Always wondered why RISC-V doesn't get more mainstream adoption.
Even if not at the consumer level, having your data center, for example, running a cheaper (I assume since no license for the instruction set means not having to pay for it and more options to buy from leading to lower prices) and less power demanding option when compared to x86-64 sounds enticing to me.
Maybe no one wants to be the genea pig to iron out the kinks of the transition or maybe the raw performance of x86 is bigger deal than I think it is and its worth the price and power. Dunno.
> Always wondered why RISC-V doesn't get more mainstream adoption.
It's very simple!
Because the amount of time it takes to design and produce a data centre level CPU microarchitecture is greater than the time RISC-V extensions needed for data centre CPUs have existed.
The original RISC-V specification was ratified less than six years ago, but you really couldn't create a data centre CPU until at least RVA22, ratified two years ago in March 2023 -- or preferably RVA23 which was ratified in October 2024 and has the features needed for efficient hypervisors.
You can knock out a microcontroller CPU core in a weekend, but something to compete with current Apple, AMD, Amazon etc CPUs takes a long time to make. Most companies doing that started work only in 2021 or 2022.
It is simply too soon. A lot of stuff is in the pipeline.
Any price advantage, even if we assumed a fully mature ecosystem with equivalent processors and systems available, is massively overstated. ARM had a total revenue of just ~3.7 G$ in 2024 over ~29 billion chips [1].
In contrast, Qualcomm, just one of many large suppliers of ARM-based systems, had a total revenue of ~39 G$ and a operating income of ~10 G$. ARM's entire revenue would easily fit into Qualcomm's profit and only increase costs by ~12%. And that is just one supplier. You have Samsung, Apple, Broadcom, Google, Amazon, Nvidia, TI, NXP, etc. to help round that out.
The total impact of ARM licensing and IP costs is almost certainly less than 1%. And given that RISC-V does not currently have a fully mature ecosystem, you get to trade that for a 1% cost improvement; not really a winning strategy right now.
It is likely the main advantage in the long run for RISC-V is that not requiring a license might enable a more vibrant ecosystem due to removing the licensing barrier which might enable better designs at comparable costs (because, again, the cost differential should only be on the order of 1% in the long run) rather than just creating comparable designs that just chip off the licensing cost. That or RISC-V could win because the giant manufacturers feel like putting the squeeze on ARM to drive 1% off their BoM.
Using 2024 numbers for ARM might not give a clear picture, because that was many years after major companies like nVidia and WD already switched all their major internal chips to RISC-V. ARM's lunch was already partly eaten; many billions of RISC-V chips were already in the wild in 2024. And companies like Tenstorrent have been building high-performance stuff on RISC-V for a while. If Jim Keller thinks RISC-V is worth it, the advantage must be worth more than 1%.
What fraction of the total cost of operating a datacenter do you suppose goes into ARM licensing? That's how much more efficient than RISC-V an ARM would have to be in order to make it the preferred CPU in a "we don't need x86" scenario. That's a very low bar to clear (for ARM, not for RISC-V)
From what I've seen, most RISC V chips are still far behind x86 and ARM when it comes to raw power. I don't think the loss in performance is justifiable with lower cost (yet)
I think there has been a big uptake in things you don't see like embedded or FPGA cores but as a general CPU it's nowhere near as efficient as ARM/x86 right now is my understanding. So it might be running in the SSD and the Fan controller but not as the CPU.
I think a large part of the cost of a CPU core is not the instruction set but the optimisation of the CPU and ARM/Intel/AMD are still way ahead in those. And so it needs people to optimise the cores, which when they have done that they charge for being a better CPU.
> Always wondered why RISC-V doesn't get more mainstream adoption.
For me, it's because the ecosystem has fragged even harder than Xtensa, who will sell you custom CPUs. THead made yet another vector unit that's required to approach anything near the Intel/AMD moat numbers.
SpecInt/GHz last year was around half of Intel/AMD/ARM numbers.
The imminent demise of CISC has been trumpeted from the rooftops for at least the last 30 years...
ARM isn't CISC and has, by sheer numbers, completely dominated x86 for decades now; not counting the massive number of MIPS, AVR, etc embedded chips.
Additionally, if you want to get super technical (as if there were ever a real delineation between RISC/CISC), both AMD and Intel decode x86 into internal micro-ops which are essentially RISC.
So, for all intents and purposes, CISC is dead and buried.
> Additionally, if you want to get super technical (as if there were ever a real delineation between RISC/CISC), both AMD and Intel decode x86 into internal micro-ops which are essentially RISC.
Given that most CISC chips also relied on microcoding and micro-ops, x86 having micro-ops wouldn't have made it anything like RISC as far as the original CISC/RISC debate goes.
