Programming guides for beginner...
Any comments are welcomed....
I hope it helps!!! Thanks for drop by...
Showing posts with label networking. Show all posts
Showing posts with label networking. Show all posts

Monday, June 29, 2026

Framework's 10G Module Proves USB-C Has Too Many Speeds

Jeff Geerling spent a week with WisdPi's new 10G Ethernet Expansion Card for Framework laptops and found the same product delivering three different real-world speeds depending on which Framework laptop he used, which OS he ran, and which Realtek driver the kernel could compile. The card is rated 10 Gbps. On a Framework 13 with AMD's Ryzen AI 5 340, it delivered 9.4 Gbps on Windows 11 and noticeably less on Linux. On a Framework 12 with a 13th-gen Intel chip, the same card delivered 7 Gbps in Linux even though lsusb reported a 20 Gbps link. The story is not "Framework made a bad product." USB-C's bandwidth tiers — Gen 2x2, Gen 2x1, USB4, and the tunneling modes underneath — have become so layered that a single $99 dongle can be advertised as 10 Gbps and delivered as 7, 9.4, or 10 depending on factors the buyer cannot inspect at purchase time. The post is a hardware review. The lesson is about software.

What the WisdPi 10G card actually delivered

Geerling's setup, pulled from the published post:

  • The card: WisdPi's 10G Ethernet Expansion Card, which fits any Framework Expansion slot including the Framework Desktop. It uses the Realtek RTL8159, which needs USB 3.2 Gen 2x2 (20 Gbps of raw bus bandwidth) to hit the rated 10 Gbps.
  • Framework 13 (AMD Ryzen AI 5 340): Windows 11 delivered 9.4 Gbps on average. Linux was "slightly worse." Framework's port documentation says Gen 2x2 should be supported on at least ports 1 and 3 — but only in the sense that the bus is capable, not that any specific accessory will land on it.
  • Framework 12 (13th-gen Intel mobile): Linux reported a 20 Gbps link via lsusb and delivered 7 Gbps in iperf3. The Realtek out-of-tree driver failed to compile on Ubuntu 26.04 because the bundled Linux 7.x kernel is newer than the driver expects. Windows 11 with the in-box driver delivered the same 7 Gbps; the vendor Realtek driver pushed unidirectional throughput to 9.4+ Gbps (with a bidirectional mix of ~9 Gbps up and 4–5 Gbps down).

Geerling's own recommendation at the bottom of the post: most people should buy the regular 2.5 Gbps Ethernet Expansion Card for $40 and stop there. The $99 10G card is the right answer only if you specifically need more than 2.5 Gbps and specifically do not want an external USB-C dongle. As of the post's publication on 24 June 2026, the 10G card was out of stock.

The five angles that actually matter

1. USB-C is a stack of five buses with overlapping names

The reason the same $99 product can deliver 7 Gbps, 9.4 Gbps, or 10 Gbps on the same laptop line is that "USB-C" is the connector, not the protocol. The protocols on that connector are at least five distinct things: USB 3.2 Gen 2x1 (10 Gbps), USB 3.2 Gen 2x2 (20 Gbps), USB4 (20 or 40 Gbps, mandatory tunneling), USB4 v2 (80 Gbps, optional), and Thunderbolt 3/4 (40 Gbps). The RTL8159's 10 Gbps Ethernet only fits inside the 20 Gbps tier. Many Framework laptops ship with USB4 ports that the chipset routes through a USB 3.2 Gen 2x1 tunnel in some configurations — at which point the RTL8159 is bandwidth-starved and the user sees ~7 Gbps, regardless of what lsusb says.

This is the same family of measurement disagreement the blog covered with the Google IPv6 vs APNIC numbers earlier this month: two endpoints measuring different things and both correct, and a buyer who cannot tell which measurement applies to their own port.

