Intel Z3770 geekbench sighted

[Nothingness]

The Antutu benchmarks comparing Baytrail clocked at 1.4GHz (45,000) to Snapdragon 800 (30,000) i.e. the best SoC from ARM camp this year, shows an impressive lead of Baytrail over ARM, even when using an UNDERCLOCKED Baytrail

Later this year, we'll see the brainless tech blogging sheeps telling the exact opposite story from last year's "Intel is dead, ARM is king" nonsense to "Intel makes a comeback, challenge the ARM's throne". Go Intel!!!

Some replies are so funny they deserve to be quoted :biggrin:

Now I hope this was supposed to be humourous... Tell me it is!

 

[Exophase]

Way too low scores for the Z3770 imho. This CPU should easily crush a 2 core 1,0 Ghz Temash.

1.46GHz, not final clock speeds..

I know some people think that it's turboing up to full speed but that makes no sense to me, engineering samples have lower clock speeds because they don't want you to run at those clock speeds, not because they're trying to keep power consumption very low.

 

[Arachnotronic]

Either Geekbench is screwed up, or Intel really screwed up and underestimated the competition with Silvermont, especially on floating point.

Wouldn't doubt either at this point; smells a lot like Bulldozer at this point, but I hope I'm wrong and that Intel isn't foolish enough to put out an incredibly weak processor when it needs leadership just to break into this new industry.

 

[shady28]

The Antutu benchmarks comparing Baytrail clocked at 1.4GHz (45,000) to Snapdragon 800 (30,000) i.e. the best SoC from ARM camp this year, shows an impressive lead of Baytrail over ARM, even when using an UNDERCLOCKED Baytrail

Later this year, we'll see the brainless tech blogging sheeps telling the exact opposite story from last year's "Intel is dead, ARM is king" nonsense to "Intel makes a comeback, challenge the ARM's throne". Go Intel!!!

Not sure where you're drawing that conclusion based on the benchmarks the OP linked.

In those, the Samsung GT-I9500 (Exynos Octa) is whipping Bay Trail across maybe 80% of the benchmarks. Mostly by small margins - but in some cases by very wide margins (esp in FPU).

I would point out, the Exynos Ocata is here today, in a phone that you can buy. Bay Trail is not here yet, and when it is here it isn't going to be a in phone it'll be in a tablet.

I'm not a cheerleader for ARM or intel, but I am looking at getting a Win8 tablet - was considering waiting for Bay Trail - and those benchmarks don't look good.

The blue bars at this link are where the Z3770 lost that benchmark vs Octa :

http://browser.primatelabs.com/geekbench2/compare/2166084/2068301

 

[Exophase]

Either Geekbench is screwed up, or Intel really screwed up and underestimated the competition with Silvermont, especially on floating point.

Wouldn't doubt either at this point; smells a lot like Bulldozer at this point, but I hope I'm wrong and that Intel isn't foolish enough to put out an incredibly weak processor when it needs leadership just to break into this new industry.

Definitely wouldn't read a lot into the FP scores. I was already mostly ignoring those for ages.. like you showed, they're suspicious against other platforms too. And it's not just a subset of tests. Even if they're legitimate they're pretty micro-benchy and they're not being compiled to very good code.

The integer code is more representative (but not of everything, not even close!) because it runs actual libraries and VMs that are in use. But even that relies on the input data set being a good representation. I have my doubts about the program they're using for the Lua test..

But despite all of this geekbench is still one of the best mobile benchmarks in use Here's hoping version 3 is a good update.

Anyway, the numbers for the integer part aren't bad at all. perf/MHz may be a little lower than A15 in some cases (although it's higher in others!) but that's not at all a bad position to be in if they're really using substantially less power while clocking even higher. Let's take a massive leap and say the integer part really is representative of performance across the board - if they can take this and hit 2GHz at ~750mW/core or less then they'll be in a very strong position. And there are some people who think the power consumption will be even lower than that; this I will need to see for myself.

I think what could end up holding Intel back the most is not the broad engineering side of things but the area constraints they're working in. They'll probably be spending a lot less die area on GPU than some of their competitors (at least Qualcomm and most likely Apple) and I suspect they'll be paying for it. There's also the matter of using only 2 cores in their phone parts when a lot of their competitors will be using 4.

 

[Khato]

Either Geekbench is screwed up, or Intel really screwed up and underestimated the competition with Silvermont, especially on floating point.

