Hands On & Initial Benchmarks With An Ampere eMAG 32-Core ARM Server

[Travis Lazar]

Michael -

Thanks for taking the time to play with our hardware and share these thoughts; your Phoronix posts are always well-written and informative. The Calxeda project is an interesting case-study on this architecture and a great look at where things were just a short 6 years ago.

I'm happy to see that you had an easy time getting your system up and running, despite the cabling we included in the box. We're eager to read more when you've had time to run some of the bigger benchmarks, and we're always open to feedback from developer communities like this one.

---

Travis Lazar
Software Projects - Ampere Computing
[email protected]

---

[Michael]

Michael -

Thanks for taking the time to play with our hardware and share these thoughts; your Phoronix posts are always well-written and informative. The Calxeda project is an interesting case-study on this architecture and a great look at where things were just a short 6 years ago.

I'm happy to see that you had an easy time getting your system up and running, despite the cabling we included in the box. We're eager to read more when you've had time to run some of the bigger benchmarks, and we're always open to feedback from developer communities like this one.

---

Travis Lazar
Software Projects - Ampere Computing
[email protected]

---

Cabling was just more of a pleasant shock... Seeing a USB serial cable and fearing lengthy setup when in reality, power it up and it's all working great! Smooth process certainly made it easy getting out these initial tests.

[Michael]

Don't you have access to some ThunderX2 machine? This should be competitive against this Ampere chip.

As written in the article, I don't have any ThunderX2 access. Even with the X1, it was just remote access for a short time.

I did try getting access to X2 from Cavium, but they claimed PTS is biased against ARM which is clearly just crap.... Considering it's just collecting interesting open-source/upstream benchmarks and can be seen from the great Ampere performance here, etc.

[juanrga]

comparing it with Xeon D would be a much more useful comparison.

This is not a microserver. This doesn't target Xeon D, but Xeon E5 class

[ldesnogu]

As written in the article, I don't have any ThunderX2 access. Even with the X1, it was just remote access for a short time.

I did try getting access to X2 from Cavium, but they claimed PTS is biased against ARM which is clearly just crap.... Considering it's just collecting interesting open-source/upstream benchmarks and can be seen from the great Ampere performance here, etc.

How disappointing from them. If they feel some of these open source packages are not good enough for ARM, then they should invest in fixing them rather than dismissing all of them. Hopefully now that they've been bought by Marvell the situation will improve.

[ms178]

How disappointing from them. If they feel some of these open source packages are not good enough for ARM, then they should invest in fixing them rather than dismissing all of them. Hopefully now that they've been bought by Marvell the situation will improve.

I agree, maybe they were reluctant to be part of a competition so soon. The ARM ecosystem is still evolving fast. Also there is still room for a lot of software and platform optimizations. Glibc got some performance improvements merged lately. ZLIB received some patches from Google some time ago but the merging process is very painful there.

The 32 cores, 8 channels of DDR4, and 42 PCIe lanes puts it firmly in Threadripper/EPYC territory (on paper anyways) and solidly ahead of anything from intel. It's pretty amazing they managed all this in just 125w TDP.

Indeed! And the 7nm iteration is also just around the corner. I also vaguely remember them being mentioned in a Microsoft talk about Windows 10 on ARM as hinted at in my first post, so that would come in handy if they wanted to enter the Workstation market at a later date. Microsoft's ARM software ecosystem is still lagging behind though with a different focus, entering the notebook market first. That is a safer bet, I have to admit but I'd argue that it would be far more interesting to have a competitor with a different ISA in the Workstation market as well. But I can understand that this needs a lot of work on both the hardware and software side to get it right and too much risk right now. Growing the customer base for ARM64 (on both Windows and Linux), maturing the software and hardware along the way, is the safer bet before getting a real chance to compete against Intel and AMD.

[wizard69]

Nice machine! Impressive performance but not unexpected given what we see from modern ARM processors.

In any event Michael is like to see testing done as an engineering workstation not a server. Floating point performance and the ability to add in a good graphics card is the key to usefulness as an engineering work station. What makes this interesting is the low power meaning the potential for a quiet computer under load. Manufactures comments about such uses would be interesting also. The dream here is reasonable performance that doesn’t require an air conditioner blowing on the PC.

[edwaleni]

I did try getting access to X2 from Cavium, but they claimed PTS is biased against ARM which is clearly just crap....

When the maker blames the tester, then they don't have anything to show.

If Cavium is that insecure about their market and the fact a benchmark may counter sales talk, then they have the issue.

"bias against ARM" sheesh.

[speculatrix]

A Intel xeon 6142, which has a 16 core CPU, retails for over $3100, so this emag cpu is quite competitive in price if you're only counting cores.
I'd be interested what the price of an equivalent Intel or AMD device is when comparing performance.

[juanrga]

I agree, maybe they were reluctant to be part of a competition so soon. The ARM ecosystem is still evolving fast. Also there is still room for a lot of software and platform optimizations. Glibc got some performance improvements merged lately. ZLIB received some patches from Google some time ago but the merging process is very painful there.

ThunderX2 has been tested by ServeTheHome on server worklaods and by Bristol University people on HPC workloads

http://uob-hpc.github.io/assets/cug-2018.pdf