Bay Trail benchmark appears online, crushes fastest Snapdragon ARM SoC

[TuxDave]

How do you figure? Even full OoOE doesn't need register renaming. It's just there to alleviate false dependencies and increase available ILP.

Good point. You can technically get away with no register renaming on an OoOE CPU and take the big performance penalty of serializing false dependencies.

 

[Exophase]

Good point. You can technically get away with no register renaming on an OoOE CPU and take the big performance penalty of serializing false dependencies.

Right, and IMO the number of false dependencies that'd be resolved between an FALU + FADD queue and an all-other-SIMD queue is not enough to warrant renaming.

It could still have it; David Kanter has reported other information that goes beyond what's reported in the manual (and is probably not based on misunderstandings), although he didn't report that the FP queues were in-order. It just seems odd to me that Intel would use PRF-based renaming only for this purpose, given the extra logic that has to be there to recover it from false speculation.

 

[Nothingness]

I guess the SunSpider test mentioned here (direct link to video) with a very disappointing 958 ms result is as "interesting" as the AnTuTu that we keep on hearing. At least it offers a counterview... or not.

I wonder if that even was a Bay Trail at all. I guess we should get ready for many fakes, FUD, and other BS for weeks to come, until real devices reach unbiased reviewers.

 

[Arachnotronic]

I guess the SunSpider test mentioned here (direct link to video) with a very disappointing 958 ms result is as "interesting" as the AnTuTu that we keep on hearing. At least it offers a counterview... or not.

I wonder if that even was a Bay Trail at all. I guess we should get ready for many fakes, FUD, and other BS for weeks to come, until real devices reach unbiased reviewers.

Considering Clover Trail outperforms this Bay Trail result, it makes sense that something wasn't exactly right here.

 

[Exophase]

Considering Clover Trail outperforms this Bay Trail result, it makes sense that something wasn't exactly right here.

Score seems believable to me for a ~1.1GHz Silvermont. That's 22% better clock normalized vs the CT score. For anyone expecting > 50% IPC on every last Silvermont test I have bad news for you.

What, you mean you guys still think this is a 1.1GHz part that was actually running at 2GHz while doing tests? Come on think about it for a second. Engineering samples running at much lower speeds is nothing new. Why would it have such a low base speed but an unrestricted turbo? This particular sample could have been less than 1.1GHz (or more). Of course there are also other possible explanations, the score could be bad for other reasons. And of course Sunspider is a terrible benchmark.

BTW, listing your Nexus 7's score in that post seems like a pretty deliberate misdirection. You could have given what the same AT review gives the same Tegra 3 hardware (maybe slightly different clocks) when running on the same OS/browser the Atom numbers are using.. a score over half as high as your Nexus 7 score..

(also who rounds 714.9 to 714, sloppy )

 

[Nothingness]

Score seems believable to me for a ~1.1GHz Silvermont. That's 22% better clock normalized vs the CT score. For anyone expecting > 50% IPC on every last Silvermont test I have bad news for you.

According to this video at 50 s, the CPU is reported as 1.66 GHz. Granted it doesn't prove the chip was running at that frequency.

And of course Sunspider is a terrible benchmark.

Definitely terrible, as much as AnTuTu. If it had been a longer/better JS benchmark, I would have countered your claim that a 22% improvement was good, given that I'd expect JIT to benefit from OoOE more than many benchmarks.

 

[Face2Face]

I hope to see some Baytrail tablets @ $499 with Windows Pro to replace my wife's iPad.

 

[Arachnotronic]

Score seems believable to me for a ~1.1GHz Silvermont. That's 22% better clock normalized vs the CT score. For anyone expecting > 50% IPC on every last Silvermont test I have bad news for you.

Heh, fair enough!

What, you mean you guys still think this is a 1.1GHz part that was actually running at 2GHz while doing tests? Come on think about it for a second. Engineering samples running at much lower speeds is nothing new. Why would it have such a low base speed but an unrestricted turbo? This particular sample could have been less than 1.1GHz (or more). Of course there are also other possible explanations, the score could be bad for other reasons. And of course Sunspider is a terrible benchmark.

I would really expect that this close to launch that the company would have full-speed parts ready to go...

BTW, listing your Nexus 7's score in that post seems like a pretty deliberate misdirection. You could have given what the same AT review gives the same Tegra 3 hardware (maybe slightly different clocks) when running on the same OS/browser the Atom numbers are using.. a score over half as high as your Nexus 7 score..

*goes and checks what the AT review gave that hardware*

Huh...981ms on WinRT for a Tegra 3, but an Exynos 5 dual (much more capable A15's) get 1384ms on Android/Chrome.

Interesting. That would suggest that a 1.1GHz BYT is roughly on par with a 1.3GHz Tegra 3/Cortex A9.

(also who rounds 714.9 to 714, sloppy )

Heh, probably should have rounded to 715 rather than truncating!

 

[Arachnotronic]

I hope to see some Baytrail tablets @ $499 with Windows Pro to replace my wife's iPad.

