anand's jaguar article

[RoarTiger]

Due diligence is about generating data firsthand.

It is taking a car for a test drive before buying it.

It is looking at a house with your own eyes, and paying for a home inspection report, before buying it.

It is getting referrals for a family doctor or dentist from credible individuals who have been customers in the past and that you also happen to personally know or trust.

Would you let a dentist for whom you have no firsthand knowledge perform dental surgery on your children? Not at all, you'd perform due diligence beforehand to vet the dentist. Make sure their academic credentials check out, that they are registered with the medical boards they claim to be registered with, etc.

It is fact checking, and if facts are lacking or come from questionable sources then you pursue avenues for generating facts you can trust.

Due diligence does not mean buying upfront and waiting to see if you develop buyers-remorse after the fact.

At the most basic level just find someone you trust and ask them to perform a test-drive for you with your software of choice with the hardware that you are interested in.

For example Aigo generated Nehalem data for me when I was thinking about buying a Nehalem (chose not to based on the results). Likewise Slowspyder generated data for me with both a Phenom and a Phenom II (again the results convinced me not to bother buying at the time).

Are you trying to change your own definition of due diligence? I know perfectly well the meaning. Previously you listed your own testing of a purchased item as an example. I can only assume that you intend to infer now that people should take a car for a "test drive" by purchasing it then dyno testing it and dissasembling it to test tolerences. Presumably before returning it for refund and then making their actual car purchase decision?

I questioned how due dilgence as cited by you pertained to ICC compiler issues when the information was not available. Is your due diligence defense of compiler issues that users should buy the CPUs and test this for themselves since the major tech news sites were not covering the issue? I just dont understand how you believe due diligence affects ICC issues when none of the major tech sites including this one cover the issue. Your new defense if I understand it isnt due diligence at all but asking users to purchase the items themselves and test. Maybe Intel and AMD will loan me the prerequisite hardware since the reviewers who actually do receive it free of charge will not report it.

 

[CTho9305]

IDC, that's a cop out. Most people don't have time to do full due diligence in every aspect of their lives. At best, interested people will check the ingredients labels on food; of those who do, few will spend the time to understand what each one means. Even fewer will understand the science, even if they try. Virtually none will take steps to verify the accuracy of the ingredients list (horse meat, melamine, etc). It's just not a realistic way to live. One can argue that an "implied warranty" shouldn't be a safe assumption, but I'm skeptical that someone who isn't attempting to rationalize something would make that argument.

The ICC web page just has a note at the very bottom "For more complete information about compiler optimizations, see our Optimization Notice.(<-was a clickable link)". The actual notice only includes one sentence that might raise a red flag to me if I were reading it without knowing the truth already, and then refers you to yet another external document (which isn't even linked to). It gets sleazier - the optimization notice is a PICTURE rather than text (presumably so Google and others don't index the text; it also lacks meaningful "alt" text). Be honest, how many people will even see the notice at the bottom of the first page (assume somebody told them, "try ICC, it produces fast code")? Of those who do, how many will assume it's just technical details about optimization that they don't need to worry about (I know I haven't dug into how gcc and llvm optimize code, I just know they do).

 

[Vesku]

Oh wow, at the GIF. Did they have a lunch meeting with Apple on that one? Seems they are both fans of making sure notices don't appear without scrolling down.

http://www.androidpolice.com/2012/1...msung-apology-behind-sneaky-uk-homepage-code/

If they just moved that "Certain optimizations not specific to Intel microarchitecture are reserved for Intel microprocessors." to the top and made the notice searchable text rather than an image, I'd give them a thumbs up even if they left the scroll down to the notice aspect on the main page.

More proof the FTC phoned this one in.

Again, the economic relationship isn't supposed to be entirely stacked against the consumer. If you pay to have your own independent home inspector examine a property but he misses foundation cracks because the owner actively patched them without disclosing their existence, well in most states that's not a "too bad your guy missed it, sucks to be you" situation. The seller is usually required to disclose any known major structural issues. So you can seek remedy if you've been hoodwinked. In the case of giant corporations that are hard for an individual to sue, that's a gap government oversight is supposed to fill in. It's also an area in need of major reform, imo.

