5 ms·

You might check out the NanoPi Fire3, which I think came out in early 2018: https://www.friendlyarm.com/index.php?route=product/product&product_id=206 https://

You might check out the NanoPi Fire3, which I think came out in early 2018: https://www.friendlyarm.com/index.php?route=product/product&product_id=206 https://www.friendlyarm.com/index.php?route=product/product&... Has an eight-core 1.4GHz CPU, and is otherwise mostly comparable to the new Pi: 1G RAM, 1Gb ethernet. Has an SD card slot and no built-in storage.

vardump 7y ago
8 slow in-order A53 cores. I'd much rather take RPi4's four wide out-of-order A72 cores.
- rbanffy 7y ago
  It really depends on your use. If you plan to build a desktop replacement, thread performance matters. If you want a small cluster node for educational purposes, more slow cores are better.
  - vardump 7y ago
    Clock for clock, A72 is 2-4x faster than A53. But yeah, perhaps it's better to have a lot of small cores for a cluster.
    - megous 7y ago
      That's great, but if you need crypto (because you have LUKS storage and send AES encrypted data over the network, like in a usual NAS case), A53 with aes instructions may be faster than this A72 without one. And so far it seems these cores don't have aes instructions.
      - vardump 7y ago
        Seriously? They left out Cryptography Extensions from the RPi4 SoC? Any source / references? http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.100097_0002_04_en/way1395175472464.html http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.... Edit: Ouch if this turns out to be true... https://www.raspberrypi.org/forums/viewtopic.php?t=243410 https://www.raspberrypi.org/forums/viewtopic.php?t=243410 Edit 2: Ah, also RPi3B and 3B+ didn't have those extensions. Oh well. Edit 3: Just tested my bog standard RPi3B+ running Raspbian. It could do "openssl speed -elapsed aes-256-cbc" 43 MB/s AES-256-CBC. Tried also with "-multi 4" and it resulted 144 MB/s using all 4 cores. So perhaps RPi4 will be fast enough with CPU only crypto... would still love to have HW assist. Edit 4: RPi4 running 32-bit OS can do about 65 MB/s per core aes-256-cbc. 85 MB/s per core for aes-128-cbc. So by using two CPU cores for encryption (+ heatsink + fan :-)) 1 Gbps ethernet can be saturated.
        megous 7y ago
        Nice investigation!
        vardump 7y ago
        I believe it might be possible to use VideoCore VI for crypto acceleration. I think it would be pretty tricky to get LUKS to use it, though. At least writing a kernel module, but most likely LUKS patch would be required. You'd probably have to choose between 3D acceleration and full disk encryption. Looks like VideoCore VI is getting some compute shader support: https://gitlab.freedesktop.org/anholt/mesa/tree/v3d-cs/src/gallium/drivers/v3d https://gitlab.freedesktop.org/anholt/mesa/tree/v3d-cs/src/g...
      - fulafel 7y ago
        Maybe. It has a lot of CPU power relative to IO bandwidth, it should do ok in pure software. Hmm, anyone know if the VideoCore VI would be any good for crypto?
        vardump 7y ago
        Quick Googling [0] revealed VideoCore VI might have at least some support for Vulkan and/or OpenCL. If this turns out to be true, then yes, VideoCore VI might be able to function as a crypto-coprocessor. We'll see. [0]: https://www.phoronix.com/scan.php?page=news_item&px=Broadcom-VC5-VC6-Hmm https://www.phoronix.com/scan.php?page=news_item&px=Broadcom...
        fulafel 7y ago
        People have also written some native code for the VC IV in assembler (both QPU and VPU) in the past.
        megous 7y ago
        Depends on your use case. If you have SSD attached over USB3, you can easilly pull 400MB/s. This would severely limit that, if that is an encrypted drive. On Allwinner H6, for example, you could get that speed on oncrypted drive while still having other cores free for other stuff. So if you need to go through/process gigabytes of data on encrypted drive, having to do encryption in SW will slow you down massively, especially if you also need to process the retrieved data somehow.
      - rbanffy 7y ago
        I would never expect (or demand) a cluster designed for educational purposes out of sub $100 nodes to be fast. In fact, being slow can be considered a feature in this scenario.
        megous 7y ago
        Yet, you can have the same priced board that can do 200-800MB/s AES-128 encryption (per core), leaving much more time to the actual useful work, or idle (less power consumption). https://github.com/ThomasKaiser/sbc-bench/blob/master/Results.md https://github.com/ThomasKaiser/sbc-bench/blob/master/Result... See PineH64 for example.
rbanffy 7y ago
It's interesting, but it's an 8-way in-order part. Shipping from them used to be horrendously expensive, but I'm happy to notice that changed. Thank you.
Matsta 7y ago
Yes, these are A53 cores which the RPi 3 used. The RPi 4 uses A72 cores, which are a lot faster. A53 are efficiency optimised cores and A72 are performance optimised cores. The RPi has such a big community that I imagine performance and optimisation will quickly surpass other chips, eg. Rockchip RK3399 [1] (2xA72 and 4xA53) and will probably end up on par with an Amlogic S922x [2](4xA73 and 2xA53). [1] http://rockchip.wikidot.com/rk3399 http://rockchip.wikidot.com/rk3399 [2] https://www.cnx-software.com/2019/02/01/amlogic-s922x-benchmarks/ https://www.cnx-software.com/2019/02/01/amlogic-s922x-benchm... To get the full performance out of the RPi, you are going to need a beefy heatsink or it will thermal throttle very quickly.