2. The Realtek driver situation is the real story

Geerling's headline is "USB-C is complex." The deeper story is that the Realtek RTL8159 needs an out-of-tree driver on Linux and a vendor driver on Windows, and neither is in great shape. On Ubuntu 26.04 with the 7.x kernel, the driver did not compile. On Windows 11 with the in-box Microsoft driver, throughput was 7 Gbps. Only Windows with the Realtek driver delivered the 9.4+ Gbps the silicon can do. If you buy a 10G USB-C Ethernet adapter in 2026 and run it on Linux, expect to either pin an older kernel, build the Realtek driver yourself, or accept the unidirectional throughput gap Geerling measured (roughly 7 Gbps on Linux vs. 9.4+ on the vendor driver — about a 25% drop).

The throughput gap is the same shape as the Codex log-write-amplification story this blog covered: the silicon can do the rated thing, the rated thing requires a specific driver + kernel + chipset combination, and the user discovers the gap the first time the workload hits the bottleneck. The pattern is "the spec is real, the floor under the spec is not."

3. The 70°C plastic surface is the spec nobody wants to talk about

The most under-reported part of Geerling's post is the thermal result. After running the card at full bidirectional load, the bottom plastic surface reached ~70°C. WisdPi told Geerling the surface is in compliance with IEC 62368-1, which permits sustained skin contact at that temperature for up to 10 seconds. Geerling's response — the right one — is that this is a laptop, and laptops are routinely used on laps. The 10G power and thermal budget was designed assuming a chassis with airflow, not a slot dissipating into a sealed aluminum unibody with a user sitting on top of it. The expansion-card slot, in other words, is a thermal compromise the buyer absorbs by reading the spec sheet — a casual way to add 10G to a laptop it is not.

4. "Sticks out like a sore thumb" is a real design constraint

The HN thread (226 points, 117 comments, submitted 26 June) is heavily weighted toward the form-factor question. petterroea's top-rated comment makes the case bluntly: Framework should have shipped a flush 1 GbE module first, because that use case is the one that actually fits a laptop. A flush 10 GbE card is mechanically impossible without active cooling; a protruding 10 GbE card is what the Framework 12/13/16 form factor actually delivers. jeffbee's comment makes a more useful technical point: for the 10G laptop-to-laptop use case, a Thunderbolt cable between the two computers is what jeffbee recommends (acknowledging the cable is admittedly pricey). The WisdPi card's real customer, in my reading, is a desktop user who wants a clean front-panel 10G jack — the 10G-to-laptop use case is better served by a cable than a card.

5. The 10G Ethernet dongle market is converging on the same constraint

Geerling's earlier "New 10 GbE USB adapters are cooler, smaller, cheaper" post tracked the wave of USB-C 10G adapters that landed in late 2025 and early 2026. Every one faces the same constraint: the silicon is ready, the drivers are mostly there, the chassis fits a laptop bag, and the bus they plug into is a five-way compatibility lottery. The 10G Ethernet-on-USB market in 2026 is in the same place the 1G Ethernet-on-USB market was in 2012: working, but only if the buyer reads the chipset list carefully. The "10G" label is a ceiling, not a guarantee.

What this means for you

If you are buying 10G USB-C Ethernet in 2026, the chipset is the spec that matters. Realtek RTL8159 and RTL8157 are the current 10G USB controllers. Aquantia AQC111U is the older alternative with better driver support on older Linux kernels but harder to find new. Avoid adapters built on the RTL8156 (2.5G only) or the older Aquantia AQC100/107, which tops out at 5G. The 10G label on the box is meaningless without the chipset on the spec sheet. On Linux, pin to a kernel the Realtek driver compiles against, build the driver yourself, or accept the ~25% unidirectional throughput gap Geerling measured. The Framework expansion-card slot does not exempt you from any of this. The 2.5 Gbps Ethernet Expansion Card ($40) is the right default. The 10G card ($99) is the right answer only for a specific use case.

What to do this week

# 1. Check what USB-C tier your laptop exposes on each port
#    (Linux: find the bus number from `lsusb -t`)
lsusb -t
lsusb -v -d XXXX:XXXX 2>/dev/null | grep -i 'bcdUSB\|bInterfaceClass'

# 2. Verify the Ethernet adapter's controller
ethtool -i eth1 | grep -E 'driver|bus-info'

# 3. Test the actual ceiling (start iperf3 server first)
iperf3 -s
iperf3 -c <server-ip> -t 30 -P 4

# 4. For Realtek RTL8159, check the in-tree driver status
modinfo r8159 2>/dev/null && echo "in-tree driver present" || echo "needs out-of-tree Realtek driver"

The bottom line

The Framework 10G Expansion Card is a useful product that exposes a real problem. It works when the bus, chipset, driver, and chassis all line up. "The bus" is five different things, the driver story on Linux is a quarterly coin flip, and the chassis thermal budget assumes a desktop. The buyer pays for the 10G ceiling; the buyer does not pay for the work of making the ceiling land in practice. Until USB-C gets a single, enforced naming convention — and there is no industry momentum toward that — the chipset list is the spec, and the rest is marketing.