Edit: Going to do my part to help Intel with their disclosure. Note regarding Intel compiler optimizations - Certain optimizations not specific to Intel microarchitecture are reserved for Intel microprocessors.

 

[monstercameron]

maybe this compiler debate needs it own thread...

 

[Cerb]

http://www.behardware.com/articles/847-15/the-impact-of-compilers-on-x86-x64-cpu-architectures.html

There are substantial differences. Take a look at the scores, though. The poor AMD CPU running, "crippled," code can still crunch on that data faster than the, "not crippled," GCC or MS. That 2nd to last page is also a great presentation, with the indexed scores, showing the indexed differences just between compilers and settings.

The dispatch stuff is something to know about, if you plan to use ICC, but it's nothing like the 386 v. P3/P4+SSE (that the Athlon XP could run just as well, if not for the dispatch check) of old, and, knowing about it, any sane developer would either go for compatibility right off (base x86-64 feature target only, FI), or do a few different builds and test them a bit (if the differences are substantial, the chances of that changing across anything less than a near-rewrite are slim, so it's not like it's going to need to be redone with every possible new build).

 

[Exophase]

http://www.behardware.com/articles/847-15/the-impact-of-compilers-on-x86-x64-cpu-architectures.html

There are substantial differences. Take a look at the scores, though. The poor AMD CPU running, "crippled," code can still crunch on that data faster than the, "not crippled," GCC or MS. That 2nd to last page is also a great presentation, with the indexed scores, showing the indexed differences just between compilers and settings.

The dispatch stuff is something to know about, if you plan to use ICC, but it's nothing like the 386 v. P3/P4+SSE (that the Athlon XP could run just as well, if not for the dispatch check) of old, and, knowing about it, any sane developer would either go for compatibility right off (base x86-64 feature target only, FI), or do a few different builds and test them a bit (if the differences are substantial, the chances of that changing across anything less than a near-rewrite are slim, so it's not like it's going to need to be redone with every possible new build).

This is a good example of the unfair auto-dispatcher not being used. Otherwise they wouldn't be able to highlight the differences between compiling targeting different instruction sets.

The funny thing is that they found the base Qax mode (auto-dispatch) is even generally slower on Intel processors. Which is another reason to avoid it altogether and use your own separate binaries.

 

[mrmt]

The big caveat is that Intel was forced to inform ICC customers about it's very strange way of deciding instruction set support when the FTC finally decided to look into the complaints from AMD, Via and Nvidia.

The FCC ordered Intel to tell everyone what should be a fair assumption to everyone: That ICC is biased.

Still, kind of silly to use a FTC mandated action as proof Intel plays nicely.

And when I did say that Intel plays nice? Quite the opposite, I've been saying here that Intel plays ruthlessly, that what they do which is good for them isn't always good for the consumer and that they *never deserve the assumption* they will be nice with you.

Maybe we are looking the things from two different POVs here. Maybe you are the vanilla customer seeing the world through pink lenses or an idealistic person like Agner Fog, who thinks the x86 standard should be steered like the Linux. I, OTOH, worked in banking and O&G, two business where cutthroat competition is a norm, and exclusivity/rebate deals like the ones Intel made with their customers are just the tip of the iceberg, and I'm not even talking about breaking the law here.

Where you see a smaller competitor being treated unfairly, I see a whiny company trying to compete in a game it cannot afford to.

 

[galego]

I'm not mixing up nor making a mess of anything. It doesn't matter if a benchmark is bad because it's unfair or bad because it's just bad. If it's bad it should be thrown out. Yet you automatically give it the benefit of the doubt while knowing nothing about it. It also doesn't matter if you're not taking it "as final." You were happy to use it to characterize performance without knowing a damn thing about it (including not knowing if it's fair or not) and are happy to happy to repeat it over and over again.

But unfortunately, as soon as the discussion turns to technical merit you lose all interest. I shouldn't even be arguing with someone who has no problem claiming that Jaguar has better per clock single-threaded performance than Ivy Bridge. :|

It matters! A fair benchmark is not benefiting one concrete branch (Intel) over the rest branches in any possible occasion. The biased benchmarks cited before make that.

