MiSTer FPGA Forum 2026: 3DO Core, AI Slop Split

The MiSTer FPGA project has always kept two front doors. One is a Discord where the betas drop at 2am; the other is a phpBB forum where the receipts are filed. In 2026 that forum crossed 10,690 registered members and briefly held 42,092 users online at once, then spent its summer arguing about whether a chatbot should be allowed to write its cores. This is a report on the state of the town square: who is shipping what, which cores can never be perfect, and why the most active board in retro hardware is quietly at war with itself.

The Forum Is Still the Project's System of Record

10,690 members, 106,431 posts, one spike of 42,092

Strip away the mythology and misterfpga.org is a stock phpBB board with a stock blue theme. The numbers are not stock. The live board index reports 10,690 members, 106,431 posts across 8,048 topics, and a most-users-online record of 42,092, set on Fri Jun 19, 2026 at 1:20am. A pinned thread still sitting near the top is titled "Over 5000 MiSTer Forum Members!" which means the community has roughly doubled since anyone last stopped to celebrate its size. For a hobby that expects you to know what a lookup table and a block RAM are, that is a genuine crowd, and it is still growing in the middle of a supposed emulation-is-solved era.

A changelog board is a distribution channel

The forum's defining feature is not conversation, it is logistics. The MiSTer Updates and Changelog board is where developers post release notes for new and updated cores directly, which quietly turns a message board into a package index. A core is announced here, the downloader databases register the change, and the device pulls the new bitstream over SSH the next time you run the updater. The forum is less a cafe than a registry with opinions attached.

# A core gets announced on the Updates and Changelog board;
# the device pulls it. SSH in (default root / 1) and run:
cd /media/fat/Scripts
./update_all.sh

# update_all reads downloader .ini databases that mirror the
# MiSTer-devel distribution. Newly announced cores land in
# the right folders automatically:
/media/fat/_Console/    # console cores (.rbf)
/media/fat/_Computer/   # computer cores
/media/fat/_Arcade/     # arcade cores (plus .mra)
/media/fat/games/       # ROMs, BIOS, saves

# Odelot's RetroAchievements build deploys side-by-side
# without breaking update_all, via manyhats-mike's tool.

Mind which mirror you are quoting

One footnote for anyone citing forum stats: a secondary mirror at mail.misterfpga.org serves a snapshot that trails the main board by weeks. A March 8, 2026 capture there listed the "Discussion About The Forum" section at 58 topics and 353 posts, and the "New Users Forum" at 337 topics and 2,685 posts, figures the live index had already blown past. The two hostnames disagree by design, so quote the domain, not just the number.

Inside the Board Index

The sections that carry the weight

The layout is conventional and, for once, that is a compliment. Up top sits News and Announcements, home to Updates and Changelog, Discussion About The Forum, and the Project Wiki. Below it, General collects the New Users Forum, Development for MiSTer, Guides and Videos, and the sprawling General Discussions board. Then come the parts that make MiSTer MiSTer: dedicated per-system Console Core, Computer Core, and Arcade Core forums, where a single console can accumulate years of bug reports, timing-diagram arguments, and save-state regressions. This is where the accuracy work is actually adjudicated, thread by thread.

The addon boards point straight at GitHub

The Hardware area is where the money lives: an FPGA development boards discussion, a Hardware Setups board, and a Sellers board restricted to official addon boards rather than a general flea market. The "Official Project Addon Boards" section links directly to MiSTer-devel/Hardware_MiSTer on GitHub, which tells you everything about the project's structure. The forum is not a walled garden, it is a front end bolted to a developer repo. Anyone can read the schematics; the discussion just happens to have a login.

The board that blocks robots while arguing about robots

And here is the joke the community has not fully appreciated: sometime around mid-2026 the forum hardened itself against AI crawler traffic and now returns a 403 to automated fetchers hammering it for training data. So the board that will not let a bot read it is, on its own front page, hosting the loudest fight in its history over bots that write. The irony is doing a lot of unpaid work.

srg320's 3DO Core: 55% of the Chip, Zero Percent Guaranteed

The developer who said the quiet part out loud

Sergiy "srg320" Dvodnenko is the name behind MiSTer's SNES and Sega Saturn cores, which is to say he has already solved two of the hardest problems on the platform. In 2026 he turned to the 3DO Interactive Multiplayer, the notoriously awkward 1993 console, and then did the thing FPGA developers almost never do: he told everyone up front that it would never be perfect. In an interview with ReadOnlyMemo on April 12, 2026, he was blunt: "The main problem is VRAM emulation. This is likely what prevented other developers from working on the 3DO core. Accurate VRAM emulation requires a certain amount of internal FPGA memory, which it doesn't have. Therefore, unfortunately, the core cannot be accurate on the MiSTer."

