical Transceivers: A Complete Spec Guide to High-Speed Optical Modules in the AI Era

In 2021 I wrote this article: https://www.drflyout.com/post/400g-optical-transceiver-spec Thank you all for the support and feedback. As we now enter the 800G era, I have updated it accordingly.

Moving from 400G to 800G is not just a doubling of speed — it is the optical communications industry's full response to AI clusters, GPU SuperPODs and scale-up / scale-out architectures. 800G has become the mainstream procurement spec for hyperscale data centers, and the most critical technical battleground across the entire optical supply chain.

This article is my reorganized overview of 800G optical module specifications, written for engineers, investors and anyone interested in optical communications, to quickly understand the latest form factors, packaging technologies, reach classes and industry direction.

1. Why 800G?

Traffic demand from AI clusters is exploding over 2024–2027:

• Each GPU needs several Tbps of I/O

• Tens of thousands of GPUs form hyperscale clusters

• East-west traffic accounts for more than 80% of the total

These demands are directly pushing 800G to become the mainstream optical module spec for short-to-mid reach (SR / DR).

The most important shifts:

• 400G → 800G: speed doubles, but power and thermal pressure surge

• 8-lane PAM4 → 4-lane 200G/lane / 100G/lane

• Packaging moves from DSP + EML → LPO / CPO / SiPh (silicon photonics)

2. Common 800G Form Factors: QSFP-DD / OSFP / OSFP-XD

QSFP-DD 800G

Form factor characteristics:

• Dimensions: 18.35 mm × 89.4 mm (double-row gold fingers)

• Lane configuration: 8×100G PAM4 (8-lane)

• Power capacity: up to ~16–18W (about 14–16W in practice)

• Signal interface: uses 8×50/100G data lanes

• Thermal design: mainly front-end heatsink + top-cover heat conduction
  

• Easy to plug, highly compatible, and the easiest to deploy on top of existing 400G infrastructure.

Scalability today:

• Practical limit is around 1.0–1.2T; beyond that, thermals and size constrain further scaling

• So beyond 1.6T, QSFP-DD is no longer the mainstream choice

Positioning: high volume, low cost, highly interchangeable

Hyperscalers (Meta / Google) used it heavily in the past, but AI clusters are increasingly moving to OSFP.

OSFP 800G

Form factor characteristics:

• Dimensions: 22.58 mm × 107.8 mm (larger than QSFP-DD)

• Lane configuration: 8×100G or 4×200G PAM4 (depending on SR8 / DR8 / FR4)

• Power capacity: up to ~20–24W (far more thermal headroom than QSFP-DD)

• Thermal design: supports thicker heatsinks, airflow channels and side heat conduction

• Electrical interface: more stable high-frequency SI (signal integrity)

Why do AI switches mostly choose OSFP?

• Greater thermal capacity, able to support LPO, DSP, WDM and SiPh

• For 1.6T: OSFP has introduced OSFP 4.0, supporting 200G/lane

• Broadcom Tomahawk and NVIDIA Spectrum series are primarily OSFP

Scalability:

• 800G → 1.6T is the mainstream path

• 3.2T is still possible, but thermals become very tight (leaning toward CPO/OIO)

Positioning: the mainstream form factor of the 800G–1.6T era

In real AI switch deployments, it is more popular than QSFP-DD.

OSFP-XD (Extreme Density)

Form factor characteristics:

• Dimensions: 32.0 mm wide × 141 mm long (noticeably larger)

• Lane configuration: designed to support 16×100G / 16×200G (for 1.6T / 3.2T)

• Power capacity: up to 38–45W (depending on version)

• Thermal expansion: can use larger heatsinks, thicker fins and ducted airflow designs

Why is OSFP-XD needed?

• As speeds rise to 1.6T / 3.2T, standard OSFP runs short on power and thermal capacity

• Especially for the combination of WDM silicon photonics + high-power CW laser + high-speed driver/TIA
  

• XD trades more physical space for better SI and thermal management

Scalability:

• Specified directly for future 1.6T / 3.2T / 6.4T needs

• Several vendors have already shown OSFP-XD 3.2T demos (Coherent, InnoLight, etc.)

Positioning: a high-density module spec truly aimed at 1.6T–3.2T

For 800G it is a "reserved upgrade path", not the mainstream spec.

