[HN Gopher] Nvidia to ship a billion of RISC-V cores in 2024
       ___________________________________________________________________
        
       Nvidia to ship a billion of RISC-V cores in 2024
        
       Author : ohong
       Score  : 22 points
       Date   : 2024-10-24 19:38 UTC (3 hours ago)
        
 (HTM) web link (www.tomshardware.com)
 (TXT) w3m dump (www.tomshardware.com)
        
       | ohong wrote:
       | Announced during Nvidia's presentation at this year's RISC-V
       | Summit.
        
       | mixmastamyk wrote:
       | Great news that I've never heard before. Good to know R5 has at
       | least one friend with _deep_ pockets. Wouldn 't have guessed
       | NVDA, who has been somewhat antagonistic to open computing.
        
         | jsheard wrote:
         | These RISC-V cores are an internal implementation detail of
         | their GPUs, so although the architecture is open they aren't
         | really open in the sense that you can actually program them.
         | Nvidia is doing it to save a buck because they no longer have
         | to roll their own architecture (as they did previously) or pay
         | for ARM licenses.
        
           | almostgotcaught wrote:
           | They're also not the SMs but the control processors - they're
           | basically replacing esp32s with riscv based thingamajigs.
        
       | camel-cdr wrote:
       | What really surprised me to learn is that use some VLEN=1024 RVV
       | 1.0 cores:
       | https://x.com/FlorianWoh/status/1848771326791848438/photo/3
        
         | jsheard wrote:
         | That neatly lines up with their actual GPU cores which are also
         | 1024 bits wide (32 lanes of FP32) but I have no idea why they
         | would need to go that wide in an auxiliary management core.
        
           | camel-cdr wrote:
           | I noticed that as well, but it's still not obvious to me why,
           | they must do a decent amount of complex computation on them.
           | Otherwise, they could've also used VLEN=128 with LMUL=8 to
           | get effectively four 1024-bit vector registers.
        
       | knowitnone wrote:
       | Meh. They've spent a bunch of money engineering these but could
       | have bought a microcontroller from the numerous companies for
       | pennies. Yes, they have more than enough money to waste which is
       | exactly what they have done here. If they have further plans for
       | RISC-V, ie compete with ARM, AMD, Intel, Amazon, then that is
       | something to cheer.
        
       ___________________________________________________________________
       (page generated 2024-10-24 23:01 UTC)