Why VRAM is the wall

The constraint is physical, not a matter of effort. The Cyclone V simply does not carry enough on-chip block RAM to hold the 3DO's video memory: only about a third fits internally, and the rest has to spill into external DDR in burst mode, with a page buffer papering over the latency. srg320 is quick to point out the logic itself is cheap. "The core itself is not complex and does not consume a lot of FPGA resources (it currently uses about 55% of the resources)," he said, so this is not a gate-count problem. It is a memory-bandwidth-and-timing problem, and those are the ones you cannot brute-force. He also flagged the 3DO's sprite engine, which renders most of the 2D and 3D output and whose internal FIFO capacities were never documented, meaning, in his words, "absolute accuracy won't be achievable."

Shipped anyway, through Discord, on a dual-SDRAM board

None of that stopped it. An early beta surfaced in March 2026 and, tellingly, it was shared through the MiSTer Discord rather than posted to the forum first, according to GenerationAmiga. It also requires a dual-SDRAM board, putting it out of reach of anyone on a single-RAM setup. That combination, cutting-edge yet honest about its limits and gated behind both a chat app and a hardware upgrade, is the perfect snapshot of where MiSTer development actually happens in 2026.

The 'AI Slop' Core War

When a chatbot ports MAME

The summer's real story had nothing to do with any single console. A wave of new cores appeared that had not been traced from silicon at all; they were generated by pointing a large language model at existing MAME source and asking for register-transfer logic. Jotego, the developer behind the respected Capcom-era jtcores and a man who has spent years reverse-engineering arcade boards by hand, flagged a batch of Gaelco arcade games running on MiSTer and said the quiet part on X: "All these are AI-generated ports of MAME using the JTCORES framework." The framework was his. The cores were not.

Jotego's three-count indictment

His objection, as reported by Time Extension, is not sentimental. He argues AI-generated cores do three concrete things to the platform: they turn it into "a second class MAME, that tries to emulate what the emulator does," they drive human talent away, and they discourage the original hardware research that made FPGA cores worth having in the first place. The entire value proposition of an FPGA core is that it is a simulation of the hardware, not a copy of someone else's software emulator. An AI port of MAME is, by definition, a hardware description of an emulator, a photocopy of a photocopy running on a $200 board.

The forum tries to zone the problem

The community response has been anything but unified. One forum thread is titled, with zero ambiguity, "Segregate Ai-Slop Cores to a Dark Corner of the Forum or Ban Them Altogether." Time Extension's own headlines capture the mood: one piece leads with a reader's verdict that "You Might As Well Play On A Raspberry Pi," another simply calls the trend "Extremely Disappointing." But there is a real counter-argument underneath the sneering, and it is availability. Nobody wants to wait months for a perfectionist to release a cycle-accurate core for a 35-year-old arcade game when Claude Code or Codex can produce a playable one from MAME tonight. That is the same tradeoff that has always separated FPGA purists from the crowd that just wants to spin up two hundred software cores in an afternoon and start playing. The AI cores did not invent the accuracy-versus-availability fight; they industrialized it.

DreamSTer: The 'Impossible' Dreamcast, Held Together With Tape

Not a core. A hijack.

If AI slop is the argument for doing less work, DreamSTer is the argument for doing insane amounts of it. Released as "r1" on July 17, 2026 by skmp, a developer previously known for the Reicast Dreamcast emulator, DreamSTer brings experimental Sega Dreamcast support to hardware most people assumed could never touch it. The catch is in how. The project describes itself as "polly2-rtl, minicast and a python TUI rolled together," and warns that it is "very experimental and does not follow MiSTer conventions" and "is not a MiSTer core." It splits the console in half: the ARM CPU on the DE10-Nano runs the Dreamcast's main processor in software, while an FPGA implementation of the PowerVR2 handles the graphics.

# DreamSTer is NOT a MiSTer core. It hijacks the fabric:
1. kill the running MiSTer process
2. take over the FPGA (load the PowerVR2 RTL)
3. launch minicast (ARM-side Dreamcast emulation)
4. on exit: reload menu.rbf, restart MiSTer

Why it works and why it will not be standard

The mechanism is gloriously unconventional. Per GenerationAmiga, DreamSTer kills the running MiSTer process, seizes the FPGA fabric to load its PowerVR2 logic, launches its minicast emulator on the ARM side, and then reloads menu.rbf and restarts MiSTer when you quit. It treats the DE10-Nano as two computers taped together rather than one accuracy machine. Compatibility, performance, and HDMI output are all limited, and by the author's own framing this is not a MiSTer core and does not pretend to be one. It is proof that the "impossible" is merely difficult, which, on this platform, is a distinction the forum takes very seriously.

