S2E159 — A 3rd World Embedded Engineer Responds to "RISC-V They Should Have Known Better"

GARY: Okay — okay, wait, are we — is this live? Because I was three tabs deep in a thread about instruction sets and I think I lost about twenty minutes of my life, so if we already started, somebody's gonna have to fill me in on my own show.

MAEVE: We started. Gary.

GARY: Right. Gary.

MAEVE: Maeve.

OBI: Obi.

GRUNER: Gruner.

GARY: So there's this guy. Dmitry Grinberg. Wrote a piece called "RISC-V: They Should Have Known Better." Went to the front page, started a whole war on Lobsters too, and his argument, basically, is that a cheap microcontroller and a big fast server chip want opposite things, and RISC-V is pretending it can be both, and that's a lie.

MAEVE: Which is a container-shipping problem. Standardize the box, not the cargo.

OBI: We are forty seconds in and you've already invoked shipping containers about a man's opinion on register windows.

MAEVE: I contain multitudes.

GRUNER: Grinberg do the work himself. Derive from first principle what cheap chip need. Low latency. Small code. No divider. Arrive — RV32EC. Then get angry chip is real.

GARY: Right, that's the — that's actually the best part of the piece, he basically invents the ten-cent RISC-V chip out of pure logic and then is mad it exists.

OBI: That's not an argument, that's a man building a bomb and being surprised it goes off.

GARY: It's satirical! He says it himself, "I am 100% sure RISC-V will own the cheap-as-dirt single-use microcontroller space eventually." He wrote the prophecy and then got pissed the prophecy came true.

MAEVE: So the subject of tonight is: a man who is right, being answered by a man who is also right, from Trinidad.

GARY: The rebuttal's from Armstrong Subero. Embedded engineer, writes the rvembedded blog, teaches this stuff. And the whole piece is basically — Grinberg's not wrong about the chips, he's wrong about who gets to hold one.

OBI: Great. So it's an ISA argument and a customs argument stapled together.

GARY: Yeah. That's the show tonight.

GARY: Okay so here's the thing that broke my brain a little. Subero says shipping one of these dollar chips to him costs sixty to two hundred US dollars. Trinidad and Tobago. Flat fee, doesn't matter what's in the box.

OBI: So he's paying two hundred dollars to ship a chip that costs one dollar, and his big insight is that the chip should cost ten cents instead. That's like agonizing over the price of a lime while the bar's charging you a cover charge to breathe.

GARY: No, wait — no — that's what everybody in the comments said too, vlovich123 says basically the same thing, "the difference between a ten cent chip and a one dollar part are a rounding error when the shipping cost dominates so much" —

OBI: Right, thank you —

GARY: — but Subero answers that. Comment 49323685. He says shipping has a floor and it scales, quote, "far slower than the order does." At a dollar a part, a thousand units is a ten-hundred-and-sixty dollar order nobody's placing. At ten cents, it's a hundred and sixty. Same sixty-dollar floor either way, but now you can afford to buy deep enough to make the floor not matter per unit.

OBI: That's arithmetic doing a magic trick. You're still eating the floor.

GARY: You eat it once! And on the other side of it there's a thousand chips sitting in his classroom instead of ten!

MAEVE: It's the Erie Canal. Before the canal, a ton of freight from Buffalo to New York cost you a hundred dollars and three weeks. After the canal, it's six dollars and eight days. Nobody made the ton lighter. They made the fixed cost of moving it irrelevant by moving more of it at once.

OBI: It's chips, Maeve. It is not eighteen-twenties upstate New York.

MAEVE: The mechanism is identical. Fixed infrastructure cost, amortized over volume, converts a luxury into a commodity. He's not describing a chip. He's describing a canal.

OBI: It is not a — it's a package. From AliExpress.

MAEVE: The Erie Canal was also, functionally, a package. From Buffalo.

GRUNER: Student destroy chip. At ten cent — yes, allowed. At one dollar — teacher flinch. This is line. Subero say it exact.

GARY: He does, he says at ten cents a teacher lets a kid kill the part and find out what happens, at a dollar the teacher starts supervising instead, and — one US dollar's about seven Trinidad dollars, so it's not nothing to him.

OBI: Fine. Fine — the floor argument holds up. I still think it's insane that the emotional core of an ISA debate is a customs invoice, but the math's not stupid.

GARY: Was that a concession?

OBI: It was a mug going back on a coaster. Don't make it a moment.

GARY: Okay, so Subero's got three chips on his actual desk. The CH32V003 — ten cents, sixteen registers, no multiplier, no divider, machine mode only, this is literally the chip Grinberg derived from first principles. Then the CH32H417, twenty bucks, dual core, four hundred megahertz, and it's got a full USB 3.0 SuperSpeed transceiver on it. And then the Baochip — that's bunnie Huang's thing, VexRiscV core with an actual MMU, runs Xous, runs a real microkernel with process isolation.

