Register for free to join our community of investors and share your ideas. You will also get access to streaming quotes, interactive charts, trades, portfolio, live options flow and more tools.
Register for free to join our community of investors and share your ideas. You will also get access to streaming quotes, interactive charts, trades, portfolio, live options flow and more tools.
Those SD835 Cinebench results look good to me: this certainly goes through
emulation, and given that it's very CPU intensive, it's a worst case. In
fact I'm surprised by how good that score is! I'd expect the native score to
be at least twice better, but we'll never know.
What this shows on the other hand is that if x86 compatibility of CPU
intensive programs is needed, Windows on ARM is surely not the way to go. For
other apps time will tell.
Heise.de has run SPEC CPU2006 using various compilers on both Epyc and Xeon (in German).
It seems people forget Intel had a very similar issue on Skylake.
Intel Skylake bug causes PCs to freeze during complex workloads
It was patched by a microcode update.
The news is up and it has nothing to do with Intel:
Also see how 176.gcc is 2000 for both A53 and A57, while it's 3000 and 4000 for 253.perlmbk. Statistically that's almost impossible to have 4 multiples of 1000
Definitely dubious results.
I hope my answer to Paul clarifies what I think.
To add one point: I don't care at all about Geekbench MT score. As you
wrote those 8 cores (or even 4 cores for phones) SoCs are dumb.
You'll understand that there are some things I can't obviously say due to
NDA.
That being said, I'm fully aware of Geekbench shortcomings as they are
pretty obvious:
- too many compression/decompression benchmarks
- too many benchmarks using dedicated instructions
- too small datasets.
I still find it useful as long as you examine individual scores (Lua and
Dijkstra are interesting IMHO) and keep in mind that drawing a conclusion
on a chip performance based on a single benchmark (no matter what that
benchmark is) is utterly dumb. I think John Poole is really trying to
improve his benchmark and accusing him of having an agenda isn't fair.
My understanding is that he's taking inputs from various companies for
GB4 and I guess our preferred one is being listened to.
As far as SPEC 2006 goes memory requirements did not allow Apple phones
to run it (64-bit requires 2GB as you know). Also one has to be careful
with SPEC, given that many companies have dedicated huge efforts to tune
their compilers (icc on libquantum is the obvious example), so to use it
to compare different CPU's it would be fair to use gcc. Perhaps we'll
see someone qualified do that at last (not a random Anandtech guy
unable to properly compile SPEC as was the case in the Exynos 7
review...).
So you won't read me saying how great Geekbench is, but you won't read me
saying it's a PoS. It provides some useful information. Just one data
point, far from enough to draw any conclusion about an alleged catch up
of Intel by Apple.
I don't need input from Linus or any other about Geekbench. As a
benchmarking engineer working in a company that bought a license, I have
access to the source code
I think you're being a little too hard at Geekbench author. Here is what
he wrote on Realworldtech:
In our testing the performance difference between the mobile and the desktop workload sizes is less than 1% on the same hardware.
The two datasets was the only way we were able to produce a benchmark that ran on both desktop and mobile and had a reasonable runtime on both given the huge delta between mobile SoCs and desktop CPUs. This delta has decreased dramatically and so Geekbench 4 will remove the distinction between mobile and desktop workload sizes.
Let's give him the benefit of the doubt.
You still can't deny it was partly broken by several compiler vendors. That's
the price to pay for the source being available
Also only relying on SPECCPU alone is very misleading. As an example, the
Icache footprint is ridiculous, much lower than a simple browser benchmark.
There are much better Snapdragon 810 Geekbench MT score close to 5k: http://browser.primatelabs.com/geekbench3/1882896
No matter how bad AnTuTu is (and it really is bad, even Geekbench is
better), the scores posted by HotHardware make little sense: see how
the Z3770 tablet is significantly worse for CPU Int than then Dell Venue
while it is much better for CPU Float, even though the cores are the same.
Also other sites have reported >55K AnTuTu scores for NVIDIA SHIELD Tablet.
IMHO AnTuTu definitely is a useless benchmark.