It also matter that I did not take unixbench "as final" and clearly wrote that I would be waiting for more benchmarks to see if the performance of jaguar is that or is not. I am repeating this to you and you continue misinterpreting my position. That has a name.

 

[insertcarehere]

It matters! A fair benchmark is not benefiting one concrete branch (Intel) over the rest branches in any possible occasion. The biased benchmarks cited before make that.

What if, for example, Processor A has extensions which are supported by Benchmark X which make it faster than Processor B in that benchmark, even if Processor B has identical performance otherwise? Would benchmark X be a benchmark that you would classify as 'fair'?

 

[Exophase]

It matters! A fair benchmark is not benefiting one concrete branch (Intel) over the rest branches in any possible occasion. The biased benchmarks cited before make that.

It also matter that I did not take unixbench "as final" and clearly wrote that I would be waiting for more benchmarks to see if the performance of jaguar is that or is not. I am repeating this to you and you continue misinterpreting my position. That has a name.

If a benchmark is bad you don't use it, regardless of why it's bad. If you realize a benchmark is bad you retract what you said about it. You aren't doing this. You instead keep going on about things like benchmarks where FX-8350 is 42% faster than i7-3770K (I did a summary of all of Phoronix's benches in their article and never found anything that fast) and ignore all the ones where it's slower. You say other benhcmarks are cheating but you automatically assume that some totally anonymous score that you can't even cite a reference for is not just legitimate but also fair. You don't know this.

You never even clearly wrote anything about your anonymous unixbench scores not being final or applicable. This is all you ever said about it:

One single jaguar core in kabini gives slightly more performance per cycle than a SB core in an i3.

Unixbench single-core index
Jaguar (kabini) @ 2.0 GHz: 403.8
SB (i3-2100T) @ 2.5 GHz: 495.5

Look to the single-core benchmark given. 201.9 vs 198.2

It was the stuff about theoretical GFLOPS that you said wasn't final. Can't even keep your own arguments straight...

 

[Khato]

Had Intel not been found guilty during the antitrust cases someone like mrmt & a few others would have just said that "since it has not been proven in a court of law Intel is not guilty" so AMD/fanbois stop your whining !

And what court of law has found Intel guilty exactly? Last I checked there has yet to be a judicial ruling against Intel. I won't claim that Intel hasn't engaged in some questionable behavior, but when it comes to legality... well, I sure don't trust those with political motivations to make unbiased decisions in the matter.

 

[R0H1T]

And what court of law has found Intel guilty exactly? Last I checked there has yet to be a judicial ruling against Intel. I won't claim that Intel hasn't engaged in some questionable behavior, but when it comes to legality... well, I sure don't trust those with political motivations to make unbiased decisions in the matter.

I wouldn't go out of my way to prove something that is common knowledge, if you're arguing semantics i.e. "Intel not proven guilty in a court of law" then fine go ahead & you win but don't try to distort facts ! If however you're saying that Intel did not pay bribes to OEM's during the period in question, there were multiple cases filed against Intel including the ones in EU/US/Japan et al, then you're either ignorant, willfully or not is upto you to decide, or somehow lying to cover up the facts ! Now as far as the actual facts relating to the matter are concerned AFAIK they were never made fully public after AMD settled out of court so you'll have to contact either Intel or AMD to know the full extent of the malaise btw good luck with that :whiste:

 

[Phynaz]

That's the problem with people that say Intel broke consumer protection laws. If they had, the issue wouldn't go away just because the paid off (bribed) AMD to make the case go away.

 

[Idontcare]

I wouldn't go out of my way to prove something that is common knowledge, if you're arguing semantics i.e. "Intel not proven guilty in a court of law" then fine go ahead & you win but don't try to distort facts ! If however you're saying that Intel did not pay bribes to OEM's during the period in question, there were multiple cases filed against Intel including the ones in EU/US/Japan et al, then you're either ignorant, willfully or not is upto you to decide, or somehow lying to cover up the facts ! Now as far as the actual facts relating to the matter are concerned AFAIK they were never made fully public after AMD settled out of court so you'll have to contact either Intel or AMD to know the full extent of the malaise btw good luck with that :whiste:

What is wrong with bribery? We do it all the time, in spirit, only we don't like to call it what it is because it kinda takes the wind out of sails as hypocrites when we want to castigate others for doing it.