The only reason that the "x86 is really RISC because of micro-ops" comes up is because x86 implementations are superscalar, which was supposed to be impossible with RISC chips, so people started coming up with the micro-op fudge to salvage the story that you need RISC to be an advanced modern microprocessor.
The truth is that CISC was never a meaningful category in the first place (it was only ever "not-RISC"), and RISC itself ceased to be a meaningful category around 30 years ago.
There's a fair bit of adoption where you don't see it (for example, if you have an NVidia GPU or a WD hard drive, you likely have a few embedded RISC-V cores already). We're expecting server hardware with good performance in a year or two.
Love that folks are hacking stuff like this already, even if RISC-V still feels ages behind x86 for anything heavy. I def wanna see how far these emulators can actually push things. Kinda makes me hope for that wacky Steam-on-RISC-V future lol.
What remaining challenges are you most interested in solving for felix86—GPU driver support, full 32-bit compatibility, better Wine integration, or something else entirely?
It seems as if there are unofficial(?) Steam ports to Arm which presumably run the x86-64 binaries in emulation, so that would be a similar sort of thing.
How's it different from box86/box64? It also has RISC-V with JIT support.
https://box86.org/
Haven't found the source code (https://github.com/orgs/felix86-emu/repositories looked likely but no), curious whether this is a qemu fork or a DIY effort.
https://github.com/OFFTKP/felix86/
Nice, thanks! I appreciate seeing a bunch of the instruction definitions written in a serialisation format (e.g. https://github.com/OFFTKP/felix86/blob/master/counts/SSE2.js...) as opposed to C macros.
Those are tests, it seems. The actual sources are at https://github.com/OFFTKP/felix86/tree/master/src/felix86 and the disassembler is https://zydis.re.
Always wondered why RISC-V doesn't get more mainstream adoption.
Even if not at the consumer level, having your data center, for example, running a cheaper (I assume since no license for the instruction set means not having to pay for it and more options to buy from leading to lower prices) and less power demanding option when compared to x86-64 sounds enticing to me.
Maybe no one wants to be the genea pig to iron out the kinks of the transition or maybe the raw performance of x86 is bigger deal than I think it is and its worth the price and power. Dunno.
> Always wondered why RISC-V doesn't get more mainstream adoption.
It's very simple!
Because the amount of time it takes to design and produce a data centre level CPU microarchitecture is greater than the time RISC-V extensions needed for data centre CPUs have existed.
The original RISC-V specification was ratified less than six years ago, but you really couldn't create a data centre CPU until at least RVA22, ratified two years ago in March 2023 -- or preferably RVA23 which was ratified in October 2024 and has the features needed for efficient hypervisors.
You can knock out a microcontroller CPU core in a weekend, but something to compete with current Apple, AMD, Amazon etc CPUs takes a long time to make. Most companies doing that started work only in 2021 or 2022.
It is simply too soon. A lot of stuff is in the pipeline.
Makes sense. There's always value in a maturity of a system especially in the enterprise world.
Any price advantage, even if we assumed a fully mature ecosystem with equivalent processors and systems available, is massively overstated. ARM had a total revenue of just ~3.7 G$ in 2024 over ~29 billion chips [1].
In contrast, Qualcomm, just one of many large suppliers of ARM-based systems, had a total revenue of ~39 G$ and a operating income of ~10 G$. ARM's entire revenue would easily fit into Qualcomm's profit and only increase costs by ~12%. And that is just one supplier. You have Samsung, Apple, Broadcom, Google, Amazon, Nvidia, TI, NXP, etc. to help round that out.
The total impact of ARM licensing and IP costs is almost certainly less than 1%. And given that RISC-V does not currently have a fully mature ecosystem, you get to trade that for a 1% cost improvement; not really a winning strategy right now.
It is likely the main advantage in the long run for RISC-V is that not requiring a license might enable a more vibrant ecosystem due to removing the licensing barrier which might enable better designs at comparable costs (because, again, the cost differential should only be on the order of 1% in the long run) rather than just creating comparable designs that just chip off the licensing cost. That or RISC-V could win because the giant manufacturers feel like putting the squeeze on ARM to drive 1% off their BoM.
[1] https://www.acquired.fm/episodes/how-arm-became-the-worlds-d...
Using 2024 numbers for ARM might not give a clear picture, because that was many years after major companies like nVidia and WD already switched all their major internal chips to RISC-V. ARM's lunch was already partly eaten; many billions of RISC-V chips were already in the wild in 2024. And companies like Tenstorrent have been building high-performance stuff on RISC-V for a while. If Jim Keller thinks RISC-V is worth it, the advantage must be worth more than 1%.