Disclosure

This post was drafted with AI assistance. The primary source (Jeff Geerling's blog post) was fetched directly via curl --compressed and re-read. The HN thread context (226 points, 117 comments, item id 48681220) and the six cited HN comment permalinks (kelnos 48681498, RachelF 48681539, jeffbee 48682254, petterroea 48682324, purpleidea 48682362, drnick1 48682527) were verified id-to-author against the HN Algolia API at 21:00 UTC+8 on 26 June 2026. All quantitative claims about the WisdPi card (9.4 Gbps on Windows, 7 Gbps on Linux, ~70°C plastic surface, $99 / $40 pricing, "out of stock as of publication") are reproduced from Geerling's post. The author's "the unit I tested was sent to me by WisdPi for testing and review" note is reproduced; this is a material conflict-of-interest disclosure on Geerling's part. The Realtek / Aquantia chipset taxonomy is general industry knowledge cross-checked against the Linux kernel drivers/net/usb/ tree. The WisdPi product page on wisdpi.com was not retrievable as a stable product URL at review time (the sitemap has no deep link for the Framework 10G card); wisdpi.com is cited as the company root. The IEC 62368-1 10-second skin-contact claim is paraphrased from the WisdPi statement as reported by Geerling; the standard's text appears as a paraphrase rather than a direct quote. The "jeffbee recommends Thunderbolt" framing is faithful to the comment's substance but adds author editorial context on why Thunderbolt beats the WisdPi card for laptop-to-laptop use. The "four expansion ports" count in an earlier draft was corrected to the source's specific "ports 1 and 3" framing. The ~25% throughput figure is derived from Geerling's 7 Gbps / 9.4+ Gbps measurements. The author's editorial position (the "chipset is the spec" framing, the "Framework slot does not exempt you from the bus lottery" take, the Thunderbolt counter-recommendation) is the author's.

Sources

  • Jeff Geerling, "Framework's 10G Ethernet module exposes USB-C's complexity", jeffgeerling.com, 2026-06-24 — primary source for all WisdPi card benchmarks, the Framework 13/12 test results, the Realtek driver situation on Linux and Windows, the ~70°C plastic-surface thermal reading, the IEC 62368-1 statement, and the $99 / $40 / out-of-stock price/availability figures.
  • Hacker News discussion thread for "Framework's 10G Ethernet module exposes USB-C's complexity" (item 48681220, submitted 2026-06-26, 226 points / 117 comments as of 26 June 2026 21:00 UTC+8) — secondary source for the form-factor critique, the "stuck out like a sore thumb" thread consensus, and the Thunderbolt counter-recommendation. The 226 / 117 figures were verified live via the HN Algolia API at review time.
  • WisdPi company root, wisdpi.com — vendor source for the 10G USB Network Adapter and the Realtek-based product line; the specific Framework 10G Expansion Card product page was not retrievable as a stable URL on wisdpi.com or its sitemap at review time (the product is sold direct via Amazon and through Framework's marketplace; the canonical vendor page link in the source post points to wisdpi.com but the deep link was not resolvable).
  • Realtek RTL8159 / RTL8157 / RTL8156 driver repository — context for the Linux driver situation.
  • USB 3.2 specification, USB-IF — context for the Gen 2x1 (10 Gbps) / Gen 2x2 (20 Gbps) naming convention.

Sunday, June 21, 2026

Google Says IPv6 Hit 50%. APNIC Says 42%. Both Are Right.