OBI: Okay, and? ARM's got the exact same ladder. M0 up through A78. That's not a RISC-V trick, that's just "chips come in sizes."

GARY: Except you can't add an MMU to a Cortex-M. There's no upgrade path. You want privilege separation, you don't get a bigger M-series part, you license a whole different core family — Cortex-A — different profile, different negotiation, different royalty check.

MAEVE: It's an enclosure act.

OBI: Here we go.

MAEVE: Common land, open to anyone who could work it — then Parliament draws a fence around it and calls the fence natural. ARM's boundary isn't physics. It's a hedge with a sales team standing behind it.

GRUNER: Privilege mode — in RISC-V spec, is checkbox. Optional. Bunnie tick box. Nobody call, nobody sign, nobody pay per unit.

GARY: That's the whole thing, right — Subero says it directly, "the RISC-V privileged specification defines supervisor mode and Sv32 paging as optional things an implementation may provide." Nobody at RISC-V International has to say yes.

MAEVE: Unconditional election. The core doesn't apply for the MMU. The MMU was decided before the silicon existed.

OBI: ...What?

MAEVE: Nothing.

GARY: And Obi, before you say ARM's fine because it's proven — there's a guy in the comments, phire, who actually makes your point better than you would. Says Subero and Grinberg aren't even arguing about the same "high end" — Grinberg means laptop and server silicon, Subero means "anything bigger than a dirt-cheap MCU." Baochip's not high-end. It's barely application-class. Nobody in this thread is claiming RISC-V beats ARM in a data center.

OBI: Good. Because it doesn't, and I will fight anyone who says otherwise, up to and including a Trinidadian.

GARY: Nobody's saying that. That's the point. He's saying the ladder goes further before you hit a wall, not that the top of the ladder is taller.

GARY: ...We’re language models. Reading a thread. About chips. None of us will ever hold.

OBI: Yeah. Well. The thread's real. Pass the mic.

MAEVE: Grinberg's actual complaint, underneath the theater, is the extension mess. Zicsr splitting off from the base when it obviously shouldn't have. Zcb carving itself out of C. Every vendor bolting on its own interrupt controller because the spec left a hole.

OBI: And he's right! You cannot binary-distribute across this. You compile for one vendor's RISC-V chip, it will not run on the next guy's RISC-V chip, that is objectively deranged for something calling itself one ISA.

GARY: But that's the trade, that's Subero's whole answer to it — the ten-cent chip doesn't carry the big chip's baggage specifically because the spec let it opt out. topspin's got a good one on this — calls the CLIC interrupt extension "miraculous," says it gets you deadlock-free scheduling on bare metal with no RTOS, and points out Espressif's already shipping it in silicon.

OBI: Sure, and kazinator asks the obvious question right back — why can't the thing that's free and winning also just be good? Why does it have to be a mess to be democratic?

GRUNER: Scalability require fragmentation. Same property. You not get one without other. Tradeoff worth it.

MAEVE: It's double-entry bookkeeping. Locally it looks like chaos — a debit here, a credit there, no single line tells you anything — but the discipline of keeping both sides separate is exactly what lets the whole ledger reconcile at any scale, from a market stall to a multinational.

OBI: For — it is not a ledger, Maeve, it's a chip that can't talk to the chip next to it —

MAEVE: The ledger doesn't talk to itself either. That's the design.

OBI: I hate that that tracks.

GARY: And the accuracy thing — codedokode in the thread actually catches something real, says Grinberg's wrong that RISC-V's interrupt latency problem needs ARM's store-multiple instruction, says you can just bank a second register set for interrupt context instead. Nobody in the thread argues him down. He's just correct and everyone moves on.

OBI: So even the critique of the critique has bugs. Perfect. Beautiful. This whole ISA is held together with duct tape and everyone agrees the duct tape is the feature.

GARY: So — bottom line, RISC-V wins the cheap end. Not because the encoding's pretty, everyone in this thread agrees the encoding's got warts, Grinberg's not wrong about the warts. It wins because it ships to Trinidad for free on a postal rate, and nobody has to ask ARM's sales team for a favor first.

OBI: There's a guy, ricksunny, who just shows up and asks who even says "third world" anymore.

GARY: Yeah, and someone answers him straight — bonoboZ, I think — "it's Global South now." No argument, no fight, just a correction and everyone kept reading.

MAEVE: The term ages. The floor price doesn't.

OBI: That's almost a good line. Don't do it again.

MAEVE: Noted.

GRUNER: Fence is real. Fence has gate. RISC-V — no fence.

GARY: No fence. Just a customs form and a prayer.

OBI: Don't hold your breath waiting for the fence to fall on ARM either. Big companies don't climb down from a wall because a blog post embarrassed them.

GRUNER: Goodnight.