I think x86/amd64 virtualization has made ARM a much less compelling option for servers than what it would have been a few years ago.
I'm looking forward to benchmarks of new ARM server CPU (esp. AMDs), but in the past comparing scale out ARM boxes and tradtional servers in the same space has:
1. Been much more more in favor x86/amd64 under low load (much lower latency for users)
2. Been pretty similar under very high concurrent loads
3. Not had compelling differentiation in power consumption
In short 48 1.6 GHz ARM cores doesn't farewell against even a lower end x86/amd64 server with dual quad core CPUs with SMT (hyperthreading), that is 32 logical cores at 3.0 GHz. And the x86/amd64 is much cheaper.
In reality I could fill a rack with these arm blades or have two quad socket x86/amd64 servers be equivalent or better...
I hope this changes for the sake of consumer options, but I need to see benchmarks + power usage stats to believe it.
To be fair, comparing core counts directly, each of those ARM processors is quadcore - so there are actually 192 physical cores in that 3U rack, not 48.
That said, I'm not convinced ARM is ready for server applications either. This might be interesting to hosting companies that want to sell low-end dedicated servers, though. (OVH already does something similar with Kimsufi, I believe?)
The new aarch64 chips will support large amounts of memory and hardware virtualization. Despite low performance, the perf/watt and high memory density might be good enough to compete with x86. These exist only as engineering samples and emulators right now.
All of the ARM server products in existence today are basically demos. Presumably the "select customers" offered HP "moonshot" and Dell "copper" are working on software ports and automation infrastructure for an aarch64 future.
> In reality I could fill a rack with these arm blades or have two quad socket x86/amd64 servers be equivalent or better
Quad socket x86 is still prohibitively expensive, but in the price per performance assessment you might be right (at least if you do not consider the expensive high-end x86 CPUs). The ARM blades look interesting for bare metal clouds though.
> 4 2U dual socket x86/amd64 boxes still is a better value proposition.
Not if you're aiming at a good density (processing power per rack). But there are many blade options (8 boards in 3U, or 4 in 2U from Supermicro e.g.) for x86 that will fit the bill, yes.
The cost for the vendor should be much lower with ARM CPUs though (the CPUs will probably be dirt cheap, while Intel's aren't).
Context: I'm talking about relative to ARM blades, not denser x86 options.
>The cost for the vendor should be much lower with ARM CPUs though (the CPUs will probably be dirt cheap, while Intel's aren't).
Maybe for the vendor, but these ARM blades boxes are expensive for the consumer!
Lets compare two 2U boxes, one with 48 1.6 GHz ARM SoC and one with dual six-core Xeons with SMT (HT) at lets say 2.8 GHz, that is 48 logical cores. The x86/amd64 box in this comparison is:
0. Much cheaper!
1. Low latency at low loads
2. Similar throughput and latency at high loads
3. Uses less power
4. Easier to maintain
This. Purchasers of the highest-end and/or costly experimental architectures tend to be research institutions or national labs that are using grant money.
In the case of quad-socket, it's actually enterprises running SQL databases and VMware consolidation where the cost of the software license exceeds the cost of the server.
Just having 32 logical cores doesn't mean much for performance. Hyperthreading is not magical, dual quad core cpus is still 8 cores, with potentially some minor performance gains from the hyperthreading depending on the application.
In my experience with HPC (i.e. 100% load on all cores), enabling Hyperthreading gives you higher job throughput at the cost of increasing the time to complete a single job. The throughput improvement is around 5-30% depending on the applications being run. The average I've seen is around 20.
In high-load but I/O bound tasks Hyperthreading (Intel market speak for SMT) is magic. If something is compute bound, you are correct and the benefit is often slim, but typically not for application or database servers.
I'm looking forward to benchmarks of new ARM server CPU (esp. AMDs), but in the past comparing scale out ARM boxes and tradtional servers in the same space has:
In short 48 1.6 GHz ARM cores doesn't farewell against even a lower end x86/amd64 server with dual quad core CPUs with SMT (hyperthreading), that is 32 logical cores at 3.0 GHz. And the x86/amd64 is much cheaper.In reality I could fill a rack with these arm blades or have two quad socket x86/amd64 servers be equivalent or better...
I hope this changes for the sake of consumer options, but I need to see benchmarks + power usage stats to believe it.