but literally beating the flagship desktop chips in single-core performance
See, this is what I despise about x86. AFAIK it’s literally RISC on the bare metal but there are hundreds of “instructions” running microcode which is basically just a translation layer. You’re not allowed to write code for the actual RISC implementation because that’s a trade secret or something. So obviously single core performance would be shit because you’re basically running an emulator all the time.
RISC-V can’t come fast enough. Maybe someone will even make a chip that’s RISC-V but with the same instruction/microcode support as x86. So you can run RISC-V code directly or do the microcode thing and pretend you’re on x86. Though that would probably get the shit sued out of them by Intel because god forbid there’s actual innovation that the original creator can’t cash in on.
RISCV would be a huge step forward, and there are projects like this one working on making a high performance architecture using it. But I’d argue that we should really be rethinking the way we do programming as well.
The problem goes deeper than just the translation layer because modern chips are still contorting themselves to maintain a fiction for a legacy architecture. We are basically burning silicon and electricity to pretend that modern hardware acts like a PDP-11 from the 1970s because that is what C expects. C assumes a serial abstract machine where one thing happens after another in a flat memory space, but real hardware hasn’t worked that way in decades. To bridge that gap, modern processors have to implement insane amounts of instruction level parallelism just to keep the execution units busy.
This obsession with pretending to be a simple serial machine also causes security nightmares like Meltdown and Spectre. When the processor speculates past an access check and guesses wrong, it throws the work away, but that discarded work leaves side effects in the cache that attackers can measure. It’s a massive security liability introduced solely to let programmers believe they are writing low level code when they are actually writing for a legacy abstraction. on top of that, you have things like the register rename engine, which is a huge consumer of power and die area, running constantly to manage dependencies in scalar code. If we could actually code for the hardware, like how GPUs handle explicit threading, we wouldn’t need all this dark silicon wasting power on renaming and speculation just to extract speed from a language that refuses to acknowledge how modern computers actually work. This is a fantastic read on the whole thing https://spawn-queue.acm.org/doi/10.1145/3212477.3212479
We can look at Erlang OTP for an example of a language platform looks like when it stops lying about hardware and actually embraces how modern chips work. Erlang was designed from the ground up for massive concurrency and fault tolerance. In C, creating a thread is an expensive OS-level operation, and managing shared memory between them is a nightmare that requires complex locking using mutexes and forces the CPU to work overtime maintaining cache coherency.
Meanwhile, in the Erlang world, you don’t have threads sharing memory. Instead, you have lightweight processes, that use something like 300 words of memory, that share nothing and only communicate by sending messages. Because the data is immutable and isolated, the CPU doesn’t have to waste cycles worrying about one core overwriting what another core is reading. You don’t need complex hardware logic to guess what happens next because the parallelism is explicit in the code, not hidden. The Erlang VM basically spins up a scheduler on each physical core and just churns through these millions of tiny processes. It feeds the hardware independent, parallel chunks of work without the illusion of serial execution which is exactly what it wants. So, if you designed a whole stack from hardware to software around this idea, you could get a far better overall architecture.
Erlang isn’t special because it’s functional, but rather it’s functional because that was the only way to make its specific architecture work. Joe Armstrong and his team at Ericsson set out to build a system with nine nines of reliability. They quickly realized that to have a system that never goes down, you need to be able to let parts of it crash and restart without taking down the rest. That requirement for total isolation forced their hand on the architecture, which in turn dictated the language features.
The specialness is entirely in the BEAM VM itself, which acts less like a language runtime like the JVM or CLR, and more like a mini operating system. In almost every other environment, threads share a giant heap of memory. If one thread corrupts that memory, the whole ship sinks. In Erlang, every single virtual process has its own tiny, private heap. This is the killer architectural feature that makes Erlang special. Because nothing is shared, the VM can garbage collect a single process without stopping the world, and if a process crashes, it takes its private memory with it, leaving the rest of the system untouched.
The functional programming aspect is just the necessary glue to make a shared nothing architecture usable. If you had mutable state scattered everywhere, you couldn’t trivially restart a process to a known good state. So, they stripped out mutation to enforce isolation. The result is that Erlang creates a distributed system inside a single chip. It treats two processes running on the same core with the same level of mistrust and isolation as two servers running on opposite sides of the Atlantic.
Learning functional style can be a bit of a brain teaser, and I would highly recommend it. Once you learn to think in this style it will help you write imperative code as well because you’re going to have a whole new perspective on state management.
See, this is what I despise about x86. AFAIK it’s literally RISC on the bare metal but there are hundreds of “instructions” running microcode which is basically just a translation layer. You’re not allowed to write code for the actual RISC implementation because that’s a trade secret or something. So obviously single core performance would be shit because you’re basically running an emulator all the time.
