Investigating Cavium's ThunderX: ARM Server SoC | Hacker News Reader

Investigating Cavium's ThunderX: ARM Server SoC

anandtech.com

72 pointsdbcooper21 comments

Funny, just today I google on ThunderX boards and tried to track down the latest AMD Opteron A1100 developments. I wish there were simple, stripped down development boards for these kind of processors available. I would certainly enjoy building a personal computer from them. Sadly, there seems to be no viable competition today in terms of performance to x86/Core i7.

If you wanted to design a simple computer PCB with a RISC CPU in the GHz range (or 500 Mhz+), what would you choose?

http://www.hardkernel.com/main/products/prdt_info.php

The ODROID-C2 is a neat device, but not at all comparable with an A1100. The A1100 has 8 A57 cores, while the ODROID only has 4 A53s. Mainline kernel support for the A1100 (and HuskyBoard, etc) has been relatively solid for a while now, while the C-2's SoC only made it into 4.6, and we're seeing the beginnings of a device tree in 4.7.

edit: Some A1100s only have 4 cores. They're still 4 beefier cores than the S905.

You can buy an ODROID-C2 today, which A1100 systems are available ?

And that's the fundamental problem with ARM servers...

none, as far as I know. First likely to become available is probably going to be http://www.lenovator.com/product/103.html - they're claiming Q2 2016.

You can certainly make a usable PC from a raspberry pi, one of the varieties of beaglebones, etc. Assuming you don't mind early-2000s levels of performance they are perfectly serviceable. There isn't anything a hobbyist could reasonably be able to put together that is anywhere close to x86 performance though.

actually for some cool server performance stuff a raspberry pi is not really great, especially since he is missing at least some sata interface. Beagleboard is better but still misses sata

The new Beaglebone has eSATA, indeed it will not be that great for high-performance SSDs or many server applications but it could be perfectly fine for a PC.

Dual gigabit Ethernet ports and USB 3 round out the impressive X-15's impressive features. It's not available yet but should be by the end of the month or early July based on the wiki. If it runs OpenBSD I might choose it over Soekris.

Ah yes I forgot about the dual ethernet ports, it might be really nice paired with a good ethernet switch as a router system!

Probably the answer is AMD Cello[1]. This has a quad core Seattle. However I have had mine on preorder since April, and it still hasn't shipped (still 15 days to go in "Q2" though :-)

[1] http://www.lenovator.com/product/103.html

If this ever ships I'd be interested in getting one. I didn't realize anyone was actually making A1100 in boards you could actually buy, I thought they were all sold to integrators and they were only going to do the next iteration downmarket.

Love to see the power consumption number when all 48 cores are running at 100% at max freq.

Also, It takes hours to build ASOP android opensource project in typical x86 build server utilize all the 4/8 cores.

If 48 cores can be configure to build ASOP somehow, I love to see that number to see if there are "real" advantage to "48 cores" in that particular use cases.

Correct me if I'm wrong, but these seem to be TDP/speed/heat limited because they're based on a modified, older 28nm process design?

Process shrink requires a lot of resources and R&D spending, but surely they could have benefited from not using 2008 ARM cores in 2016?

They seem to have plans to move to a modern GlobalFoundries 14nm process, so it might be more competitive when that happens. The decision to use an older, in-order core is probably based on the fact that their cache and memory controller design is pretty bad, and unable to even feed the slower cores as it is, so faster cores would be a waste. It turns out it is really hard to shove a lot of highly performant cores in a single package, it is much easier to shove a lot of crappy ones.

I'm curious how AppliedMicro's X-Gene 3 will fare. It should be out later this year on 14nm.

It will be interesting to see how that one goes, indeed. I think their timing could be unfortunate, however, given AMD is about to reenter the server market in a big way in the same timeframe.

How so?

By the time X-Gene 3 is out I'm sure they'll be saying "wait for X-Gene 4".

Seems like if you're going to move to 14nm you need a reasonable amount of demand, or it's a big waste. The price jump for tape-out is pretty significant between 28 and 14, so if you're bootstrapping it makes sense to do a 28 and see if there's uptake, essentially making it up on price if the market is soft.

Reply on news.ycombinator.com