4 ms·

I think the fact that it will run mainline amd64 Linux [0] and is dirt cheap will make people overlook that. Benchmarks are imperfect, but it seems to be in th

by DCKing 7y ago

I think the fact that it will run mainline amd64 Linux [0] and is dirt cheap will make people overlook that. Benchmarks are imperfect, but it seems to be in the same ballpark as the Raspberry Pi 4 [1] - slightly worse in single core, slightly better in multicore. So performance wise it's competitive with competing ARM boards (you get what you pay for) but the software experience should be great in comparison with most boards, even in comparison to the RPi 4. [0]: And Windows. And FreeBSD. Perhaps even OpenBSD http://openbsd-archive.7691.n7.nabble.com/Fwd-Intel-Compute-Stick-BOXSTK1AW32SC-td294917.html http://openbsd-archive.7691.n7.nabble.com/Fwd-Intel-Compute-... . [1]: https://browser.geekbench.com/v4/cpu/compare/13701853?baseline=12971889 https://browser.geekbench.com/v4/cpu/compare/13701853?baseli... [2]: Sidenote that the benchmark is skewed because of the AES benchmark. The Raspberry Pi 4 like the Atom has AES hardware acceleration when run in 64 bit mode, but this benchmark is run on 32-bit ARM. This is likely because the official RPi 4 distribution (Raspbian) is also still 32 bit. Software support affects benchmarks too.

vardump 7y ago
RPi4's specific weak spot is heat production. It'll probably severely throttle during the benchmark if that $1 heat sink is missing. So I wonder what's the case in this benchmark? Did the test system have a heatsink or not? Low multicore results suggest not. Another issue that can skew benchmarks is library optimization situation. On x86, most things are SSE/AVX optimized, but on ARM side NEON optimizations can be slow implementations or even completely missing. AArch64 ISA has changes that eliminate some false dependency chains (more general purpose registers and NEON register layout changes). This blocks out-of-order engine from making optimal use of CPU resources. When you hit this issue, 64-bit ARM code can run quite a bit faster. Raspberry Pi's weak spot has always been rather slow memory bandwidth. RPi4 improves this (DDR4 instead of DDR2), but it's still the weakest spot. Just 32-bits wide bus to SDRAM.
- DCKing 7y ago
  I mean yeah, you can split a lot of hairs there. But even so, even if you assume the RPi 4 can be optimized to be 20% faster than the Atom across the board, it's still the same ballpark. In that ballpark mainstream amd64 support is a big advantage.
  - vardump 7y ago
    NEON optimizations can have pretty large effect (up to ~4x), just like SSE. The benchmarks with the largest differences are suspect, especially SGEMM. Also don't underestimate how much RPi4 throttles under multicore load. Even the cheapest $1 heatsink will make a big difference. In general I agree, though. It'll still be in the same ballpark.

by DCKing 7y ago

vardump 7y ago

RPi4's specific weak spot is heat production. It'll probably severely throttle during the benchmark if that $1 heat sink is missing. So I wonder what's the case in this benchmark? Did the test system have a heatsink or not? Low multicore results suggest not. Another issue that can skew benchmarks is library optimization situation. On x86, most things are SSE/AVX optimized, but on ARM side NEON optimizations can be slow implementations or even completely missing. AArch64 ISA has changes that eliminate some false dependency chains (more general purpose registers and NEON register layout changes). This blocks out-of-order engine from making optimal use of CPU resources. When you hit this issue, 64-bit ARM code can run quite a bit faster. Raspberry Pi's weak spot has always been rather slow memory bandwidth. RPi4 improves this (DDR4 instead of DDR2), but it's still the weakest spot. Just 32-bits wide bus to SDRAM.

DCKing 7y ago
I mean yeah, you can split a lot of hairs there. But even so, even if you assume the RPi 4 can be optimized to be 20% faster than the Atom across the board, it's still the same ballpark. In that ballpark mainstream amd64 support is a big advantage.
- vardump 7y ago
  NEON optimizations can have pretty large effect (up to ~4x), just like SSE. The benchmarks with the largest differences are suspect, especially SGEMM. Also don't underestimate how much RPi4 throttles under multicore load. Even the cheapest $1 heatsink will make a big difference. In general I agree, though. It'll still be in the same ballpark.