RISC-V can’t come fast enough. Maybe someone will even make a chip that’s RISC-V but with the same instruction/microcode support as x86. So you can run RISC-V code directly or do the microcode thing and pretend you’re on x86. Though that would probably get the shit sued out of them by Intel because god forbid there’s actual innovation that the original creator can’t cash in on.
RISCV would be a huge step forward, and there are projects like this one working on making a high performance architecture using it. But I’d argue that we should really be rethinking the way we do programming as well.
The problem goes deeper than just the translation layer because modern chips are still contorting themselves to maintain a fiction for a legacy architecture. We are basically burning silicon and electricity to pretend that modern hardware acts like a PDP-11 from the 1970s because that is what C expects. C assumes a serial abstract machine where one thing happens after another in a flat memory space, but real hardware hasn’t worked that way in decades. To bridge that gap, modern processors have to implement insane amounts of instruction level parallelism just to keep the execution units busy.
This obsession with pretending to be a simple serial machine also causes security nightmares like Meltdown and Spectre. When the processor speculates past an access check and guesses wrong, it throws the work away, but that discarded work leaves side effects in the cache that attackers can measure. It’s a massive security liability introduced solely to let programmers believe they are writing low level code when they are actually writing for a legacy abstraction. on top of that, you have things like the register rename engine, which is a huge consumer of power and die area, running constantly to manage dependencies in scalar code. If we could actually code for the hardware, like how GPUs handle explicit threading, we wouldn’t need all this dark silicon wasting power on renaming and speculation just to extract speed from a language that refuses to acknowledge how modern computers actually work. This is a fantastic read on the whole thing https://spawn-queue.acm.org/doi/10.1145/3212477.3212479
We can look at Erlang OTP for an example of a language platform looks like when it stops lying about hardware and actually embraces how modern chips work. Erlang was designed from the ground up for massive concurrency and fault tolerance. In C, creating a thread is an expensive OS-level operation, and managing shared memory between them is a nightmare that requires complex locking using mutexes and forces the CPU to work overtime maintaining cache coherency.
Meanwhile, in the Erlang world, you don’t have threads sharing memory. Instead, you have lightweight processes, that use something like 300 words of memory, that share nothing and only communicate by sending messages. Because the data is immutable and isolated, the CPU doesn’t have to waste cycles worrying about one core overwriting what another core is reading. You don’t need complex hardware logic to guess what happens next because the parallelism is explicit in the code, not hidden. The Erlang VM basically spins up a scheduler on each physical core and just churns through these millions of tiny processes. It feeds the hardware independent, parallel chunks of work without the illusion of serial execution which is exactly what it wants. So, if you designed a whole stack from hardware to software around this idea, you could get a far better overall architecture.
Is Erlang special in its architecture or is it more that it’s functional?
One day I’ll learn how to do purely functional, maybe even purely declarative. But I have to train my brain to think of computer programs like that.
Is there a functional and/or declarative language that has memory management features similar to Rust as opposed to a garbage collector?
Erlang isn’t special because it’s functional, but rather it’s functional because that was the only way to make its specific architecture work. Joe Armstrong and his team at Ericsson set out to build a system with nine nines of reliability. They quickly realized that to have a system that never goes down, you need to be able to let parts of it crash and restart without taking down the rest. That requirement for total isolation forced their hand on the architecture, which in turn dictated the language features.
The specialness is entirely in the BEAM VM itself, which acts less like a language runtime like the JVM or CLR, and more like a mini operating system. In almost every other environment, threads share a giant heap of memory. If one thread corrupts that memory, the whole ship sinks. In Erlang, every single virtual process has its own tiny, private heap. This is the killer architectural feature that makes Erlang special. Because nothing is shared, the VM can garbage collect a single process without stopping the world, and if a process crashes, it takes its private memory with it, leaving the rest of the system untouched.
The functional programming aspect is just the necessary glue to make a shared nothing architecture usable. If you had mutable state scattered everywhere, you couldn’t trivially restart a process to a known good state. So, they stripped out mutation to enforce isolation. The result is that Erlang creates a distributed system inside a single chip. It treats two processes running on the same core with the same level of mistrust and isolation as two servers running on opposite sides of the Atlantic.
Learning functional style can be a bit of a brain teaser, and I would highly recommend it. Once you learn to think in this style it will help you write imperative code as well because you’re going to have a whole new perspective on state management.
And yeah there are functional languages that don’t rely on using a VM, Carp is a good example https://github.com/carp-lang/Carp