3 ms·
> RISC-V implementations in the wild lack advanced features that modern CPUs rely on for speed, including sophisticated pipelining mechanisms, out-of-order exec
by amelius 2y ago
> RISC-V implementations in the wild lack advanced features that modern CPUs rely on for speed, including sophisticated pipelining mechanisms, out-of-order execution capabilities, advanced branch prediction, and multi-tiered cache hierarchies. Most commercial RISC-V chips remain in-order processors, meaning they execute instructions sequentially rather than optimizing their order for performance. This architectural simplicity creates a fundamental performance ceiling that's difficult to overcome without significant architectural changes.
This is a bit surprising given that all these techniques have been in computer architecture textbooks since at least the 90s.
- undersuit 2y agoYou don't need these tricks, you can always increase clocks and speed up memory interfaces. It was a lot harder for Intel to widen the memory interface of the P5 or increase the clocks sufficiently so they made it superscalar. RISC-V gets to take advantage of being produced in 2024 and absorbing all the clock speed and transistor advantages we get for free today because of 6 decades of transistor production.
- amelius 2y agoBut how do you explain the 25x performance gap then?
- phkahler 2y agoIt costs a lot of money to develop an advanced CPU and deploy on an advanced node. You need a market for the chips to justify that so it's a bit of chicken or egg problem. The gap is slowly closing and I'm looking forward to a RISC-V linux laptop in the next few years.
- undersuit 2y agoRISC-V hasn't maxed out. A new 2.4Ghz clockspeed at 7nm(dev board actually 1.4Ghz) when my old Ryzen 7 5800X at 7nm gets 4.5Ghz. https://www.theregister.com/2024/04/09/sifive_riscv_hifive/ https://www.theregister.com/2024/04/09/sifive_riscv_hifive/
- librasteve 2y agoi would guess that over 50% of the design investment in modern OOO CPUs is in branch prediction, cache tuning and so on … with multiple speed /cost points available to fine tune and balance prefetch, decode, instruction unit, blah blah to max out benchmark scores (the bible according to Hennessy et al). i would be more interested to see how fast you can make the literally 1000s of “snitch” cores go with the right software.