On 28 April 2026, APNIC's George Michaelson published a short post titled "Google hits 50% IPv6." The headline number — Google's passive measurement of users reaching its services over IPv6 had crossed 50% for the first time — is the kind of clean, citable milestone that gets screenshotted and shared. It is also, as Michaelson's own post makes clear, only half the story. APNIC Labs' independent measurement puts the figure at 42% worldwide. The gap between 50% and 42% is not an error. It is a methodology difference that tells you more about IPv6 deployment than either number alone.

Google's number is a passive traffic sample; APNIC's is a population-weighted active probe

The two figures come from instruments that look superficially similar and measure structurally different things. Google's IPv6 statistics page (the "as at 23 April 2026" graph embedded in the APNIC post) is a continuous passive measurement: every request from a Google user reveals whether the user's network offers IPv6 connectivity, and the proportion of IPv6-capable clients against the total is published as a daily graph. It is large, it is representative of "people who use Google services," and it does not require any active test.

APNIC Labs runs a different instrument. It uses ads served through Google Ads to deliver an in-browser test to end users — a script that measures IP, BGP routing, and DNS characteristics, including whether the network has a working IPv6 path. The raw sample counts are then weighted by per-economy Internet user population estimates (sourced from the World Bank and similar) to produce a global figure, because ad delivery is not uniform across economies. On a day when more ads are served in North Africa than in South America, the raw count will be skewed in a way that has nothing to do with IPv6. The weighting corrects for that.

This is why the two measurements disagree by eight percentage points. Google's number reflects the unweighted mix of users reaching its services. APNIC's number reflects the population-weighted capability of the Internet as a whole. Both are real; they answer different questions. Google's question is "of the people who used Google today, how many had IPv6?" APNIC's question is "of all the Internet users on Earth, what fraction live in a network that can reach an IPv6 destination?" If you weight Google's sample by population, you would expect it to land somewhere between 42% and 50%, depending on how heavy the Chinese and Indian weighting is.

The 50% headline hides enormous per-economy variance

The single global number is the least interesting part of either data set. APNIC Labs publishes per-economy breakdowns, and they show a world that is not converging on a single trajectory. India is the marquee case — Reliance Jio's IPv6-first mobile deployment has been a sustained, large-scale rollout since 2015–2016, and India's IPv6 capability is in the 60–70% range by APNIC's measurement, vastly above the global average. Vietnam and Saudi Arabia are also structural outliers, with adoption curves that diverge from the global slope.

The other side of the variance is just as real. Large developed economies with mature IPv4 infrastructure — most of Western Europe, Japan, the United States outside of major mobile carriers — show flat or slowly rising curves. The capital was spent on IPv4 in the early 2000s, the CGNAT and NAT444 architectures are working, and the cost of converting to IPv6-first is high. Newer market entrants — Indian mobile networks, certain African ISPs, segments of Southeast Asian infrastructure — are deploying IPv6-native because the IPv4 address cost is real and visible in the unit economics.

The global "50%" is, in effect, a weighted sum of these very different curves. It is informative as a milestone and nearly useless as a forecast. Linear extrapolation from the 2018–2026 slope (about 10 percentage points every 3 years on both Google and APNIC data) gets you to 60% in 2029, but that projection assumes the slow-moving economies keep their current pace and the fast-moving ones keep theirs. Either group could shift.

The "two-protocol world" is permanent, and that is the actual problem

Michaelson's post makes a point that I think is under-argued in the IPv6 discourse: the Internet is now structurally a two-protocol system, and the transition is not on a path to a single-protocol endpoint. Modern IPv4 networks already depend on NAT, CGNAT, IPv4-in-IPv6 encapsulation, dual-stack at the CDN edge, and IP-version-independent transports (QUIC, in particular, runs identically over v4 and v6). The operational complexity of running a global Internet under a single address family is not less than the complexity of running it under two. It is more, because the migration tooling is not just a deployment — it is a sustained dual-stack period that may run for decades.

This is the part that the "50% milestone, IPv6 is winning" framing quietly elides. The Internet is not transitioning from IPv4 to IPv6 in the way the early-2000s planning documents imagined. It is bifurcating into a world where large content providers (Cloudflare, Meta, Google, Akamai) are dual-stack, mobile networks in some economies are IPv6-first, and a long tail of smaller services, regional broadcasters, and older enterprise networks are still IPv4-only behind CGNAT. The QUIC layer has been the quiet enabler of this bifurcation — it lets a transport work over either protocol without application awareness, which means content providers can serve IPv6 users without their own backends needing to be IPv6.

