On August 28, 2018 3:04 AM, Joel Wirāmu Pauling <j...@aenertia.net> wrote:
> On 28 August 2018 at 05:26, Joseph Mayer joseph.ma...@protonmail.com wrote: > > > Joel, > > Are you saying you gave up on using the PCIe at all? > > There's a 4-lane PCIe connector on the Rock64 right, aren't those > > dedicated lanes, and, if they'd somehow be shared with any other > > hardware, then you should still have supposedly >90% of the 16gbps > > capacity available? > > Did you try connecting some multiport 1gbps or 10gbps PCIe NIC? > > That's where my Wireless card goes - so it's not very useful. In other words you'd have more use of two two-lane PCIe ports. (Which I don't know if the RK3399 could drive in itself. PCIe switch chips which convert four PCIe lanes upstream to 2x or 3x four PCIe lanes downstream, seem to be generally pricy.) I'd wonder what performance you got if you put the wired ethernet on PCIe and wireless on USB. > My general experience with the Arm boards SBC's vs the Intel SoC's is > the Arm SBC's generally have screwed up some fundemental Bus sharing > in their designs (either placing the USB controller on so it shares > bandwidth with the GMAC) or Exposing only a Single X4 PCI lane (which > inevitably get's used by Wireless - unless you want to stick with the > onboard wireless which suffers from the first problem). I'd like to learn to know if-how-how much the "screwed up some fundamental" applies to the RK3399 and the MacchiatoBin/EspressoBin boards. > > Geekbench figures indicate that RK3399 and Celeron N3160 should perform > > fairly similarly. > > https://browser.geekbench.com/v4/cpu/search?q=rk3399 > > https://browser.geekbench.com/v4/cpu/search?utf8=✓&q=n3160 > > https://ark.intel.com/products/91831/Intel-Celeron-Processor-N3160-2M-Cache-up-to-2_24-GHz > > The first test I normally subject my boxes too is localhost flent > (which is mostly netperf/iperf3/iperf under the hood) which provides a > bunch of test suites which are relatively good at finding out where > the Packet gen/receive limits are and if there is jitter - and provide > a starting point for further investigation. Irtt (which is written in > go) is also a really useful tool to figure out latencies (it times > sleep accuracy) on the SoC's. > > Note the Arm boards don't have AES-NI either, so once you start > playing with VPN it get's pretty bad in comparison. (Some of the newer > arms do have AES offloads but - implementations are varied, the H3/H5 > sunxi platform is where I am focused on at the moment - but not for > network stuff) (Note, Sunxi don't do PCIe so you're stuck with XHCI there. Also Sunxi's Allwinner A64 has around 25% of the performance rating of the RK3399 on Geekbench.) AES performance is an interesting point.
