Cazalan :
juanrga :
I wrote 'sponsors' not sponsors.
As mentioned before, there are three (3) Intel persons in his research group (one collaborator plus two students), not only one (1).
You're really grasping at straws here. The paper was written by university students, not Vertical Research group. They just credit Vertical Research group for helping review the paper.
Are you kidding? The paper is signed by three people and the senior author is Karthikeyan Sankaralingam, head of the Vertical Research Group, which is closely related to Intel.
Cazalan :
juanrga :
No sure why you mention Broadcom and Cisco. Do they have a historical record of dishonest competition/cheating? Also, do you believe that because there is people from other companies the possibility of aggressive 'sponsoring' is eliminated? AMD, Nvidia, and VIA were members of SYSMARK consortium and still Intel cheated the SYSMARK benchmarks for artificially increasing the score of its products.
Your reason for assuming the research group is biased is because there are some Intel people in it.
No. I started explaining first why the article was biased pro-Intel. Then driven by curiosity I did search info about the authors and found that are closely related to Intel.
You ignored the early part of my arguments, the part about the article defects, and you took only the last part of my arguments, that part about the authors, which you reproduced and replied in the first place, inverting my real argument and even the order of my message. I wonder why...
Cazalan :
juanrga :
If that article that you mentioned was right and Intel (x86) is able to compete with ARM face to face, then Intel had not been caught using its ICC compiler to cheat the Antutu benchmarks this year.
The cheated benchmark gave ARM processors a mere 50% upgrade in performance whereas Intel chips were seeing increases of up to 292%.
Now you're just going off on an ICC tangent, when the paper used GCC for compiling. Is GCC now biased for Intel too?
No. I have said you that _if_ the article was right and Intel did really compete with ARM, then Intel had not cheated benchmarks this year.
The article is biased pro-intel. And Intel couldn't compare its new chip against ARM under fair rules. That is why Intel did cheat Antutu benchmarks this year.
Are you kidding again? First I am not saying that any x86 maker cheats. I am saying that _Intel_ does. Therefore, if you have _some_ ARM maker that cheats report it, but not try a anti-ARM argument like that.
Second, that is Anandtech. The same site that uses biased benchmarks as SYSMARK for favouring Intel, the same site that has been for years announcing how Intel did beat ARM (but never did), the same site that has presented Intel mobile benchmarks this year, where the Intel chip was running inside a refrigerated room for allowing the chip to run benchmarks at maximum turbo freq. OK?
Third, the article says something completely different to what you pretend. It is not about cheating benchmarks code (as Intel did with Antutu), it is about if products run at turbo freq. when running benchmarks, or not.
Fourth, I can see lots of chipmakers with with a big "N" all the table cells. The "N" mean _NO_. Look at Nexus 4 or to Tegra 4 for example. They both play honest.
Cazalan :
juanrga :
I have submitted conference and contest papers a pair of months in advance. However, there are other objections to your hypothesis. The first, that they cite references published so late as 2013. The second, that they mention the A15 and even give some numbers for it. The third, that they claim that they evaluated but rejected more modern SoCs such as Tegra 3 before starting the lab work.
I'm sure they had their reasons.
I also. Tegra 3 is about 2x faster and about 40% more efficient, for instance. AMD bulldozer would rise power consumption and ruin the efficiency that they report for x86...
Cazalan :
The fact is they used parts that were comparable for the time frame. If they used an Ivy and just an old A8 you'd have a point, but they didn't. They used A8/A9, an old atom and an old i7 (Sandy).
Recall this was made at 2012. Tegra 3 was available and they rejected it.
The 'old' SB i7 is only slightly behind the modern IB i7, and this one is basically on tie with the last HW i7. On the other hand a A15 is _much_ faster and efficient than A9, and A57 is _much_ faster and efficient than A15.
E.g. the A15 is ~70% faster than A9, whereas reducing power consumption.
The conclusions in their paper are wrong.
