The Ampere Altra Review: 2x 80 Cores Arm Server Performance Monster
by Andrei Frumusanu on December 18, 2020 6:00 AM EST- Posted in
- Servers
- Neoverse N1
- Ampere
- Altra
SPEC - Single-Threaded Performance
Starting off with SPECint2017, we’re using the single-instance runs of the rate variants of the benchmarks.
As usual, because there are not officially submitted scores to SPEC, we’re labelling the results as “estimates” as per the SPEC rules and license.
We compile the binaries with GCC 10.2 on their respective platforms, with simple -Ofast optimisation flags and relevant architecture and machine tuning flags (-march/-mtune=Neoverse-n1 ; -march/-mtune=skylake-avx512 ; -march/-mtune=znver2).
While single-threaded performance in such large enterprise systems isn’t a very meaningful or relevant measure, given that the sockets will rarely ever be used with just 1 thread being loaded on them, it’s still an interesting figure academically, and for the few use-cases which would have such performance bottlenecks. It’s to be remembered that the EPYC and Xeon systems will clock up to respectively 3.4GHz and 4GHz under such situations, while the Ampere Altra still maintains its 3.3GHz maximum speed.
In SPECint2017, the Altra system is performing admirably and is able to generally match the performance of its counterparts, winning some workloads, while losing some others.
In SPECfp2017 the Neoverse-N1 cores seem to more generally fall behind their x86 counterparts. Particularly what’s odd to see is the vast discrepancy in 507.cactuBSSN_r where the Altra posts less than half the performance of the x86 cores. This is actually quite odd as the Graviton2 had scored 3.81 in the test. The workload has the highest L1D miss rate amongst the SPEC suite, so it’s possible that the neutered prefetchers on the Altra system might in some way play a more substantial role in this workload.
The Altra Q80-33 ends up performing extremely well and competitively against the AMD EPYC 7742 and Intel Xeon 8280, actually beating the EPYC in SPECint, although it loses by a larger margin in SPECfp. The Xeon 8280 still holds the crown here in this test due to its ability to boost up to 4GHz across two cores, clocking down to 3.8, 3.7, 3.5 and 3.3GHz beyond 2, 4, 8 and 20 cores active.
The Altra showcases a massive 52% performance lead over the Graviton2 in SPECint, which is actually beyond the expected 32% difference due to clock frequencies being at 3.3GHz versus 2.5GHz. On the other hand, the SPECfp figures are only ahead of 15% for the Altra. The prefetchers are really amongst the only thing that come to mind in regards to these differences, as the only other difference being that the Graviton2 figures were from earlier in the year on GCC 9.3. The Altra figures are definitely more reliable as we actually have our hands on the system here.
While on the AMD system the move from NPS1 to NPS4 hardly changes performance, limiting the Altra Q80-33 from a monolithic setup to a quadrant setup does incur a small performance penalty, which is unsurprising as we’re cutting the L3 into a quarter of its size for single-threaded workloads. That in itself is actually a very interesting experiment as we haven’t been able to do such a change on any prior system before.
Facebook
Twitter
RSS
148 Comments
View All Comments
Wilco1 - Saturday, December 19, 2020 - link
SMT gives very little benefit (only 15% faster on SPECINT and 3.5% *slower* on SPECFP), adds a lot of area and complexity, and results in very bad per-thread performance.It's always better to use 2 real cores instead of 1 SMT core. So if you have small cores like Neoverse N1, adding SMT makes no sense at all.
mode_13h - Sunday, December 20, 2020 - link
Whether SMT makes sense depends on the workload. Many tasks have greater branch-density and less computational density, in which case SMT is a massive win. I'm compiling code all the time and see huge speedups from SMT.That said, of course I'll take two real cores instead of 2-way SMT, all else being equal, but that always costs more. If SMT really made as little sense as you say, then it wouldn't be nearly so widespread.
Wilco1 - Monday, December 21, 2020 - link
There are certainly cases where SMT helps, but having some wins doesn't mean it is worth adding SMT. All too often people talk up the upsides and ignore the downsides. Let's see how Altra Max does vs Milan next year, that should answer which is best.Note almost none of the billions of CPUs sold every year have SMT (even if we exclude embedded). Adding another core is simpler, cheaper and gives more performance.
mode_13h - Monday, December 21, 2020 - link
> Let's see how Altra Max does vs Milan next year, that should answer which is best.I disagree. That's a bit apples and oranges. The differences between Zen2/Skylake and N1 are too big. You need to look at the overhead of SMT for a particular core vs. the benefits for that core.
These x86 cores are large not just because of SMT, but also the x86 tax, their wide vector units, and other things. It could be that SMT adds just 5% overhead, and that's not enough to increase your core count hardly at all, if you drop it.
Wilco1 - Monday, December 21, 2020 - link
It's never going to be a perfect comparison - nobody will design SMT and non-SMT variants of the same core! There are many differences because they use different design principles. However it will clearly show which of these designs works out best for top-end server performance.Spunjji - Monday, December 21, 2020 - link
SMT is a clear win where you already have large cores with a lot of execution resources, whereby the extra resources required aren't a large proportion of overall die area. It also helps if your tasks are focused on overall performance for a given number of threads, rather than performance-per-thread.Where the cores are this small, though, simply adding more of them seems to be the better option.
mode_13h - Monday, December 21, 2020 - link
Well said.mode_13h - Sunday, December 20, 2020 - link
Plus, per-thread performance is only an issue if that's how you're paying for CPU time, and then what you actually care about is thread performance per dollar, which would compensate for any cost differences due to SMT, as well.Wilco1 - Monday, December 21, 2020 - link
And Graviton 2 shows which is cheaper and faster.mode_13h - Monday, December 21, 2020 - link
That comparison is only valid for Amazon customers and in the short term. It can't be used to support a broader conclusion about SMT, because we lack transparency into the cost structure of Amazon's hosting, like whether they're subsidizing Graviton2 servers or even just charging enough for them to simply break even on the hardware.