| 08:47 | topi` | it seems cnx-software has some inside info on upcoming boards from solid-run ... ARMADA 8040 networking board, is this true or just a false rumor? |
| 08:47 | jnettlet | topi`, we announced it last week. |
| 08:48 | jnettlet | it is a developer board for SDN's |
| 08:48 | jnettlet | if you go to our website to the armada page you will find it at the bottom |
| 08:48 | topi` | oh, there is a pre-order page |
| 08:48 | topi` | what is a SDN? |
| 08:48 | jnettlet | software defined network |
| 08:49 | jnettlet | basically top of the rack type networking |
| 08:50 | topi` | right :) |
| 08:50 | topi` | sounds fun, can I have one? |
| 08:50 | jnettlet | but it can be used for anything where fast ARM soc's and high bandwidth will be useful |
| 08:50 | jnettlet | storage, build server etc |
| 08:51 | topi` | yeah |
| 08:52 | topi` | it'll be cheaper than a SoftIron AMD A1100 server :) |
| 08:55 | topi` | how many preorders do you need to get an initial product run? |
| 08:55 | topi` | if there's one missing, PM me and I'll try to find some extra $300 ;) |
| 09:00 | jnettlet | topi`, the board is being developed in partnership with Marvell...so it is getting made regardless of preorders |
| 09:01 | jnettlet | it is the developer platform for their new SOC |
| 09:37 | topi` | ok, cool |
| 09:37 | topi` | good to hear that you have working relationship with marvell :) |
| 09:38 | topi` | do they have a kernel engineering team? They did at least back in 2008 |
| 09:38 | topi` | I met some of their kernel engineers in Linuxconf Ottawa |
| 09:42 | jnettlet | Yes they do. We have direct collaboration with them |
| 10:35 | topi` | I hope we could eventually see 8-core versions of the 8040 :) |
| 10:35 | topi` | it would probably kick the butt of a 8-core Avoton cpu- and I/O-wise |
| 10:36 | topi` | you should perhaps try to offer that board to the folks at scaleway.com? They previously built their own compute node using ARMADA |
| 10:38 | topi` | sadly the single-core perf of their ARMADA nodes is underwhelming... even operations like apt-get install require some patience |
| 10:41 | jnettlet | topi`, they will be launching 8, 16, 32 core variants during the first quarter of 2017. http://www.marvell.com/company/news/pressDetail.do?releaseID=8317 |
| 10:43 | suihkulokki | topi`: scaleway uses armada xp which was launched in 2011 |
| 10:44 | jnettlet | yep the armada a38x is very different |
| 10:44 | suihkulokki | topi`: at that point it was pretty much the fastest arm core around (iirc the alternatice was pandaboard with 2x cortex-a9) |
| 10:48 | vpeter | haha - commnet on cnx: "In SolidRun we trust. Don't screw up this board, guys." |
| 10:48 | jnettlet | :) |
| 11:58 | rabeeh | nice comment |
| 11:58 | rabeeh | we will try not to screw it :) |
| 11:58 | rabeeh | suihkulokki: that was using Marvell PJ4B internally developed processor |
| 11:58 | rabeeh | A38x and beyond are standard ARM processor |
| 11:59 | rabeeh | In a robust high end environment i tend to use standard ARM processors since finding all the processor bugs really depends on how many engineers puts his hands on the processor |
| 12:00 | rabeeh | when using standard instruction set but proprietary implementation the audience of engineers that tries different workloads and stumbles with issues is far less than using standard instruction set and standard implementation |
| 12:01 | rabeeh | imagine a production border gateway that gets attacked because of buggy pipeline |
| 12:03 | suihkulokki | rabeeh: I'd be more concerned about obscure crashes than getting attacked |
| 12:04 | rabeeh | suihkulokki: oh yeah |
| 12:04 | rabeeh | been there, done that. thank you no more of it |
| 12:04 | rabeeh | :) |
| 12:05 | suihkulokki | otoh at least the current armada xp's been super reliable - I guess it took quite a few iterations to get there ;) |
| 12:06 | rabeeh | so; just reviewed the aramada xp product again; it's called PJ4C |
| 12:06 | rabeeh | Armada XP is super reliable thanks to millions of millions of tests that done in pre and post silicon |
| 12:07 | rabeeh | but that's exactly what i'm saying; if the target application of the core is smartphone; then you don't mind resetting it one a week |
| 12:07 | rabeeh | if you use the same core in enterprise environment then it's totally different story |
| 12:31 | topi` | so Marvell is one of those guys who have also attempted a custom ARM core? |
| 12:32 | topi` | I'd reckon the ARM cores are good enough and cheap enough that why bother designing a custom core, especially considering the huge cost of that kind of attempt |
| 12:33 | topi` | I meant the cortex-* arm cores |
| 12:33 | topi` | it's actually amazing how many cores ARM designs, develops and debugs |
| 12:34 | topi` | there's been a8, a5, a7, a9, a15, a12/a17 , a53, a57, a72, a35, a73? |
| 12:34 | topi` | that's 11 unique cores. oof |
| 12:34 | topi` | I'm pretty certain the amount of engineers in ARM is just a tiny fraction, say, 10%, of companies like Intel |
| 12:36 | topi` | anyhow, if I get enough RAM, ethernet and SATA out of your 8040 eval board, then I'm going to be a customer :) |
| 12:42 | jnettlet | well RAM is a single DIMM slot...so in theory you can up it to more than 4GB |
| 12:42 | jnettlet | it is a 64-bit processor |
| 12:42 | topi` | on my scaleway machine, /proc/cpuinfo says CPU implementer : 0x56 |
| 12:43 | topi` | that's ascii code for 'M' so I guess it's marvell's custom core |
| 12:43 | topi` | single DIMM slot means that the ram access is 32 bit wide? or is it 64bit? |
| 12:43 | jnettlet | 64-bit, just single channel |
| 12:44 | topi` | right. |
| 12:45 | topi` | back in 1997, I used to own an Alpha AXP machine. It had only 64-bit bus to RAM (pairs of SIMMs installed) whereas the more expensive 21064 models had 128bit bus (you added SIMMs in groups of 4) |
| 12:46 | topi` | the Cortex-A53 reminds me of those Alphas. Lots of similarities. Dual-issue, in-order, 64-bit, short pipeline, prefetch hints for LD/ST units |
| 12:47 | topi` | of course it was 30-60 watts against 200 milliwatts, 300 MHz against 2000 MHz and so on.. times change. |
| 12:48 | jnettlet | yes they do |
| 13:01 | topi` | damn, now our performance bonuses have been bound to the amount of NEW customers we can acquire the coming year |
| 13:01 | topi` | this will mean more projects, and less time to handle individual projects (well) ... lots of deadline misses are going to result of this :( |
| 13:12 | topi` | how much theoretical bandwidth would we be getting out of that single 64-bit DDR bus? |
| 13:12 | topi` | some mobile SoCs can do around 25 GB/s |
| 13:12 | topi` | but that's with LPDDR4 |
| 13:21 | rabeeh | those are max theoritical |
| 13:21 | rabeeh | i.e. frequency; multiplied by width :) |
| 13:22 | rabeeh | in practice, given the higher CL devices you get in a good day 60% of that |
| 13:22 | rabeeh | i mean in standard compute workload. |
| 13:22 | rabeeh | gpu and video workloads are different stories. |