Such is only the case if the reported clock speed/expected Z3770 clock speeds are not correct. Assuming that the reported clock speed is correct then Silvermont is looking pretty good - comparable integer IPC to A15 despite only being a 2 issue architecture? Sure geekbench reports it as being notably slower when it comes to FP performance, but how much does that matter in its intended markets exactly? What FP-heavy workload is going to be executed on these devices with the expectation of high performance? Especially if the software side can stop slacking and allow more of those sort of workloads to execute on GPU resources. Basically, the FP performance looks bad in comparison, but how much of a difference will that make in practice? It appears to be fast enough that the occasional FP instructions aren't going to bog down a program.

It all comes down to frequency and power. If a Silvermont core running at 2.4 GHz uses the same or less power executing the same workload as an A15 running at 1.6 GHz then Intel has a solid 1.4x lead in integer. Whereas if Silvermont power consumption is only comparable at same frequency then yeah, Intel wouldn't have a very convincing product and certainly would be delivering on its promises.

 

[shady28]

..
It all comes down to frequency and power. If a Silvermont core running at 2.4 GHz uses the same or less power executing the same workload as an A15 running at 1.6 GHz then Intel has a solid 1.4x lead in integer. Whereas if Silvermont power consumption is only comparable at same frequency then yeah, Intel wouldn't have a very convincing product and certainly would be delivering on its promises.

The Octa can go from 200Mhz up to 1.8Ghz (about 35% faster than the one in the OP's link).

You're unlikely to see an Octa at 1.8 Ghz, and equally unlikely to see a Z3770 at 2.4Ghz.

What we see with existing Octa / Clovertrail chips is far more likely : a 1.4Ghz Octa, and a 1.8 Ghz Z3770.

I did a bit more research and I see the Z3770 is a significant improvement over existing Z2760 parts though, about 60% faster per mhz :

http://browser.primatelabs.com/geekbench2/compare/2068301/2167566

It does well against the Snapdragon 600 :

http://browser.primatelabs.com/geekbench2/compare/2167998/2068301

But as we saw not so well against the Snapdragon 800 aka 'Octa'.

 

[zlatan]

Whats with the single threaded FP scores?? At around 50% higher clockspeed it is lower than Kabini in most of the tests!!

The Jaguar core is an FP "monster". Even support AVX. I think AMD was do this for the consoles. On the other hand Intel didn't really speed up Silvermont on the FP side, compared to Saltwell.

 

[sweatbysd]

I suppose the use of the number 3770 depicts the top of line baytrail sku

 

[NTMBK]

Jaguar has 128-bit wide FPUs, and Silvermont doesn't. Not really sure why it's a surprise that it has considerably higher floating-point IPC.

 

[pelov]

Jaguar has 128-bit wide FPUs, and Silvermont doesn't. Not really sure why it's a surprise that it has considerably higher floating-point IPC.

This bugs the crap out of me.

While forum-goers complain about the lack of software utilizing newer ISAs on brand new shiny silicon, the x86 landscape is fragmented with respect to features (virtualization, TSX) and ISAs-extensions (AVX, SSEx.xx before that) and makes having any sort of consistent set of software that stretches across devices nothing but a pipe dream.

I understand that it's not worth the die space and added cost and power on the Atoms, but because the vast majority of consumer-oriented development occurs for the lower end chips - and the lowest common denominator - we're all getting screwed at the top.

:|

 

[insertcarehere]

• Intel Atom Z3770 1.46GHz vs. VIA QuadCore L4800 1.46+GHz (1.6GHz capable)
source: http://browser.primatelabs.com/geekbench2/compare/2068301/2124912


• AMD A4-5000 1.5GHz vs. VIA QuadCore L4800 1.46+GHz (1.6GHz capable)
source: http://browser.primatelabs.com/geekbench2/compare/2139838/2124912

The Atom is doing pretty well for itself IPC wise as a 2-wide architecture (not even fuly Ooo) against the three-wide architectures from VIA and AMD, considering that the VIA is a 27.5W TDP processor w/o igpu (40nm).

 

[USER8000]

The Jaguar core is an FP "monster". Even support AVX. I think AMD was do this for the consoles. On the other hand Intel didn't really speed up Silvermont on the FP side, compared to Saltwell.

I wonder how an A4-5000 mini-ITX based system would compare to a Pentium J2850 for HTPC and general use:

http://liliputing.com/2013/07/leake...eron-and-pentium-chips-use-less-than-10w.html

 

[Exophase]

Jaguar has 128-bit wide FPUs, and Silvermont doesn't. Not really sure why it's a surprise that it has considerably higher floating-point IPC.

Silvermont has 2x128-bit integer ALUs, 1x128-bit FADD, and 1x64-bit FMUL and IMUL. Same as Saltwell. Very different from Bobcat that's 64-bit for all of those. Please don't just say it's not 128-bit

 

[NTMBK]

The Atom is doing pretty well for itself IPC wise as a 2-wide architecture (not even fuly Ooo) against the three-wide architectures from VIA and AMD, considering that the VIA is a 27.5W TDP processor w/o igpu (40nm).

