Intel CPO Alive in Papers but Trails Nvidia on Link Quality [2026]
Monday, September 21, 2026By Indie Kings | September 21, 2026
Updated September 21, 2026: Intel is silicon photonics group is publishing co-packaged optics papers again after attrition and repeated re-orgs had one analyst writing the team off as dead, per Irrational Analysis on August 18. The same teardown grades Intel is 8-wavelength link a C+ against Nvidia is A+, with an Nvidia Lumentum laser paper and an Aeva amplifier paper as bonus reads. Papers prove activity, not a shipping product.
Note: no topical image verified 200 with referer at draft time, so this file posts without art rather than an unlabeled plot. Image filenames on the source page are hashes with no captions, which fails our alt-text honesty rule.
Why co-packaged optics matters
Per 3D InCites packaging primer on August 20, copper stays competitive for short links but channel loss and equalization grow painful as rates and distances rise, which is why optical links earn their place at bandwidth and distance extremes. Co-packaged optics moves the optical engine next to the switch ASIC to shorten the electrical path, per the same primer.
Per the same primer, the trade is packaging pain: heat from hundred-watt ASICs against temperature-sensitive photonics, micron optical alignment, plus test and yield across every die in the package. Relevance to our CS-4 draft is direct: Cerebras attacks the same power and interconnect wall with wafer scale instead of optics.
Per Nvidia March 2026 news, the stakes drew $2 billion each into Lumentum and Coherent plus a multibillion purchase commitment for laser capacity. Photonics is procurement strategy now, not lab curiosity.
- Win: Shorter electrical paths, less equalization, WDM sharing.
- Cost: Thermal co-design, alignment tolerance, package yield.
- Metric: System energy per bit, not waveguide physics alone.
What the Intel papers show
Per Irrational Analysis, Intel is link amplifies 8 modulated wavelengths through one semiconductor optical amplifier, built on a hybrid-integration laser with gratings etched in silicon in a Scintil-like approach. The author notes a rumor that Intel and Scintil reviewed patents and found no infringement, flagged as rumor.
Positive notes per the author: plus or minus 10 GHz spacing accuracy called insanely good, 5 dB extinction on ring modulators against a typical 3 to 4 dB, and a 75 dB TIA gain stage called damn nice. Average 1.3 dB coupling loss over two lenses is rated ok for a periscope-style attachable connector tradeoff.
Per the author, no laser wavelength control is mentioned, which puzzles him since co-integrated lasers cannot use TEC temperature tuning. Fabrication accuracy must carry the burden, per his read.
| Intel result per teardown | Author grade |
|---|---|
| Spacing accuracy plus or minus 10 GHz | Praised |
| Ring extinction 5 dB all lanes | Praised |
| Coupling 1.3 dB two lenses | Acceptable tradeoff |
| 16-wavelength follow-up | Doubted, interleaved not real |
Where the analyst flags gaps
Per the author, one shared SOA across 8 wavelengths spawns four-wave-mixing tones that hide under each carrier and wreck RIN, with true side-mode suppression near 30 dB rather than a claimed 50 dB. Spectrum-clipping and cherry-picked lane pairing allegations are the author is opinion from plot reading, not independently verified facts, so treat them as disputed.
Per the author, bit-error claims below 1e-12 come from bathtub curves at perfect timing, while Nvidia specs 1e-12 at a wide 0.47 UI window with clock forwarding that Intel lacks. His generous 0.2 UI read of Intel lands near 1e-5. Methodology dispute, not a lab replication.
Per the author, the 16-lambda design keeps 8-wavelength Tx buses for thermal reasons but uses three cascaded MZI stages on receive, which costs link budget and forces a second SOA plus polarization control. His verdict words are harsh. Read the architecture critique, discount the adjectives.
| Method split per teardown | Nvidia per author | Intel per author |
|---|---|---|
| Timing strategy | Clock forwarded, low jitter | No clock forwarding |
| BER window | 1e-12 at 0.47 UI | Sub 1e-12 at perfect timing claimed |
| Amplifier path | Per-channel SOA avoids mixing | Shared SOA invites mixing |
| Author grade | A+ | C+ to B- |
Nvidia Lumentum and Aeva contrast
Per the teardown of the Nvidia Lumentum paper, the monolithic DFB array with one SOA per channel runs about 100 mW per channel in a MOPA layout that dodges multi-wavelength mixing. The author still dislikes monolithic DFB arrays for yield and spacing, calling the 2.25 by 1 mm chip poor for yield as his own modeling take.
Per the same teardown, measured RIN near minus 137 dBc per Hz with a relaxation oscillation spike reads unusable for real links in his judgment, against excellent power flatness, spacing and SMSR. Single TEC means no thermal spacing control, with a one-isolator package as a genuine cost win. Nvidia research framing around the RIN is dismissed by the author as cope, which is rhetoric, not data.
Per the teardown, Aeva is SOA earns real praise: 22 percent efficiency at 650 mW and 40C, ASE noise floor 40 dB down, FIT under 5. Bonus modeling alpha from Ben Lee notes a 10 percent laser efficiency threshold for CPO energy sense plus ring self-heating warnings, per the author is reading.
- Nvidia Lumentum: Clean architecture, RIN miss per author.
- Aeva: Strong amplifier numbers per author.
- Intel: Alive and publishing, link quality questioned per author.
FAQ
Is Intel is CPO group dead?
Per the author is reversal, no. New papers show activity after attrition and re-orgs, though papers are not products.
How does Intel trail Nvidia?
Per the author, timing without clock forwarding plus shared-SOA noise plus BER method gaps. Single-analyst read, unverified by replication.
Did Intel cheat measurements?
Alleged by the author from plot reading only. Treat as disputed opinion, not established fact.
What about the laser tech?
Per the author, hybrid integration like Scintil differs from Openlight and Tower single-wavelength 20 mW paths. Patent non-infringement is rumor.
Why does Nvidia invest billions?
Per March 2026 announcements, $2 billion each into Lumentum and Coherent plus purchase commitments to lock laser supply for AI optics.
Is this financial advice?
No. Hardware analysis only. The source author discloses semiconductor positions and states opinions are his own.
Bottom Line
Intel photonics is alive on paper with respectable spacing and extinction, but the noise, timing and 16-lambda architecture questions in this teardown keep it behind Nvidia is approach until independent data lands. Read papers as progress reports, buy on shipping systems.
Related: Intel confirms Arc future with Xe3P | Intel Arc Battlemage failure or future from 2025