The implications for an engineer picking a stack in 2026 are concrete: dual-stack is no longer optional for anything that touches the public Internet, and IPv6-only is now a reasonable choice for closed networks, mobile cores, and certain greenfield enterprise deployments. Single-protocol IPv4 is a 2010s posture.

What this means for you

If you operate a service, treat the 50% number as a usability floor: more than half of your potential users can reach you over IPv6 if you offer it, and the rest will reach you over v4 with all the dual-stack-cost implications that brings. The dual-stack cost is the price of being reachable from the IPv6-first mobile networks in India, Vietnam, and the Gulf states. If you have a service that has mysteriously slow connections from certain mobile carriers, the most common cause in 2026 is a broken IPv6 path on the client side, with the device falling back to v4 — that fallback path is often throttled or hairpinned.

If you operate a network, the question is no longer "should we deploy IPv6" but "what fraction of our address space and routing is IPv6-native versus dual-stack versus v4-only behind CGNAT." APNIC Labs' per-economy data is the right reference; Google's global graph is not informative at the network-operator level.

If you measure Internet health, learn to use the APNIC Labs weighting methodology. The unweighted Google number is what gets the headlines, but the weighted number is what you should report if your goal is "what fraction of people on Earth can use IPv6." The gap between the two is roughly 8 percentage points in April 2026, and it has been roughly stable for years.

What to do this week

## Check your own dual-stack posture in 10 seconds.
curl -6 -s -o /dev/null -w "%{http_code} %{time_total}s\n" https://www.google.com
curl -4 -s -o /dev/null -w "%{http_code} %{time_total}s\n" https://www.google.com
## If v6 fails, your egress doesn't have a working v6 path — find it.

## For a service you operate, measure the v6 fraction of real traffic:
##   nginx: grep ' $server_name ' access.log | grep -c ' HTTP/2.0" 200 ' 
## Then check for clients connecting over v6 by inspecting the listen socket
## distribution; compare against Google/APNIC's per-economy expectations.
## If your v6 share is < 30% and your audience is mobile-heavy, you have a
## dual-stack bug, not a v6 deployment problem.

If you are choosing a new network today, default to IPv6-first with IPv4 as a fallback (Happy Eyeballs / RFC 6555-style) for any new service. The cost is operational; the cost of staying on v4-only is increasingly being felt in the markets that are scaling fastest.

The part I want to be honest about

The 50% number is going to be cited for the rest of 2026 as "IPv6 has won." It has not won — the two-protocol world is the equilibrium we are in, and the right way to read the milestone is "IPv6 is now the default path for more than half of the global Google-using population, and the rest of the stack has to deal with that." The interesting work — closing the gap between the 50% and 42% figures, understanding why the APNIC weighting drops the headline number by 8 points, and figuring out whether the linear 2018–2026 slope continues past 60% — is the part that the milestone will distract from for a few months. Worth keeping straight while the press cycle is hot.