Cazalan :
juanrga :
- They didn't test Bulldozer FX or similar ones but the best possible x86 chips.
Why would they bother? It is Intel's ISA after all. They also didn't use VIA chips which have been really low power for a while now.
They would bother because would ruin their conclusions about x86. And this would ruin Intel plans to sell x86 phones, servers, tablets...
Cazalan :
juanrga :
- Their compiler choices favoured Intel.
How does GCC favor Intel?
I didn't say that GCC favours intel. Their choices when compiled code did.
Cazalan :
juanrga :
- «they believe the results would look much the same — in terms of the relationship between power and performance» when looking at newer ARM designs, including the 64-bit ARMv8 architecture.
The only reference to ARMv8 is that the 64-bit parts are still under development. You're grasping at straws here.
We know that ARMv8 improves performance and efficiency. Their claim to __news__ sites that it doesn't matter reflect either ignorance or bias.
Cazalan :
juanrga :
- They have 3 Intel people in their small research group, but nobody from ARM or AMD.
All together implies bias pro Intel.
I'll say it again. The paper was written by 3 students.
Repeating it will not make it true. The paper was written by 3 authors one of them is Karthikeyan Sankaralingam, who is _not_ a student, but the head of a research group closely related to Intel.
Cazalan :
Google Android is what 99% powered by ARM cores? Broadcom makes ARM core parts. So exactly how would these astute people let Intel bias slide without question?
I already explained you how Intel cheated SYSMARK benchmark with AMD, Nvidia and VIA being part of the SYSMARK consortium. Therefore, your argument is lacking.
Cazalan :
juanrga :
Take the old Titan, double it, and you recover the #1 (approx.). But as explained to you before, this pure brute approach doesn't work for exascale supercomputing. You cannot take a x86-based supercomputer and scale it 1000x times because you cannot generate the immense power needed for that beast. That is why HPC experts are developing an ARM supercomputer for exascale.
You keep saying this 1000x times number but you're overlooking that there is already a 50 PetaFLOP (Peak) supercomputer. That means to get to an ExaFLOP you only need to get 20x faster than what is out there today, not 1000x.
Suddenly that number doesn't look so big anymore.
1000x is the scale factor between peta and exa. The goal is not to obtain one exaflop and stop there. The goal is to move the entire petaflop range of supercomputers to the exaflop range. This means 1 peta --> 1 exa and 50 peta --> 50 exa approx.
Cazalan :
juanrga :
Also you seem to not understand that the performance of the Xeon Phi has very little to do with x86, but is directly related to an _extension_ of the x86 ISA which is exclusive for Phi. Follow my advice, don't be fooled by Intel marketing ;-)
There's nothing to be fooled by. It's a parallel compute card. It's done differently than NVidia/AMD compute/graphics cards but that is how they chose to implement it. Intel is vertically integrated so they deemed that the best way to go.
And they're being aggressive with Phi by putting it on their 14nm node next year and leaving Haswell desktop at a 22nm refresh.
Now you look as an Intel representative ;-) but my point remains: x86 is a marketing trick for the Phi, because it doesn't run the same native x86 binaries that an 3770k, for instance, and because most (all) of its performance comes from using Intel Phi specific _extension_ of the x86 ISA.
Cazalan :
juanrga :
I doubt any supercomputer will be powered by Tegra 6. Who said you that?
You're telling me after all this time talking about Nvidia HPC and ARM supercomputers you suddenly forgot about Montblanc that was planning to use future Tegra mobile chips?
Nope. I have said you that Nvidia "project Denver" _core_ is aimed to supercomputers. No that Tegra 6 was.
You can have the same _core_ in different chips aimed to different markets. We will see phone chips with one or two A57 cores inside. But we will not see any phone with an AMD Seattle chi, despite Seattle uses A57 cores, because Seattle is for servers.
Montblanc project is using available mobile chips for prototyping. The final supercomputer will use ARM chips similar to the ARM servers, but _no_ mobile chips.
Facebook
Twitter
Instagram