Form factor	Dimensions	Power capacity	Lanes	Target speeds	Thermal capacity	Roadmap
QSFP-DD	18.35×89.4 mm	16–18W	8×100G	400G / 800G	★★★☆☆	Stops at 1.0T
OSFP	22.58×107.8 mm	20–24W	8×100G / 4×200G	800G / 1.6T	★★★★☆	Mainstream 1.6T
OSFP-XD	32×141 mm	38–45W	16×100G / 16×200G	1.6T / 3.2T	★★★★★	Aimed at 3.2T+

3. Three Mainstream 800G Module Types (by Reach)

800G SR8 (short reach: ≤100 m)

• 8×100G / 850 nm

• Multimode fiber (MMF)

• Components: VCSEL, MLA (multi-lane fiber array)

• Use case: rack-to-rack within the row

800G DR8 / DR8+ (mid reach: 500 m – 2 km)

• 8×100G / 1310 nm

• Single-mode fiber (SMF)

• Components: EML, DFB, silicon photonics

• Use case: AI switches, leaf-spine architectures

800G 2xFR4 / FR4 (long reach: 2 km – 10 km)

• Uses WDM (CWDM4)

• 4×200G PAM4

• Usually paired with silicon photonics + CW laser

4. 800G Internal Architectures: DSP / LPO / SiPh

DSP-based 800G (mainstream)

• Reliable and mature

• Higher cost, power around 14–16W

• Suited for: DR8 / FR4

LPO (Linear Pluggable Optics)

• No DSP → significantly lower power

• But places very high linear-equalization demands on the host side (NIC/ASIC)

Typical LPO specs:

• Power can reach the 8–10W range

• Mainly SiPh transceivers + CW laser

CPO (Co-packaged Optics) / OIO (Optical I/O)

• The ultimate solution: place the optics right next to the ASIC

• NVIDIA / Broadcom / Intel / Ayar Labs are all pushing it

5. Technical Spec Comparison of 800G Module Types

Spec	Lane rate	Fiber	Wavelength	Lane configuration	Reach	Common technology
800G SR8	8×100G PAM4	MMF	850 nm	8 Tx + 8 Rx	70–100 m	VCSEL
800G DR8	8×100G PAM4	SMF	1310 nm	8 Tx + 8 Rx	500 m–2 km	EML / SiPh
800G DR8+	8×100G	SMF	1310 nm	8 ch	2 km	SiPh
800G 2×FR4	4×200G PAM4	SMF	CWDM4	4 ch	2 km	SiPh + CW Laser
800G FR4	4×200G PAM4	SMF	CWDM4	4 ch	10 km	SiPh

6. Power Consumption Today (2025)

Power is one of the biggest battlegrounds for 800G:

Type	Average power	Notes
DSP 800G SR8	14–16W	Mature process, in volume production
DSP 800G DR8	16–18W	Widely used for switch-to-switch links
LPO 800G	8–12W	Requires full system-level design
SiPh 800G FR4	14–16W	WDM adds extra power
CPO optical engine	35–65% lower than pluggables	Mainstay of the data center roadmap

7. The Industry View: the 800G Supply Chain

Upstream components

• EML vendors: Lumentum, Coherent, Wuhu, Accelink

• CW laser: Furukawa, Lumentum, Coherent, TXC

• Driver/TIA: Broadcom, MaxLinear, MACOM, Semtech

• DSP: Marvell, Broadcom, MaxLinear, Credo

Midstream modules

• Coherent

• Lumentum

• InnoLight

• Accelink

• Hisense

• Applied Optoelectronics

• Eoptolink

Switch ASIC & System

• Broadcom Tomahawk series

• NVIDIA Spectrum series

• Cisco Silicon One

• Marvell Teralynx series

8. Industry Trends for 2025–2027 (My Own Observations)

1. LPO and DSP will coexist rather than replace each other

Large data centers will mix them:

LPO for short reach → energy savings

DSP for mid-to-long reach → stability

2. SiPh (silicon photonics) will become mainstream for 800G

Because it is better at integrating the CW laser, transceiver and modulator.

3. Power will become the biggest pain point for AI data centers

So 800G modules will transition toward CPO / OIO.

4. 1.6T will go mainstream faster than expected

800G is a transitional spec with a "fast payback".

Summary: 800G Is the "Infrastructure" of the AI Era

Think of 800G as the road network inside the data center:

• AI clusters have long since entered the multi-Tbps era

• Switches, GPUs and NICs all need faster I/O

• 800G = the solution that best balances cost, maturity and efficiency

The key to this generation is not "speed" but:

• How to lower power

• How packaging evolves

• How silicon photonics becomes mainstream

800G will be the highest-volume spec over 2025–2027, with 1.6T / 3.2T following as the next wave.