vardump 7y ago

8 slow in-order A53 cores. I'd much rather take RPi4's four wide out-of-order A72 cores.

rbanffy 7y ago
It really depends on your use. If you plan to build a desktop replacement, thread performance matters. If you want a small cluster node for educational purposes, more slow cores are better.
- vardump 7y ago
  Clock for clock, A72 is 2-4x faster than A53. But yeah, perhaps it's better to have a lot of small cores for a cluster.
  - megous 7y ago
    That's great, but if you need crypto (because you have LUKS storage and send AES encrypted data over the network, like in a usual NAS case), A53 with aes instructions may be faster than this A72 without one. And so far it seems these cores don't have aes instructions.
    - vardump 7y ago
      Seriously? They left out Cryptography Extensions from the RPi4 SoC? Any source / references? http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.100097_0002_04_en/way1395175472464.html http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.... Edit: Ouch if this turns out to be true... https://www.raspberrypi.org/forums/viewtopic.php?t=243410 https://www.raspberrypi.org/forums/viewtopic.php?t=243410 Edit 2: Ah, also RPi3B and 3B+ didn't have those extensions. Oh well. Edit 3: Just tested my bog standard RPi3B+ running Raspbian. It could do "openssl speed -elapsed aes-256-cbc" 43 MB/s AES-256-CBC. Tried also with "-multi 4" and it resulted 144 MB/s using all 4 cores. So perhaps RPi4 will be fast enough with CPU only crypto... would still love to have HW assist. Edit 4: RPi4 running 32-bit OS can do about 65 MB/s per core aes-256-cbc. 85 MB/s per core for aes-128-cbc. So by using two CPU cores for encryption (+ heatsink + fan :-)) 1 Gbps ethernet can be saturated.
      - megous 7y ago
        Nice investigation!
        vardump 7y ago
        I believe it might be possible to use VideoCore VI for crypto acceleration. I think it would be pretty tricky to get LUKS to use it, though. At least writing a kernel module, but most likely LUKS patch would be required. You'd probably have to choose between 3D acceleration and full disk encryption. Looks like VideoCore VI is getting some compute shader support: https://gitlab.freedesktop.org/anholt/mesa/tree/v3d-cs/src/gallium/drivers/v3d https://gitlab.freedesktop.org/anholt/mesa/tree/v3d-cs/src/g...
    - fulafel 7y ago
      Maybe. It has a lot of CPU power relative to IO bandwidth, it should do ok in pure software. Hmm, anyone know if the VideoCore VI would be any good for crypto?
      - vardump 7y ago
        Quick Googling [0] revealed VideoCore VI might have at least some support for Vulkan and/or OpenCL. If this turns out to be true, then yes, VideoCore VI might be able to function as a crypto-coprocessor. We'll see. [0]: https://www.phoronix.com/scan.php?page=news_item&px=Broadcom-VC5-VC6-Hmm https://www.phoronix.com/scan.php?page=news_item&px=Broadcom...
        fulafel 7y ago
        People have also written some native code for the VC IV in assembler (both QPU and VPU) in the past.
      - megous 7y ago
        Depends on your use case. If you have SSD attached over USB3, you can easilly pull 400MB/s. This would severely limit that, if that is an encrypted drive. On Allwinner H6, for example, you could get that speed on oncrypted drive while still having other cores free for other stuff. So if you need to go through/process gigabytes of data on encrypted drive, having to do encryption in SW will slow you down massively, especially if you also need to process the retrieved data somehow.
    - rbanffy 7y ago
      I would never expect (or demand) a cluster designed for educational purposes out of sub $100 nodes to be fast. In fact, being slow can be considered a feature in this scenario.
      - megous 7y ago
        Yet, you can have the same priced board that can do 200-800MB/s AES-128 encryption (per core), leaving much more time to the actual useful work, or idle (less power consumption). https://github.com/ThomasKaiser/sbc-bench/blob/master/Results.md https://github.com/ThomasKaiser/sbc-bench/blob/master/Result... See PineH64 for example.

rbanffy 7y ago

It's interesting, but it's an 8-way in-order part. Shipping from them used to be horrendously expensive, but I'm happy to notice that changed. Thank you.

Matsta 7y ago

Yes, these are A53 cores which the RPi 3 used. The RPi 4 uses A72 cores, which are a lot faster. A53 are efficiency optimised cores and A72 are performance optimised cores. The RPi has such a big community that I imagine performance and optimisation will quickly surpass other chips, eg. Rockchip RK3399 [1] (2xA72 and 4xA53) and will probably end up on par with an Amlogic S922x [2](4xA73 and 2xA53). [1] http://rockchip.wikidot.com/rk3399 http://rockchip.wikidot.com/rk3399 [2] https://www.cnx-software.com/2019/02/01/amlogic-s922x-benchmarks/ https://www.cnx-software.com/2019/02/01/amlogic-s922x-benchm... To get the full performance out of the RPi, you are going to need a beefy heatsink or it will thermal throttle very quickly.