Here's an iPad clone w/ Baytrail running Windows 8 shown off at Computex:

Let's hope both performance and battery life are good, because I'd definitely be interested in a Windows Pro iPad-clone.

 

[Exophase]

According to this video at 50 s, the CPU is reported as 1.66 GHz. Granted it doesn't prove the chip was running at that frequency.

I think Windows gets the frequency from CPUID brand string which doesn't really mean that much in this case.

Definitely terrible, as much as AnTuTu. If it had been a longer/better JS benchmark, I would have countered your claim that a 22% improvement was good, given that I'd expect JIT to benefit from OoOE more than many benchmarks.

That's true, all but a very aggressive JIT (of which I really doubt any JS one would qualify) would benefit more from OoOE than native code due to inferior compiler scheduling. But with JS in general (not just SunSpider) there are all sorts of other places to lose performance..

I would really expect that this close to launch that the company would have full-speed parts ready to go...

That laptop shown early June is like 6 months prior to launch at least, that doesn't qualify as close to me at all. The hardware they used to build it could have been pretty much any amount older.

If this were a month before launch I'd be more inclined to agree but it isn't. Go look at what you could find for Ivy Bridge or Haswell a few months before launch. If Intel really has a near-production level laptop they should send it out to sites like AT for preview testing.. and they should be releasing it a lot sooner.

 

[blackened23]

If this were a month before launch I'd be more inclined to agree but it isn't. Go look at what you could find for Ivy Bridge or Haswell a few months before launch. If Intel really has a near-production level laptop they should send it out to sites like AT for preview testing.. and they should be releasing it a lot sooner.

IIRC, all of these products are being delayed for a month or so to coincide with the release of Windows 8.1. I've been using the (8.1) beta a bit and it is a much more compelling OS than was Windows 8, I think most will be pleased with it - and due to the lackluster response windows 8 has had, most PC makers are less than enthusiastic about releasing products using it. Especially when 8.1 is so close. For instance, the macbook air has been released for some time, but PC makers are just waiting for the back to school season and windows 8.1 release - Windows 8.1 also has more power saving features included for Haswell too IIRC.

 

[SiliconWars]

http://www.eetimes.com/author.asp?section_id=36&itc=eetimes_sitedefault&doc_id=1318857&page_number=2

Going from the 2.9.3 version of the AnTuTu benchmark to the 3.3 version, the overall AnTuTu scores increased 122 percent, and the RAM score increased 292 percent for the Intel processor, while the scores for the Samsung processor increased only 59 percent and 53 percent, respectively.

 

[Exophase]

IIRC, all of these products are being delayed for a month or so to coincide with the release of Windows 8.1. I've been using the (8.1) beta a bit and it is a much more compelling OS than was Windows 8, I think most will be pleased with it - and due to the lackluster response windows 8 has had, most PC makers are less than enthusiastic about releasing products using it. Especially when 8.1 is so close. For instance, the macbook air has been released for some time, but PC makers are just waiting for the back to school season and windows 8.1 release - Windows 8.1 also has more power saving features included for Haswell too IIRC.

Why would that impact Silvermont products that were never slated to be out before Q4 2013 at the earliest? The last roadmap I've seen lists Valleyview-M products for February 2014, do you have something different? It also happens to list engineering samples as June-August 2013, interesting isn't it..

 

[Sweepr]

Why would that impact Silvermont products that were never slated to be out before Q4 2013 at the earliest? The last roadmap I've seen lists Valleyview-M products for February 2014, do you have something different?

 

[Exophase]

BayTrail-M isn't BayTrail-T. Would like to know where that little snippet is from.. what is it with people cutting out pieces of slides these days? Context is great.

 

[Nothingness]

http://www.eetimes.com/author.asp?section_id=36&itc=eetimes_sitedefault&doc_id=1318857&page_number=2

Going from the 2.9.3 version of the AnTuTu benchmark to the 3.3 version, the overall AnTuTu scores increased 122 percent, and the RAM score increased 292 percent for the Intel processor, while the scores for the Samsung processor increased only 59 percent and 53 percent, respectively.

If only that could put a stop to usage of AnTuTu in its current incarnation

 

[Vesku]

From what I have read, including links provided, in this thread AnTuTu should be boycotted by review sites. A big part of synthetics is consistency.

 

[wlee15]

I think Windows gets the frequency from CPUID brand string which doesn't really mean that much in this case.



That's true, all but a very aggressive JIT (of which I really doubt any JS one would qualify) would benefit more from OoOE than native code due to inferior compiler scheduling. But with JS in general (not just SunSpider) there are all sorts of other places to lose performance..



That laptop shown early June is like 6 months prior to launch at least, that doesn't qualify as close to me at all. The hardware they used to build it could have been pretty much any amount older.

If this were a month before launch I'd be more inclined to agree but it isn't. Go look at what you could find for Ivy Bridge or Haswell a few months before launch. If Intel really has a near-production level laptop they should send it out to sites like AT for preview testing.. and they should be releasing it a lot sooner.

