Intel Z3770 geekbench sighted

[mikk]

I wonder if Stepping 3 is the final Silvermont Stepping. It could be considering that the first tablets should come in 2-3 months.

 

[bullzz]

anand has reported a "semi-official" cinebench score

http://www.anandtech.com/show/7263/intel-teases-baytrail-performance-with-atom-z3770-cinebench-score

 

[SiliconWars]

anand has reported a "semi-official" cinebench score

http://www.anandtech.com/show/7263/intel-teases-baytrail-performance-with-atom-z3770-cinebench-score

New thread - http://forums.anandtech.com/showthread.php?t=2340636

 

[Nothingness]

A Z3770 Geekbench 3 result is up: http://browser.primatelabs.com/geekbench3/compare/42322?baseline=52725

It looks like AES is the killer feature of BT Stream is strong too.

Overall good results, but not impressive.

 

[Khato]

A Z3770 Geekbench 3 result is up: http://browser.primatelabs.com/geekbench3/compare/42322?baseline=52725

It looks like AES is the killer feature of BT Stream is strong too.

Overall good results, but not impressive.

Eh, good versus impressive is entirely a question of how much power it uses to achieve that level of performance. Hopefully we'll finally get to see which is the case this week!

 

[Nothingness]

Eh, good versus impressive is entirely a question of how much power it uses to achieve that level of performance.

I definitely agree and in fact wanted to tell this in my previous post but kind of forgot. I should refrain from posting until I have enough coffee

Hopefully we'll finally get to see which is the case this week!

I hope we'll get figures not only from Intel, because I tend to dismiss marketing figures

 

[Arachnotronic]

A Z3770 Geekbench 3 result is up: http://browser.primatelabs.com/geekbench3/compare/42322?baseline=52725

It looks like AES is the killer feature of BT Stream is strong too.

Overall good results, but not impressive.

A fan-cooled, full-tilt quad A15 against a slim iPad-like tablet in a benchmark that is mostly a raw execution unit test...and Bay Trail still does very well.

The real question is power consumption...

 

[Nothingness]

A fan-cooled, full-tilt quad A15 against a slim iPad-like tablet in a benchmark that is mostly a raw execution unit test...and Bay Trail still does very well.

How do you know this wasn't run on a dev board with active cooling?

Oh well if you insist here it is against an S800-based *smartphone*: http://browser.primatelabs.com/geekbench3/compare/20688?baseline=52725

 

[SiliconWars]

Wonder how much AES is inflating the score by.

 

[Nothingness]

Wonder how much AES is inflating the score by.

This isn't difficult: the score is computed as geomean.

For the S800 above removing AES makes the score go up from 946.14 to 1158.85.

For the Z3770 the score goes from 1063.93 up to 1152.87.

S800 wins

EDIT: I only made the computation for single-thread.

 

[SiliconWars]

Yeah I sorta figured it might be, just wondered or hoped somebody else wasn't as lazy as I was lol.

Cheers. Yeah not looking great from a performance standpoint (still good of course) and as we know it's going to get crushed in graphics. Surely it must be quite a bit better in power terms, otherwise it best be prepared for a renaming to Fail Trail.

 

[Khato]

Oh well if you insist here it is against an S800-based *smartphone*: http://browser.primatelabs.com/geekbench3/compare/20688?baseline=52725

Thanks for the second comparison... mostly for the fact that it really makes me wonder how applicable to actual usage the averaged geekbench scores actually are. It's showing a pretty minimal difference between Shield and the Snapdragon 800 phone - 932 vs 926 single core, 2878 vs 2846 multi core. That's less than 2%. However the CPU benchmarks that Anandtech ran on both Shield and the Snapdragon 800 MDP (which should outperform most actual implementations) has the following - notation is Shield vs Snapdragon 800 (LG G2).