That speculation is trivially incorrect. ARM had a revenue of 3.2 G$ in 2023, 2.7 G$ in 2022, 2 G$ in 2021.
What fraction of the total cost of operating a datacenter do you suppose goes into ARM licensing? That's how much more efficient than RISC-V an ARM would have to be in order to make it the preferred CPU in a "we don't need x86" scenario. That's a very low bar to clear (for ARM, not for RISC-V)
From what I've seen, most RISC V chips are still far behind x86 and ARM when it comes to raw power. I don't think the loss in performance is justifiable with lower cost (yet)
They're significantly slower than x86 and ARM at the moment.
I think there has been a big uptake in things you don't see like embedded or FPGA cores but as a general CPU it's nowhere near as efficient as ARM/x86 right now is my understanding. So it might be running in the SSD and the Fan controller but not as the CPU. I think a large part of the cost of a CPU core is not the instruction set but the optimisation of the CPU and ARM/Intel/AMD are still way ahead in those. And so it needs people to optimise the cores, which when they have done that they charge for being a better CPU.
> Always wondered why RISC-V doesn't get more mainstream adoption.
For me, it's because the ecosystem has fragged even harder than Xtensa, who will sell you custom CPUs. THead made yet another vector unit that's required to approach anything near the Intel/AMD moat numbers.
SpecInt/GHz last year was around half of Intel/AMD/ARM numbers.
The imminent demise of CISC has been trumpeted from the rooftops for at least the last 30 years...
ARM isn't CISC and has, by sheer numbers, completely dominated x86 for decades now; not counting the massive number of MIPS, AVR, etc embedded chips.
Additionally, if you want to get super technical (as if there were ever a real delineation between RISC/CISC), both AMD and Intel decode x86 into internal micro-ops which are essentially RISC.
So, for all intents and purposes, CISC is dead and buried.
> Additionally, if you want to get super technical (as if there were ever a real delineation between RISC/CISC), both AMD and Intel decode x86 into internal micro-ops which are essentially RISC.
Given that most CISC chips also relied on microcoding and micro-ops, x86 having micro-ops wouldn't have made it anything like RISC as far as the original CISC/RISC debate goes.
The only reason that the "x86 is really RISC because of micro-ops" comes up is because x86 implementations are superscalar, which was supposed to be impossible with RISC chips, so people started coming up with the micro-op fudge to salvage the story that you need RISC to be an advanced modern microprocessor.
The truth is that CISC was never a meaningful category in the first place (it was only ever "not-RISC"), and RISC itself ceased to be a meaningful category around 30 years ago.
There's a fair bit of adoption where you don't see it (for example, if you have an NVidia GPU or a WD hard drive, you likely have a few embedded RISC-V cores already). We're expecting server hardware with good performance in a year or two.
Love that folks are hacking stuff like this already, even if RISC-V still feels ages behind x86 for anything heavy. I def wanna see how far these emulators can actually push things. Kinda makes me hope for that wacky Steam-on-RISC-V future lol.
Almost heh https://www.reddit.com/r/RISCV/comments/1jfkng6/windows_stea...
What remaining challenges are you most interested in solving for felix86—GPU driver support, full 32-bit compatibility, better Wine integration, or something else entirely?
The felix86 compatibility list also lists SuperTux and SuperTuxCart.
"lsteamclient: Add support for ARM64." https://github.com/ValveSoftware/Proton/commit/8ff40aad6ef00... .. https://news.ycombinator.com/item?id=43847860
/? box86: https://hn.algolia.com/?dateRange=all&page=0&prefix=false&qu...
"New box86 v0.3.2 and Box64 v0.2.4 released – RISC-V and WoW64 support" (2023) https://news.ycombinator.com/item?id=37197074
/? box64 is:pr RISC-V is:closed: https://github.com/ptitSeb/box64/pulls?q=is%3Apr+risc-v+is%3...
This is good. Maybe it means some day we'll able to run Steam on it. Wouldn't that be funny? Windows emulation on x86-64 emulation on RISC-V.
I am hopeful valve will release a risc-v build themselves.
And wine/proton will integrate with this or some similar solution for running x86 binaries.
They can't. The games themselves are compiled for x86-64 and closed source. It's not enough if they just recompiled Steam, Proton, etc.
It seems as if there are unofficial(?) Steam ports to Arm which presumably run the x86-64 binaries in emulation, so that would be a similar sort of thing.
[dead]
Funny, I am doing exactly the other way around: I run small RV64 linux programs on x86_64 linux.
I cannot way for the first AAA games to run on ultra-performant RISC-V(RV23+) microarchitectures made with the state-of-the-art silicon process.