Looking at the CCR2004-1G-12S+2XS.
So it has 12 SFP+ ports and 2 25G SFP 28 ports. However looking at the block diagram there’s a PIPE connecting to the CPU with 2 x 25G for a total of 50G throughput.
Does this mean it has a maximum throughput of 50Gb?
The test results show routing with fastpath at a maximum of 39237 Mbit.
So what is the point of of a router with 12 10GB ports if it can only route 39.2Gb? Or am I missing something?
I don’t know many routers that can route at the speed of all ports combined, so this CCR2004 is no exception.
In most situations this is not a problem.
Indeed but if there’s intense traffic across 4 interfaces could completely choke up the router. Tile CCRs while only 1Gb can router wirespeed across all interfaces.
It is true, and that PIPE thing has no switch functionality either, so it also cannot do wirespeed switching between ports.
(just like other CCRs, but in this case it would have been a nice feature due to the limited connection between PIPE and CPU)
Somewhere there has to be a tradeoff. >50Gbps forwarding capability for $595 is still a bargain.
Yes you probably need to consider different use-cases where it either is used as only a router (with 2 ports in use, or maybe there is some extra port connected to a lower-bandwidth IoT or Guest network) or another case where it routes many networks but only at low speed, e.g. the SFP+ ports are all used with 1.25 Gbps SFP modules.
It is not to be used with all ports connected at their rated speed.
It would have been nice when it could e.g. be a router in a colo cabinet where up to 12 systems would be connected using SFP+ DAC and the internet is connected to one of the XS ports, and the 12 ports are switched for fast inter-system communication, but for that it would be recommended to get a separate switch and this router (bringing it back to the first mentioned use-case).
Of course we would not have had this discussion when it was released as a CCR2004-1G-4S+2XS but then probably prople would ask for “more ports” 
Probably 
. The real solution to this of course is to build something with ARM64 as control plane processor and PresteraDX as data plane forwarder (merging the CCR2004 & ROS7 CRS317, as it were). Wire speed simple IP routing (and getting less ‘simple’ with software upgrades as MT gets to grips with the L3(+) capabilities of the PresteraDX ASICs) and enough oomph for more complex tasks.
I think those devices like the 98dx8216 are positioned as L3 switching devices with the processor only intended for use as management/monitoring processor.
(to run a configuration UI and provide services like SNMP)
MikroTik uses the processor for full-featured routing, and this often creates confusion because first-time buyers who are familiar with devices described as “router/switch” expect L3 switch capability: limited routing functionality but at wirespeed. They are disappointed when the routing is never near to wirespeed.
Fortunately now for this class of devices this is being rectified and this wirespeed routing will become a reality, and we will be at the next square where home-users need to be educated that such a device still cannot be reasonable be used as a router on their superfast internet connection, because L3 routing switches lack the functionality required for that (firewalling etc) so it cannot be enabled, at least not on the internet-facing interface, and the processor still is the bottleneck.
For that, indeed a combined model would be very useful. Maybe there is a chip manufacturer somewhere that makes or will make a chip with both the switching engine and a higher-performance processor…?
Yes, they are. The CPU in the SOC is meant purely as a control plane: CLI/web interface, SNMP, Spanning Tree, routing protocols, stuff like that. The actual forwarding is done by the ASIC. Which is why those CPUs are pretty low-powered as a rule; such tasks don’t take much. With ROS7 & the CRS317 (and, I assume, the rest of the PresteraDX-based CRS3xx line to follow) MT is finally starting to tap the full potential of the chips, but I’m not sure how far that goes. It’s not clear to me what the L3(+) capabilities of the ASIC are beyond simple L3 forwarding, and the Marvell product briefs don’t really say.
But I think we may safely assume that things like NAT & ACLs/firewalling will be beyond the capability of the ASIC, which is going to create a dichotomy between ‘fast, cheap, but low functionality’ routers (ASIC-driven CRS) and ‘slow, more expensive, but can do anything you want’ routers (CPU-driven CCR). Pick your poison.
They could build something that could do both, ‘fast for simple things’ (ASIC), ‘slow for complex things’ (CPU) by gluing an AL32400 on top of a 98DX (the ‘merge CCR2004 and CRS317’), but I don’t know how easy that would be. Or how cheap; it may be simpler and cheaper to live with the split product line.
Probably they consider the processor on this chip very powerful. I have managed many L3 routing switches from the wellknown manufacturers in that field, and most of them have a much slower control plane processor than this (or else they have written their firmware in interpreted BASIC…). Those management functions and also the SNMP is usually very slow on those switches, to the point where it is becoming irritating. That is not the case on those MikroTik devices 
True . So have I, and your’re right. The CPU in most MT devices is actually pretty decent. I have worked with Cisco 7200VXR’s, way back when, and the fastest routing engine you could buy for those (the NPE-G2) was about as fast as a $69 hAP ac². On a good day.
Problem is that bandwidth usage has grown even faster . Who could think, back then, that 100+ Mbps internet connections would become normal? Gigabit ones commonplace? 10G ones on the horizon (which is why my home router is a CCR1036-8G-2S+; my provider is working on a trial, even though I’m not part of it (yet))?
Even so, I don’t see Joe Average needing more than 1G NAT/FW anytime soon.
True, but the problem is that the routing in CRS and many other devices is not fast enough for that. We “regularly” see disappointed CRS users here who kind of expected it.
We can point them to CCR and they usually will have no problem (certainly not with CCR2004) but then they start to expect other things, like wirespeed switching.
Oh well…
True
Hi @pe1chl
What could be the use case with many multi core CPUs and are they utilized in normal use cases of a small set up of a hotspot which will only have 1Gig of WAN circuit? Also CCR2004 having only fiber ports, and limitation of SFP ports requiring auto-negotiation to be turned off, which then stops link monitoring to cause any bonded interfaces to failover and failback the traffic, where is CCR2004 positioned in for?
Thanks