4 ms·

The XuanTie C920 cores in the SOPHGO SG2042 are an older design without support for Bitmanip, cryptography or Vector 1.0. You could compile especially for it th

The XuanTie C920 cores in the SOPHGO SG2042 are an older design without support for Bitmanip, cryptography or Vector 1.0. You could compile especially for it though, and take advantage of T-Head's proprietary ISA extensions and the draft version of Vector supported, but it would still be far from comparable to a Zen core.

pengaru 3y ago
Are those extensions really what matters in comparing to a Zen core? I assumed these cores are lacking in the OoO/speculative exuction/branch prediction department. It makes them smaller so it's cheaper to have many of them, but won't be as fast as any state of the art x86 cores. Special extensions don't generally automagically get utilized even when they are present...
- Findecanor 3y ago
  I don't think Bitmanip is special but pretty much required for reaching feature-parity with ARM and x86 so that you could start comparing them. RISC-V is otherwise quite spartan, which might be desirable for embedded MCUs, but for application-class processors you'd want more. For example, an indexed load which on ARM and x86 is one instruction could on RISC-V require four dependent instructions without Bitmanip and two with. The vector extension is RISC-V's closest counterpart to SSE/AVX, although it is not used for scalar floating-point. BTW, the RISC-V foundation has published official Application Profiles for application-class processors. Those are probably going to be baseline for binary distributions in the future — they include Bitmanip and more. The RVA23 profile has Vector. Google's requirements for Android (with Google services) will also include vector cryptography.
- brucehoult 3y ago
  Of course these cores are approximately 4x slower than current x86 cores, but you get 64 of them for the price of an x86 with 12-16 performance cores. If you want 64 cores in, say, a Threadripper 7980X then you're looking at $5000 just for the chip, with motherboard and RAM etc on top of that. Yeah, it's 4x faster, but it's also 4x (or more) the price.
  - IshKebab 3y ago
    Fewer faster cores is better than many slow cores though. It's just harder to do.
  - ben-schaaf 3y ago
    From the benchmarks in the paper that I think everyone is basing this on: https://ui.adsabs.harvard.edu/abs/2023arXiv230900381B/abstract https://ui.adsabs.harvard.edu/abs/2023arXiv230900381B/abstra.... The 4x slower single-core seems to be based on a EPYC 7742. A modern 7950x has approximately double the single-core performance compared to that EPIC chip, with very close multi-core performance. So the comparison is more like 4x faster (or 8x if single-core matters) for 1x the price, since you can easily get a similar rig with a 7950x for $1500.

pengaru 3y ago

Are those extensions really what matters in comparing to a Zen core? I assumed these cores are lacking in the OoO/speculative exuction/branch prediction department. It makes them smaller so it's cheaper to have many of them, but won't be as fast as any state of the art x86 cores. Special extensions don't generally automagically get utilized even when they are present...

Findecanor 3y ago
I don't think Bitmanip is special but pretty much required for reaching feature-parity with ARM and x86 so that you could start comparing them. RISC-V is otherwise quite spartan, which might be desirable for embedded MCUs, but for application-class processors you'd want more. For example, an indexed load which on ARM and x86 is one instruction could on RISC-V require four dependent instructions without Bitmanip and two with. The vector extension is RISC-V's closest counterpart to SSE/AVX, although it is not used for scalar floating-point. BTW, the RISC-V foundation has published official Application Profiles for application-class processors. Those are probably going to be baseline for binary distributions in the future — they include Bitmanip and more. The RVA23 profile has Vector. Google's requirements for Android (with Google services) will also include vector cryptography.
brucehoult 3y ago
Of course these cores are approximately 4x slower than current x86 cores, but you get 64 of them for the price of an x86 with 12-16 performance cores. If you want 64 cores in, say, a Threadripper 7980X then you're looking at $5000 just for the chip, with motherboard and RAM etc on top of that. Yeah, it's 4x faster, but it's also 4x (or more) the price.
- IshKebab 3y ago
  Fewer faster cores is better than many slow cores though. It's just harder to do.
- ben-schaaf 3y ago
  From the benchmarks in the paper that I think everyone is basing this on: https://ui.adsabs.harvard.edu/abs/2023arXiv230900381B/abstract https://ui.adsabs.harvard.edu/abs/2023arXiv230900381B/abstra.... The 4x slower single-core seems to be based on a EPYC 7742. A modern 7950x has approximately double the single-core performance compared to that EPIC chip, with very close multi-core performance. So the comparison is more like 4x faster (or 8x if single-core matters) for 1x the price, since you can easily get a similar rig with a 7950x for $1500.