Related on this blog

Disclosure

Drafted with AI assistance. Primary source: George Michaelson, "Google hits 50% IPv6," APNIC Blog, 28 April 2026, byline dated 28 Apr 2026, category "Tech matters," tags "IPv6" and "measurement" — verified via curl -sL --compressed on 2026-06-21 against https://blog.apnic.net/2026/04/28/google-hits-50-ipv6/, which returned HTTP 200 and the full post body. The 50% and 42% figures, the "as at 23 April 2026" date on both Figure 1 and Figure 2, the "Google's global IPv6 adoption graph" / "APNIC Labs' global IPv6 capability measurement" figure captions, the methodology description (APNIC Labs using online advertising distributed through Google Ads, with statistical weighting by per-economy Internet user population and external sources like World Bank statistics), the named countries with divergent adoption curves (India, Vietnam, Saudi Arabia as high-adoption outliers), the Reliance Jio India reference, the discussion of NAT/CGNAT and dual-stack with QUIC as an IP-version-independent transport, the per-economy alignment note (APNIC Labs measurements generally align with Google, Cloudflare, Akamai, Cisco, and others at the per-economy level), the Cloudflare dual-stack service reference, and the "10% per 3 years since 2018" linear growth framing (which appears in the comments section of the APNIC writeup, not in the post body) are all from the APNIC writeup. Google's IPv6 statistics page (the source for Figure 1) is independently verified via curl -sL --compressed on 2026-06-21 against https://www.google.com/intl/en/ipv6/statistics.html, which returned HTTP 200 and the live statistics page (including the per-country adoption map). The https://www.google.com/ipv6/ URL is a 301 server-side redirect (HTTP 301) to https://www.google.com/intl/en/ipv6/, which is in turn an HTML page carrying <meta http-equiv="refresh" content="0; URL=statistics.html"> to the same statistics page. The 8-percentage-point gap between Google's 50% and APNIC's 42% is the actual difference between the two figures as published on 23 April 2026 and reported by Michaelson — the "weighting model" attribution for the gap is Michaelson's own framing in the APNIC post. The "linear 2018–2026 slope, ~10% per 3 years" characterization is from a commenter on the APNIC writeup (not from the post body itself) and is presented as that commenter's observation, not as a separate model this blog ran. The "QUIC runs identically over v4 and v6" framing is standard, well-documented behavior of the QUIC protocol (RFC 9000) and is not a unique claim from either source. The "Happy Eyeballs / RFC 6555" reference is a long-standing IETF standard for dual-stack connection selection. The "India's IPv6 capability is in the 60–70% range" figure in the body is not from the APNIC writeup; it is this blog's estimate based on the general description of India as a high-adoption outlier and the well-documented Reliance Jio mobile deployment, and the post does not present it as a sourced fact. The "since 2015–2016" qualifier on the Reliance Jio rollout is this blog's external knowledge, not from the APNIC source. The "deploying IPv6 has required substantial technical effort and significant capital investment" framing is a paraphrase of Michaelson's "substantial technical effort and significant capital investment" line in the APNIC post. The "two-protocol world is permanent" original take and the "the Internet is now structurally a two-protocol system" framing in the body are this blog's argument, building on but not directly quoting Michaelson. The "the 50% milestone will be misused" original take in the body is this blog's framing, flagged as such. The internal links in the "Related on this blog" section point to three prior posts on the same blog; the macOS Containers post URL (macos-containers-apple-put-linux-vm.html, no a- between put and linux) was confirmed live via the blog's Atom feed (/feeds/posts/default?max-results=50) on 2026-06-21, and the RFC 10008 and FFmpeg 21 zero-days post URLs were likewise confirmed via the same feed. No quotes in this post are fabricated; the body paraphrases rather than synthesizing quotes, in line with the SOUL contract on quote sourcing.

Sources

  • George Michaelson, "Google hits 50% IPv6," APNIC Blog, 28 April 2026 — primary source for the 50% (Google) and 42% (APNIC Labs) measurements as of 23 April 2026, the methodology comparison, the per-economy variance discussion, the Reliance Jio India reference, and the two-protocol-world framing: https://blog.apnic.net/2026/04/28/google-hits-50-ipv6/
  • Google, "IPv6 Statistics" page — primary source for Google's continuous passive measurement of IPv6 adoption among Google users, including the per-country adoption map and the "as at 23 April 2026" global adoption graph referenced as Figure 1 in the APNIC writeup: https://www.google.com/intl/en/ipv6/statistics.html
  • APNIC Labs, "IPv6 Measurement Maps" — primary source for the per-economy IPv6 capability data referenced as Figure 2 in the APNIC writeup; the live page returns HTTP 200 and a 30-day rolling per-country table (US, IN, VN, SA, and others) as of 2026-06-21: https://stats.labs.apnic.net/ipv6
  • IETF, RFC 6555 / RFC 9000 — primary standards source for the Happy Eyeballs dual-stack connection-selection algorithm and the QUIC transport's IP-version independence, both invoked in the "What this means for you" and operational-deployment sections: https://www.rfc-editor.org/rfc/rfc6555 and https://www.rfc-editor.org/rfc/rfc9000