The pattern: ambition outside the rulebook

DreamSTer and the 3DO core are two halves of the same 2026 mood. Both take on consoles long considered off-limits; both are honest that they bend or break the platform's own accuracy standards; and both landed first outside the forum's tidy per-system boards. The frontier of MiSTer is increasingly a place where "is not a MiSTer core" is a feature, not a disclaimer.

RetroAchievements, Bolted On by a Fork

Odelot forks the main binary

Not every 2026 story is a fight. In April, developer Odelot forked MiSTer-devel/Main_MiSTer, the core Linux-side binary that drives the whole system, to add RetroAchievements support, and the community reaction, per Time Extension, was "Wow, That Was Fast." The fork bolts an achievements layer on top of roughly a dozen cores: NES, SNES, Genesis, Master System, Game Boy, N64, PlayStation, GBA, Mega CD, NeoGeo, PC Engine, Atari 2600, and Sega 32X. Hardcore mode, the no-save-state-cheating tier, was verified first on SNES, Genesis/Mega Drive, N64, and PlayStation.

How it reads the cheevos off real cores

The clever part is that it never touches the cores themselves. The forked binary computes an MD5 hash of the loaded game, checks it against the RetroAchievements database, and then reads the emulated RAM straight out of the FPGA core, polling roughly every frame at about 60Hz to evaluate achievement conditions against live game state. Load a standard community core that lacks the achievement mirror module and the whole apparatus simply goes quiet, with no penalties and no interference. It is a bolt-on that respects the thing it is bolted to.

Kept compatible with the updater

Because forking the main binary could easily break the normal update flow, a second developer, manyhats-mike, published a deploy tool specifically to swap Odelot's build in and out without breaking update_all compatibility, and MiSTer Companion added its own achievements management in v3.8.0 on May 2, 2026. This is the forum ecosystem at its best: one person forks, another makes the fork safe to install, and a third wires it into the tooling everyone already uses.

From MiST to Sorgelig: A June 2017 Origin

Melnikov's port that outgrew its parent

None of this exists without one developer and one decision. Per Wikipedia, the MiSTer project was created by Alexey "Sorgelig" Melnikov and introduced on GitHub in June 2017 as a port of Till Harbaum's earlier MiST, an analog-only FPGA board aimed at the Amiga and Atari ST. The pivotal call was hardware: rather than commission a custom board, Melnikov targeted the mass-produced Terasic DE10-Nano, an Intel Cyclone V system-on-chip development board, which brought HDMI and, crucially, an off-the-shelf supply chain. The Verge was already covering the project as a serious preservation platform by 2021.

Why this crowd chose gates over software

The appeal is not nostalgia, it is fidelity. As Kotaku's Mike Fahey put it, "the MiSTer is the right-damn-now of retro gaming," and "By many metrics, FPGA hardware simulation surpassed software emulation a few years back." The distinction he draws is the whole philosophy of the platform: "Instead of tricking a game ROM with software, you can program an FPGA to more or less be the original console hardware." That is exactly why the AI-slop fight cuts so deep. An FPGA description of MAME abandons the one thing that justified buying the board.

The clone era arrives

Success brought imitators. By 2024 the YouTuber Taki Udon had launched Retro Remake's MiSTer Pi, a low-cost Cyclone V clone of the DE10-Nano, and the forum's own hardware board now openly discusses third-party boards from Retro Remake, Heber, and QMTech alongside the original. The reference platform is no longer the only game in town, which is healthy, and, as we will see, complicated.

The Hardware Orbiting the Forum

The DE10-Nano tax

The forum runs on a specific and increasingly expensive assumption: that you own a DE10-Nano. A bare Terasic board, Cyclone V SE with a dual-core ARM Cortex-A9 and roughly 110,000 logic elements, runs about $225, and everything else is a shell around it. Ultimate MiSTer's PRO bundle, built on a genuine "no clones" DE10-Nano, lists at €349.90; the MiSTer MultiSystem Kit, which pointedly does not include the board, is €309.90 for the case, SDRAM, RTC, and wireless; and consolized ZERO and RED variants push toward €599.90. We break the full board-and-kit math down separately in our look at the DE10 hub and its €309 to €599 kits.

The all-in-ones that ditched the sandwich

The most interesting hardware trend is the move away from the DE10-Nano stack entirely. Heber's MiSTer Multisystem 2 solders a Cyclone V directly to a custom board, with no DE10 sandwich and no RAM modules, and sells for £216 in its digital form and £264 for the analogue model with SCART, VGA, and SNAC. Reviewer Lon Seidman described it as a device that "bridges the gap between the technical flexibility of the original FPGA development boards and the convenience of a dedicated home console." Hackaday put the same idea more plainly: the Multisystem finally gets MiSTer "out of the tinker cave and into the living room," and "you don't need to be a soldering wizard" to use it.