Just the other night I took my family to Red Robin for dinner. I had to pay my waiter a healthy 20% tip just to bribe them into giving me adequate service.

I just received my bribery payment the other day in the mail, my mail-in-rebate card from Newegg came!

We are a bribe-based culture. Is it so surprising that Intel's employees do it too?

 

[R0H1T]

That's the problem with people that say Intel broke consumer protection laws. If they had, the issue wouldn't go away just because the paid off (bribed) AMD to make the case go away.

What you;re saying relates to the US antitrust rulings wherein the case was filed by AMD, unlike the EU & Japan where it was the state vs Intel btw they were fined after being found guilty over there in each case ! I don;t won't to deal with this anymore than what has been debated already because the facts are out in the public domain however the extent of the damage or bribes is sealed in the court documents that I don't have access to !

 

[monstercameron]

What is wrong with bribery? We do it all the time, in spirit, only we don't like to call it what it is because it kinda takes the wind out of sails as hypocrites when we want to castigate others for doing it.

Just the other night I took my family to Red Robin for dinner. I had to pay my waiter a healthy 20% tip just to bribe them into giving me adequate service.

I just received my bribery payment the other day in the mail, my mail-in-rebate card from Newegg came!

We are a bribe-based culture. Is it so surprising that Intel's employees do it too?

and that justifies some of the things larger corporations do to smaller ones and the ecosystem as a whole?

 

[Khato]

What you;re saying relates to the US antitrust rulings wherein the case was filed by AMD, unlike the EU & Japan where it was the state vs Intel btw they were fined after being found guilty over there in each case !

I believe you mean South Korea, not Japan? The Japanese FTC simply ruled that Intel violated their antitrust regulations and ordered them to stop with no associated fine. The Korean FTC meanwhile issued a ~$26M fine to go along with their ruling. Note that AMD did file complaints against Intel in the US, EU, and Korea - such doesn't mean that the regulatory bodies of those countries weren't already investigating Intel's practices though.

 

[R0H1T]

I believe you mean South Korea, not Japan? The Japanese FTC simply ruled that Intel violated their antitrust regulations and ordered them to stop with no associated fine. The Korean FTC meanwhile issued a ~$26M fine to go along with their ruling. Note that AMD did file complaints against Intel in the US, EU, and Korea - such doesn't mean that the regulatory bodies of those countries weren't already investigating Intel's practices though.

Yes, but unlike in the US where it was a court ruling/proceeding that was initiated by AMD, the regulators in other countries found Intel guilty & made them pay a fine so AMD was not party to such investigations as in the US, besides this I don't know fully about the procedures involved in each separate case.

 

[crashtech]

What is wrong with bribery? We do it all the time, in spirit, only we don't like to call it what it is because it kinda takes the wind out of sails as hypocrites when we want to castigate others for doing it.

Just the other night I took my family to Red Robin for dinner. I had to pay my waiter a healthy 20% tip just to bribe them into giving me adequate service.

I just received my bribery payment the other day in the mail, my mail-in-rebate card from Newegg came!

We are a bribe-based culture. Is it so surprising that Intel's employees do it too?

It's really sad that you think this. I've gained a great deal from your posts, but it seems as if you've fallen off the rails when switching from facts to opinions.

First off, a waiter gets tips after delivering service; they are not asked to wield any influence beyond their own labor. So call it fee for service if you must, but it meets no definition of bribery.

Second, a rebate card is just a possible discount of a retail price. How is that possibly a bribe? It's an incentive to purchase a retail product, that's all.

Maybe you should research the definition of bribery before you throw around the term so casually. Bribery involves the use of illicitly delivered funds to gain favors from someone who can wield influence and affect outcomes in a non-meritocratic fashion. Your examples have nothing to do with this concept.

