RISC-V AI Chips Will Be Everywhere (2022)(spectrum.ieee.org) |
RISC-V AI Chips Will Be Everywhere (2022)(spectrum.ieee.org) |
Part of the strategy of commoditizing your complement: https://gwern.net/complement
Similar thing going on right now with the rampant open sourcing of LLMs
Eventually RISC-V may have CPU cores fast enough to compete with high end ARM and x86 cores but it will take a while and incentive from the chip companies to design those high performance RISC-V cores. Some of that incentive may come from ARM raising license fees.
> Eventually RISC-V may have CPU cores fast enough to compete with high end ARM and x86 cores but it will take a while
Multiple companies with as good credentials as Nuvia have been working on similar level RISC-V cores since about 2022. Nuvia was founded in 2019, the Snapdragon X Elite chip using their core is out now. Expect RISC-V to take the same amount of time ... i.e. probably hit the market in 2027.
Is that "a while"? It's basically tomorrow in chip and software development terms.
Short term we won't see it "everywhere". Without a powerhouse to push it, I find it hard to see/say if it'll be able to catch up or what that'll look like (see Linux example).
I don't think that goal has yet been achieved. Instead, as others point out, we've seen a plethora of offerings from cheap vendors dodging the ARM "monopoly".
In the AI space, established vendors like Nvidia & AMD have no incentive to overhaul their entire stack to RISC-V. It's a huge investment and risk for unknown returns.
Disruption will have to come from outside, but it'll be a while before we see Nvidia dethroned.
RISC-V doesn't have any structural issues that would mean it can't succeed. In fact it already is very successful, but it's mostly in embedded and invisible CPUs at the moment which is why it might not seem like it.
It definitely needs some time to mature for application class processors but I can't see any reason why it won't.
A ton of instructions needed for more serious things -- vectors, hypervisor, crypto, cache control operations, more -- were published in December 2021, 2 1/2 years ago. The first chips with that spec (called RVA22) have been hitting the market in the last few months. An early Core i7 level chip with RVA22 will come out late this year, a Pentium III level (but 8 core) one is out now in the BPI-F3 and soon Milk-V Jupiter (MiniITX board).
A lot more almost as important instructions have been published since then. The RVA23 spec is about to be ratified in the next few months. That will be the likely LTS baseline for Android and other mobile and desktop OSes that you will see around 2026 or 2027, just as (hopefully) desktop-class (at least similar to Snapdragon X Elite) chips come out.
How long has RISC-V been around? Very little time. As of right now, it's four years behind Arm, with the gap closing.
Maybe you can come back to this comment in five or ten years and laugh at me, but right now it feels like RISC-V is a solution in search of a problem.
1. There's no big ecosystem of compiled ARM code to prevent people moving to RISC-V. In the embedded world everything is always compiled from source so it doesn't matter. On Android everything except games is bytecode. Maybe on Windows/Mac... but that's probably the last place we'll see RISC-V.
2. ARM and RISC-V are much more similar than ARM and x86 so it is way easier to translate ARM to RISC-V. It's also much more extensible so you can easily add instructions that you need to make the translation easy. I believe ARM has some instructions that are specifically included to make running x86 code easier - that sort of thing would be even easier on RISC-V.
You're probably right about price being the major differentiator, but I wouldn't underestimate that! Especially with ARM turning the screws on license fees.
Windows supported ARM three years before RISC-V even existed.
ARM has a headstart of dozens of years. That RISC-V is even hinting at becoming a competitor is huge.
WCH makes a microcontroller that sells for around 10 cents; it's cheaper than a 7400 quad gate, so it's bound to end up in a ton of things. It occurred to me that they are like electric motors: unglamorous but ubiquitous (there are several dozen electric motors, mostly small ones, in the room I am sitting in right now)
On Digikey the cheapest I can find SOIC14 7400s is 20c each, but you have to buy 1480 of them to get that. If you want just a few they're $1.60 each, and if you want DIP14 they're $2.
The propagation delay of using a microcontroller to implement a quad NAND gate will be a lot higher than the 7400's 14ns of course. At a wild guess I'd say 200ns or greater. Could be 1us. I don't think more than that. That's still fine for many uses.
For those who don't know, a CH32V003 is a 32 bit RISC-V CPU implementing the RV32EC instruction set (basic integer instructions, 16 registers, 2-byte instructions available for the most common operations, as well as the standard 4-byte instructions, to save 25%-30% program space. It has 2048 bytes of RAM and 16k of flash memory to hold your program. A program to emulate a 7400 would use 0 of the RAM and maybe 100 bytes of the flash (most of it would be init code, run once at power-on).