Sunspider: 536.3 vs 614.8 (931.0)
Octane: 4095 vs 2994 (2978)
Kraken: 7203.6 vs 8830.6 (10066.9)

It's far more of a delta than the less than 2% that Geekbench is reporting. While I recognize that the 'browser benchmarks' have their own flaws, they're at least indicative of one of the primary smartphone/tablet workloads. Whereas Geekbench looks to be about as 'useful' as synthetic benchmarks have been in the PC space - am I alone in paying zero attention to the wholly synthetic benchmarks (e.g. sysmark) in CPU reviews?

 

[Exophase]

While I recognize that the 'browser benchmarks' have their own flaws, they're at least indicative of one of the primary smartphone/tablet workloads. Whereas Geekbench looks to be about as 'useful' as synthetic benchmarks have been in the PC space - am I alone in paying zero attention to the wholly synthetic benchmarks (e.g. sysmark) in CPU reviews?

The stuff those browser benchmarks do is more synthetic than a lot of the Geekbench tests, which at least use real libraries used by real apps. There's no argument that compression/decompression and encryption are things actually done on phones and tablets. The FP stuff is less directly related but FP heavy code tends to fall into one of a few categories anyway, it's a lot less varied than integer code.

It's true one or two specific weaknesses in a CPU could cause every JS bench to suffer, since the nature of the JS JIT will have it run a lot of code that isn't really the meat of the computations expressed by the JS program (a lot of stuff for checking data types for instance). I give you that that works as a single relevant datapoint, but that's about it.

 

[Khato]

Agreed, and the problem with the current state of mobile benchmarking is simply that we're lacking other 'real use' data points. Which is a simple result of the fact that, well, what else is there? The majority of productivity-type applications are out of the picture, especially those that are actually CPU-bound. And then gaming introduces the issue of how to benchmark the CPU - the 3dmark physics test is the closest current benchmark I'm aware of. For some reason GLBenchmark has their offscreen 1080p setting to remove screen resolution from the equation, but they don't do an offscreen 240p setting or perhaps even lower to give an indication of where the CPU portion is in the equation.

I guess the real problem I see with the current incarnation of geekbench is that they're throwing a lot of individual scores of varying importance together into a simple average. I'll definitely agree that the individual scores can provide some amount of insight, but the lack of weighting results in a not very useful total score. (Especially if they're including memory performance in that average, unless somehow none of their other sub-tests are influenced by memory bandwidth.)

 

[Exophase]

At least Geekbench makes the subscores easily visible, and most of the time they get reported in reviews instead of just the aggregates.

One thing that could be tested that no one ever does is emulators. That would test a lot of aspects of CPU performance, especially for ones that don't use hardware rendering which is most of them. Load up some game, turn off throttling and frameskip, and see how fast it goes.. if there's no frame counter you can do a time of some game intros.. You do have to watch out for emulators that hand optimize for some platform more than another, though. Those would still be fairly good comparisons for platforms with the same CPU ISA.

It's not like no one uses them either, the most popular free ones have millions of downloads on Android, although they're not allowed on Apple's store. Maybe reviewers are afraid of pissing off game companies or worried they'll be doing something illegal even when testing games they own.

It'd be nice if they did SPEC too, but I guess that's too much work and too expensive for most to bother. That and you can't run it natively on Android, but there's ways to at least sort of run Linux on top of Android (chroot from an app), maybe those can run SPEC.

 

[Khato]

Agreed - it's clear that geekbench is trying to be a useful benchmark, and it does have relevant results. That doesn't keep some from just looking at the averaged score unfortunately.

Good point on emulators - they should indeed be a reasonably CPU-bound application with actual relevance. Granted it would have some analogs to the browser JS tests, but it's still every bit as applicable as they are and would provide another good data point.

Heh, I'm always curious about what SPEC scores would be for some of these mobile platforms, but I don't expect we'll ever get anything more than marketing materials that give some relative indication of SPEC performance. I tend to view spec as what geekbench wants to be, with the key differences being that it has a greater breadth in algorithms and balances them well - effectively a far more refined metric.