The boards the forum actually names

This matters to the forum because hardware fragmentation is now part of the conversation the moderators have to manage. The board discusses Retro Remake's MiSTer Pi, Heber's Multisystem line, and QMTech's Cyclone V boards as first-class options, a very different world from the closed, single-SKU approach of rivals like Analogue, whose closed-box FPGA Analogue 3D ships upscaling in a sealed console you cannot reprogram (Engadget's Analogue Pocket review praised that polish while noting the curation is the point). MiSTer's openness is its edge over the polished-but-locked competition, the same way a tinkerer's board will always beat a $90 Miyoo Mini Plus full of ROMs or an Android handheld like the Retroid Pocket line on fidelity, if never on price or convenience.

Forum vs Discord vs GitHub: Where the Work Actually Happens

Discord did the heavy lifting in 2026

Here is the uncomfortable truth the 2026 timeline keeps repeating: the forum is increasingly the archive, not the workshop. The 3DO beta dropped on Discord. DreamSTer lives on GitHub. The AI-slop fight went public on X before it was ever a moderated forum policy. The community's own documentation now frames the split explicitly: Discord for live help, updates, and the "very active community," the forum for what it calls classic forum discussions. In practice that means the forum is where things are recorded and litigated after the fact, while the actual releases increasingly happen in chat.

GitHub is the source of truth

The center of gravity for anything permanent is the code host. Compiling for MiSTer means the MiSTer-devel/Main_MiSTer repository, with Linux and WSL build instructions on the official docs site, and a git mirror at eaw.app described itself in March 2026 as covering "all MiSTer FPGA retro computing cores and infrastructure." When Odelot forked the main binary and skmp published DreamSTer, they published to GitHub, and the forum linked to it. The repo is canonical; the forum is commentary.

Fragmentation and the Feedspot footnote

MiSTer discussion is also not confined to one board. A dedicated MiSTer subforum lives on the Atari-Forum, and aggregators reflect the diffusion: Feedspot ranks the MiSTer FPGA forum only ninth on its list of fifteen FPGA forums, slotting it among general-purpose boards like the Intel FPGA community and Red Pitaya. For the single most active retro-hardware community on the planet to rank mid-table on a generic FPGA list tells you how niche this all still is, and how much of the conversation now happens somewhere a crawler cannot easily count.

Predictions: The Next 6 to 12 Months

Moderation catches up to the AI fight

The "segregate or ban" thread is not going away, and phpBB gives moderators exactly one obvious tool: a subforum. Expect a formal AI-generated-cores label or a quarantined board by early 2027, precisely because the alternative, case-by-case deletion of cores that technically work, is unenforceable. The purists will not get a ban; they will get a containment zone, and both sides will call it a loss.

The cores that will and will not stabilize

Concrete calls. srg320's 3DO core will reach a genuinely playable stable release within the window, still requiring dual SDRAM and still glitching on VRAM-heavy edge cases exactly as he warned; "good enough" will be the official verdict, and it will be honest. DreamSTer will improve its compatibility and HDMI output but will not be folded into the standard core repo inside twelve months, because "is not a MiSTer core" is a design stance, not a temporary state. RetroAchievements support, by contrast, is the one most likely to be effectively mainlined, either upstreamed or blessed as the semi-official companion path, with more cores earning hardcore verification.

Membership, hardware, and where betas drop

The board will clear 12,000 members before the peak-online record gets seriously challenged again, because lurkers register faster than they log in simultaneously. Third-party hardware from Retro Remake, Heber, and QMTech will keep multiplying as DE10-Nano pricing and supply stay under pressure, further eroding the "you must own a DE10" assumption. And the betas will keep dropping on Discord first, keeping the forum in its 2026 role: the definitive changelog of a project whose living edge has moved to chat.

The Verdict: A Town Square Under Strain

The archive holds; the action moved

The MiSTer forum in 2026 is simultaneously healthier and less central than ever. Healthier, because 10,690 members, 106,431 posts, and a 42,092-user spike are not the vitals of a dying community. Less central, because every headline development this year, the 3DO core, DreamSTer, the RetroAchievements fork, the AI-slop revolt, either broke on Discord and GitHub first or was decided on X. The forum has become the place the community goes to write history down, not to make it.

Accuracy versus availability, unresolved

Underneath every 2026 story is one unresolved question: what is MiSTer for? srg320 answers "accuracy, honestly labeled." Jotego answers "accuracy or nothing." The AI-core crowd answers "availability, now." DreamSTer answers "whatever works, conventions be damned." The forum cannot referee that fight because it is not really a technical dispute; it is an argument about the soul of the project, and those do not get closed threads. What the board can still do, better than any Discord or repo, is keep the receipts. In a hobby built on the premise that the past deserves to be reproduced exactly, that may be the most fitting job of all.