AMD vs Intel benchmarks for latest chips

John at InsideHPC has a pointer to an article on benchmarks of the chips. There is no doubt that Intel is doing a good job on coming out with chips in a timely manner, something AMD is not doing well.
Regardless of my criticism, what is interesting are the real world tests. I don’t care so much about winrar and other things that, generally speaking, won’t impact my or my customers lives all that much.
I am a great deal more interested in applications benchmarks with real data sets that my customers run. Few run Linpack. Quite a few run LSDyna, and Fluent, and …

What is interesting is what is on the Intel site.
Floating point throughput (picture links back to their page, and picture comes from their page).


Wow. If I believe this (no reason not to) then the 2350 and 2360 are (ignoring press hype to the contrary) quite formidable competitors to Intel’s offerings. Early benchmarks we have seen with CFD and other codes support this contention. Note that this flies in the face of the “benchmarks” that other people do. Not end user HPC application benchmarks.
So lets look at Dyna:

Note: vertical axis is wall clock. This is the topcrunch benchmark.
This pretty much explains why Penryn was pushed out early. Sadly, there aren’t other relevant benchmarks there.
The take home message is that a 2 GHz Opteron 2350 appears to seriously best a 3 GHz Xeon 5365 on a real test (about 22% faster in favor of the Barcelona). Moreover, the 2 GHz 2350 is about 6% slower than the 3 GHz Penryn in the worst case, and about the same in the best case.
Now compare that data to the hype machine screaming at us that Penryn is destroying Barcelona.
This is not a rout by any sense of the word. Had Intel not come out with Penryn when it did, it would not be winning these benchmarks, and as you can see, it isn’t by much.
That said, the integer improvements in Penryn are quite exciting. I want to see what they will do for codes like BLAST. My “tomato_15min.fsa” seems to have taken on somewhat of a standard input deck status (which is funny as it is file of A. thaliana which I mislabeled some years ago … go figure) so I will use that and the latest nt and try to get my hands on a harpertown for testing.

3 thoughts on “AMD vs Intel benchmarks for latest chips”

  1. Nice numbers. And you are right …. everyone seems to be saying AMD is dead and spreading the FUD. I am sure you would have already seen the patchy first review and a better (still not good) one by AnandTech.

  2. A good reminder also of the difficulty in answering that general “Which one is faster?” question. A surprising number of people really want there to be one answer to that question and will make huge generalizations (“X was faster in this FPS therefore my lab should have bought those”) without doing any testing.
    We’ve certainly run into this since one of our main applications is a monte-carlo simulation which needs FPU and low-latency random memory access; AMD completely ruled the roost there for a long time until the Core2 chips came out – at that point the Core2 was faster for small, single jobs and the Opterons stayed ahead for bigger jobs and concurrency, meaning that the “best” system even depended on whether someone was a developer (Core2 favors lots of builds, no more than one running job) or user (1 job per core favors AMD’s non-shared memory subsystem).
    In practice, however, performance was close enough that the decision was based on price and manageability – things like Sun’s non-garbage service processors and better pricing would give the nod over a Dell unless we had a significant performance gap. That’s the under-covered story here: right now it’s hard to lose on a CPU purchase but if the Opteron hadn’t existed we’d all be paying a lot more for slower Intel CPUs (Itanium) and antiquated system designs (how long would Intel have taken to leave a shared FSB if HyperTransport hadn’t demonstrated the advantages?).

  3. well … thats what they say … competition brings advantages to the end users. and what you said is true … there is no simple answer. Know your app profile and run some tests to find out which one works for your need.

Comments are closed.