 

[Nothingness]

Regarding score aggregation SPEC does it too, but specialists look at individual scores. And I would even say that only gcc matters

The above JS scores show how odd JS benchmarking is, given how each gets the best score in turn. And geekbench certainly has too many compression algos (and anyone who claims jpeg is synthetic will have the right to read the description of the various algos ).

No perfect benchmark exists, many scores are needed to remove obvious mistakes/cheats (cf. the AnTuTu fiasco).

 

[VengenceIsMine]

The thing with a lot of the early engineering samples of Baytrail floating around is that Turbo isn't enabled and I have seen stepping 3 with turbo enabled but not a stepping 2 part. These scores are pretty inline with what I would expect from a 1.46 with no turbo enabled, turbo it up to 2.2 or 2.4 and these scores are much higher.

 

[liahos1]

The thing with a lot of the early engineering samples of Baytrail floating around is that Turbo isn't enabled and I have seen stepping 3 with turbo enabled but not a stepping 2 part. These scores are pretty inline with what I would expect from a 1.46 with no turbo enabled, turbo it up to 2.2 or 2.4 and these scores are much higher.

forgive my ignorance but the geekbench scores shown appear to indicate a ghz number inline with 2+. Are you referring to other benchmarks or does the ghz indication not necessarily mean turbo is enabled?

 

[VengenceIsMine]

forgive my ignorance but the geekbench scores shown appear to indicate a ghz number inline with 2+. Are you referring to other benchmarks or does the ghz indication not necessarily mean turbo is enabled?

The link I clicked on shows Intel Atom Z3770 @ 1.46 GHz
1 processor, 4 cores, which is the base clock. I don't see anything in the Geekbench score that indicates Ghz beyond that and an Intel chip with Turbo off will show up as the base clock in Geekbench.

 

[SiliconWars]

The link I clicked on shows Intel Atom Z3770 @ 1.46 GHz
1 processor, 4 cores, which is the base clock. I don't see anything in the Geekbench score that indicates Ghz beyond that and an Intel chip with Turbo off will show up as the base clock in Geekbench.

http://browser.primatelabs.com/geekbench3/compare/20688?baseline=52725

 

[Exophase]

The thing with a lot of the early engineering samples of Baytrail floating around is that Turbo isn't enabled and I have seen stepping 3 with turbo enabled but not a stepping 2 part. These scores are pretty inline with what I would expect from a 1.46 with no turbo enabled, turbo it up to 2.2 or 2.4 and these scores are much higher.

There have been a few posts already of scores from stepping 3 CPUs that are suitably higher than the scores posted back at the start of the thread.

 

[liahos1]

The link I clicked on shows Intel Atom Z3770 @ 1.46 GHz
1 processor, 4 cores, which is the base clock. I don't see anything in the Geekbench score that indicates Ghz beyond that and an Intel chip with Turbo off will show up as the base clock in Geekbench.

http://browser.primatelabs.com/geekbench3/compare/20688?baseline=52725

this link? that shows 2.39ghz? the no turbo'd results were all geekbench 2 no?

 

[Nothingness]

The thing with a lot of the early engineering samples of Baytrail floating around is that Turbo isn't enabled and I have seen stepping 3 with turbo enabled but not a stepping 2 part. These scores are pretty inline with what I would expect from a 1.46 with no turbo enabled, turbo it up to 2.2 or 2.4 and these scores are much higher.

So BT will be 4 times faster than z2580. Yeah sure

And if that indeed is a 1.4 GHz without turbo then the lack of MT scaling is terrible.

And last point, the report clearly states stepping 3. Did you see stepping 3 with no turbo?

 

[Nothingness]

Found Geekbench 3 scores on a French review: 972 ST and 2998 MT

Almost exactly the same as the result previously quoted: http://browser.primatelabs.com/geekbench3/52725

This means the results were obtained on basically an Intel demo tablet similar to what reviewers put their hands on, which undoubtly reaches 2.4 GHz.