I believe that the current and max processor are stored in the SMBIOS tables so I would assume they're correct in the absence of any other information.

 

[Idontcare]

From what I have read, including links provided, in this thread AnTuTu should be boycotted by review sites. A big part of synthetics is consistency.

I'm trying to put this AnTuTu situation into context for my own digestion, so lets just say (hypothetically of course ) that I am an aging simpleton who doesn't quite fully grasp what all you young energetic kids are going on about here, if I were to say to myself "hmmm, this AnTuTu benchmark situation sounds awfully reminiscent of Sysmark" would that be too far off the mark?

 

[Arachnotronic]

I'm trying to put this AnTuTu situation into context for my own digestion, so lets just say (hypothetically of course ) that I am an aging simpleton who doesn't quite fully grasp what all you young energetic kids are going on about here, if I were to say to myself "hmmm, this AnTuTu benchmark situation sounds awfully reminiscent of Sysmark" would that be too far off the mark?

Not at all.

Aging simpleton...heh, you're one of the sharpest tools in the AT shed, IDC!

 

[Exophase]

I'm trying to put this AnTuTu situation into context for my own digestion, so lets just say (hypothetically of course ) that I am an aging simpleton who doesn't quite fully grasp what all you young energetic kids are going on about here, if I were to say to myself "hmmm, this AnTuTu benchmark situation sounds awfully reminiscent of Sysmark" would that be too far off the mark?

I'm having a hard time finding out exactly what the deal was with Bapco outside of the companies leaving it in disgust. It appears the biggest complaint was that the benchmark doesn't do anything GPU-related.. if that's really the biggest problem with it then I'd say that no, it's not similar to AnTuTu.

The other side of this is that I think that the typical PC benchmarking suite is like an order of magnitude (at least!) better than what you see on phones. The basic impression I'm getting is that sysmark was a crusty old synthetic benchmark that was never very good and looked worse and worse compared to more representative benches. In other words, it sounds like most mobile benches today

AMD outright said that they weren't going to endorse synthetics tests anymore and I don't know nVidia would have ever been interested in something with no GPU component. If I'm missing some more devious scandals or other shenanigans that were uncovered hopefully someone has some links..

 

[Vesku]

I'm trying to put this AnTuTu situation into context for my own digestion, so lets just say (hypothetically of course ) that I am an aging simpleton who doesn't quite fully grasp what all you young energetic kids are going on about here, if I were to say to myself "hmmm, this AnTuTu benchmark situation sounds awfully reminiscent of Sysmark" would that be too far off the mark?

I don't know if Intel has a relationship with the AnTuTu developers, that would bring Sysmark parallels. I was referring more to the drastic change in scores for the same Atom chips between versions. Looks like Exophase did some real digging into it in his Antutu post http://forums.anandtech.com/showthread.php?t=2330027

Using ICC optimizations for Intel but not having a NEON codepath for ARM, I'd call that a pretty poor design decision. I'd also say that diminishes its viability as a multiple ISA benchmark.

I'm having a hard time finding out exactly what the deal was with Bapco outside of the companies leaving it in disgust. It appears the biggest complaint was that the benchmark doesn't do anything GPU-related.. if that's really the biggest problem with it then I'd say that no, it's not similar to AnTuTu.

I believe you brought up how benchmarks can be gamed by changing the weighting of individual tests. So if 1A+1B doesn't show off what you want it to you use 1A+0.8B or some such. Well that's what the other participants of Bapco were accusing Intel of manipulating and the organization not introducing GPGPU tests they wanted was the final straw for them.

An interesting historical note, AMD joined Bapco to try to have more influence on the benchmarks after they gave a pretty good claim (PDF link) for choices transitioning from Sysmark 2001 to 2002 being weighted in favor of the Pentium 4. Keep in mind this is a benchmark suite that has been referenced for government procurement so there was a lot of incentive to join it rather than ignore it. It seems Intel never gave up the head position at Bapco though, which probably helps explain the eventual abandonment by AMD, Nvidia and Via.

 

[Exophase]

I believe you brought up how benchmarks can be gamed by changing the weighting of individual tests. So if 1A+1B doesn't show off what you want it to you use 1A+0.8B or some such well that's what the other participants of Bapco were accusing Intel of manipulating.

That was galego

I actually thought he was talking about changing the weighting dynamically depending on which processor is running the code. That would be unquestionably illegitimate.

But sometimes weighting really is bad and should be changed. I'm thinking of PassMark's integer performance test, that does one addition, subtraction, multiplication, and division in a loop. In other words, it was basically an integer division test, since integer division is MUCH slower than those other things in any modern moderately clocked general purpose processor. Integer division is also used far less commonly in real code. So it was really, really unrepresentative of actual integer performance and one CPU improving their hardware divider totally skewed the test.

That would be a good case of a test that should be rebalanced (or ideally, replaced with a different test that wasn't hilariously bad)