Benchmarking An ARM 96-Core Cavium ThunderX System

[edwaleni]

Gigabyte released their ThunderX2 boards in November 2017.

[Michael_S]

If you want to play with ThunderX, Scaleway offer instances up to 64 vCPU/128GB running on ThunderX.

Nice, thanks for that. The Scaleway offering is 64 vCPU ARM server, 128GB RAM, 1TB SSD, 280 Euro/month. For contrast their x86_64 12 core server has 120 GB RAM, 1TB SSD, 180 Euro/month.

That seems pretty cost-competitive, depending upon your workload and what 12-core x86_64 CPU they're using. If you're running, say, something that does C-Ray style calculations or something that matches the performance profile of Java JMH then this thing is probably worth the cost. Otherwise, no.

If I was getting a high performance home machine I'd go AMD Threadripper or Core i9 over this.

[heliosh]

Some of the results just dont make sense. You would think highly parallel things like the Linux kernel woukd do better even with relatively weak cores. This has me wondering if there are teething or configuration problems.

I would assume a lack of memory bandwidth, or inefficient way for those 96 threads to access memory. I wonder how the CPU load was.

[Kendji]

Not a cheap server, but would be sweet to have

[aaronage]

There's some great coverage of ThunderX2 on The Next Platform.

Some nice detail on the architecture and projected performance here:
https://www.nextplatform.com/2017/11...-server-punch/

And this has some numbers from Cray (they're offering ThunderX2 supers):
https://www.nextplatform.com/2017/11...essor-shakeup/

Edit: if you're interested in the world beyond x86, it's worth following Next Platform. They're the only site I know of that goes into detail on developments in ARM, Power etc. They also do nice coverage of hardware beyond that, like Google's homegrown TPU (Tensor accelerator).

(not here to spam about that site, just like what they do).

[mbello]

Michael In the article you say this is a current gen processor when it is clearly not.

ThunderX 2 is supposed to improve a lot, not only because it is 14nm based but also because Cavium was more aggressive improving the arch of on the second gen.

But in any case, for those interested in ARM server chips look at Cloudflare's benchmarks of Qualcomm's Centriq 2400, they evaluated Centriq, TX2 and latest Xeons and chose Centriq for parts of their workload. Also, Centriq performance was great and the performance per watt blew Intel out of the water.

[aaronage]

If you want to play with ThunderX, Scaleway offer instances up to 64 vCPU/128GB running on ThunderX.

[aaronage]

The performance isn't surprising if you look at the architecture of this chip (I'm having trouble finding the details about the old ThunderX architecture now, but I vaguely remember it being in-order, 2-issue and having small L1 data/inst cache per core - could be wrong). ThunderX seems tailored to certain workloads (like networking), not a jack of all trades like Zen, 'Lake, Centriq etc.

These chips launched in 2014 anyway, ancient history by this point.

ThunderX2 will be more interesting, it's a more typical wide out-of-order design so should compete nicely with all the Zen, 'Lake etc.

[wizard69]

Some of the results just dont make sense. You would think highly parallel things like the Linux kernel woukd do better even with relatively weak cores. This has me wondering if there are teething or configuration problems.

By the way i wasnt expecting superior single core performance, that isnt the point of this chip or other high core count processors. However with 96 cores you do expect better from some of those tests.

[chuckula]

It's not disappointing because of its blatant lack of single-thread performance.

It's disappointing because Phoronix teed up benchmarks that should literally fall right into the lap of an OMG 96 COAR system with gobs of memory bandwidth and a pretty large power consumption envelope that's as least as high as a dual socket Xeon or Epyc system. And even there it's clearly inferior to a desktop system you could have bought last year, much less a real server.