 

[krumme]

The FCC ordered Intel to tell everyone what should be a fair assumption to everyone: That ICC is biased.

Where you see a smaller competitor being treated unfairly, I see a whiny company trying to compete in a game it cannot afford to.

They why argue that Intel have any moral rights here? The Intel inside story about we used all those money, they didnt, therefore its our rights to it - is just pathetic , because Intel is just as dependent of AMD, NV, MS whoever to keep the x86 game going and earn the big money.

Its not like Intel is the big inventor of the fundamental things that they build their business on. They are just riping the fruit of others work, and brilliant individuals, exactly as the other big players do. Talk about whining. Who the beliewe that kind of storytelling? - its just making the company fail and fall a sleep.

Btw: The sound thing about Intel identity is they dont have this story about the hero saving the company (you know in reality a workaholic with asperger and loads of IQ)

 

[Lepton87]

IDC, that's a cop out. Most people don't have time to do full due diligence in every aspect of their lives.

I agree, most people don't even have enough time to do all of what IDC asks of them. Condoning bribes is something that would warrant it's own topic. If you did something wrong, a then bribe a judge to not suffer any consequences would that be fine? If bribes are nothing wrong then sure, that would be fine. And just some time ago IDC said it's not fine to send a CPU for RMA even if it was running slightly out of spec. You've got some strange morality. But enough with OT.
Is there any way to fool ICC compiled software to think that the code is running on an Intel CPU?

 

[Phynaz]

Yes, but unlike in the US where it was a court ruling/proceeding that was initiated by AMD, the regulators in other countries found Intel guilty & made them pay a fine so AMD was not party to such investigations as in the US, besides this I don't know fully about the procedures involved in each separate case.

So it's okay for bribes to take place as long as AMD is the one getting paid off?

 

[krumme]

I doubt the economic effect of fx. Intel bribing Dell in the p4 era, cheating and lying about the compiler stuff. I know several huge global companies where that behavior in no way would be tollerated, but be punished instead. And for good reason now;

In hinsight - Intel would have been far better off without doing the stuff. The effect have been marginal at best. The effect on the brand very minor at most, but inside the company you drain the energy and self esteem, and shift focus from what would happen in the market. You kept focus on AMD, you lost the focus of the mobile revolution that is over us now. It doesnt come free.

Intel have their claws on the OEM, the huge R&D and process machine. They could easily have avoided the tricks.

(the above dont apply to business with high levels of complexity fx. modern banking where the customers often dont know what they are buying, and the bankers themselves what they are doing. A cpu is in many ways a very, very simple product - and that includes all the tco talk)

 

[Exophase]

Is there any way to fool ICC compiled software to think that the code is running on an Intel CPU?

From what I can find, you can:

1) Configure VIA processors to report a different string. This may work through microcode patching. Unfortunately AMD has no equivalent (you'd think they would have done this precisely for this reason)
2) Run it in virtualization software, so that the hypervisor intercepts and emulates the CPUID instruction. But virtualizing can come with a performance penalty.
3) Use this utility to patch the dispatch code to check for AMD instead of Intel: https://github.com/jimenezrick/patch-AuthenticAMD But I think it only works for Linux compiled binaries.

Apparently ICC uses the dispatch on every function call, which sounds pretty bad. People should really avoid it...

 

[Lepton87]

From what I can find, you can:

1) Configure VIA processors to report a different string. This may work through microcode patching. Unfortunately AMD has no equivalent (you'd think they would have done this precisely for this reason)
2) Run it in virtualization software, so that the hypervisor intercepts and emulates the CPUID instruction. But virtualizing can come with a performance penalty.
3) Use this utility to patch the dispatch code to check for AMD instead of Intel: https://github.com/jimenezrick/patch-AuthenticAMD But I think it only works for Linux compiled binaries.

Apparently ICC uses the dispatch on every function call, which sounds pretty bad. People should really avoid it...

That's a bummer. They should have done what VIA did, that they didn't do it just goes to show how incompetent they really are. Are there any benchmarks for VIA CPUs with the default and the modified string? I'd be curious how much performance difference we are talking about. If it's 5% it's a completely different story then 40%.