Jaguar isn't 3 wide...

 

[NTMBK]

• Intel Atom Z3770 1.46GHz vs. VIA QuadCore L4800 1.46+GHz (1.6GHz capable)
source: http://browser.primatelabs.com/geekbench2/compare/2068301/2124912


• AMD A4-5000 1.5GHz vs. VIA QuadCore L4800 1.46+GHz (1.6GHz capable)
source: http://browser.primatelabs.com/geekbench2/compare/2139838/2124912

What is the power consumption of this L4800, and what is the power consumption of its northbridge and southbridge?

 

[jfpoole]

The integer code is more representative (but not of everything, not even close!) because it runs actual libraries and VMs that are in use. But even that relies on the input data set being a good representation. I have my doubts about the program they're using for the Lua test..

If you've got a minute I'd be really interested to hear what these doubts are so that we can improve the workload for v3.

 

[Exophase]

If you've got a minute I'd be really interested to hear what these doubts are so that we can improve the workload for v3.

My experience with Lua is zero so I might not be the best person to ask - I think you should poke around some Lua forums (lua.org for instance) and ask for some input, or even if they'd like to contribute code.

I just don't think a prime number sieve is representative of scripts. I'm not sure of the nature of the VM - if it's purely interpreted (as opposed to JITed) then you'll be spending a lot of time in the interpretation loop outside of even executing instructions. Still, a simple test like primality could end up only exercising a subset of the VM instructions and will at the very least end up with atypical cache footprints. Since this is the integer section you should also try to avoid something float heavy, although this might be hard to do in a Lua program where numeric types are doubles by default. Again, it depends on how well they're actually interpreting it.

I do know it's very big in game scripting. Maybe you could get a bunch of scripts from some open source game and churn through them, although that may be annoying to do in a stand-alone fashion.

 

[Nothingness]

I guess the Lua embedded in Geekbench is interpreted or it wouldn't run on iOS.

 

[tempestglen]

The Antutu benchmarks comparing Baytrail clocked at 1.4GHz (45,000) to Snapdragon 800 (30,000) i.e. the best SoC from ARM camp this year, shows an impressive lead of Baytrail over ARM, even when using an UNDERCLOCKED Baytrail

Later this year, we'll see the brainless tech blogging sheeps telling the exact opposite story from last year's "Intel is dead, ARM is king" nonsense to "Intel makes a comeback, challenge the ARM's throne". Go Intel!!!

I tell you why antutu(nbench engine) is completely useless:

1) The Interger test, atom harness SSE while Arm neon is absent because antutu disable neon for some ridiculous reasons.

2) floating test, same as Interger, no neon. Furthermore the antutu is double precision floating test which is not supported by A15/A9/Krait neon. A57 neon will comply IEEE standard of double precision floating.

So antutu is a rubbish in testing phone/pad CPUs.

 

[Exophase]

I guess the Lua embedded in Geekbench is interpreted or it wouldn't run on iOS.

That's a good point.

Would like to know the quality of the interpreter then. Going to assume it's at least interpreting bytecode.

 

[LegSWAT]

The Antutu benchmarks comparing Baytrail clocked at 1.4GHz (45,000) to Snapdragon 800 (30,000) i.e. the best SoC from ARM camp this year, shows an impressive lead of Baytrail over ARM, even when using an UNDERCLOCKED Baytrail

Later this year, we'll see the brainless tech blogging sheeps telling the exact opposite story from last year's "Intel is dead, ARM is king" nonsense to "Intel makes a comeback, challenge the ARM's throne". Go Intel!!!

Those Antutu scores are a case for the bin. Here's why:
http://forums.anandtech.com/showthread.php?t=2330027

As it slowly comes to light, given its fab advantage, Silvermont is quite a duck. Luckily Intel sits on mountains of cash for marketing and bribery :whiste:

 

[mikk]

Forget Geekbench, especially Floating point results. It is flawed.

 

[itsmydamnation]

VIA Nano (Isaiah microarchitecture codename CN) has 2 Media units with 128-bit wide datapath, supporting 4 single precision or 2 double-precision operations => Floating Point Performance

Floating Point

VIA Nano X2 U4025 @ 1.20 GHz
1 processor, 2 cores
Processor / Microarchitecture Codename: CN / Isaiah
Unveiling: 2008
2860 (139,92%)

vs.

Intel Atom Z3770 @ 1.46 GHz
1 processor, 4 cores
Processor / Microarchitecture Codename: ValleyView-T / Silvermount
Unveiling: 2013
2044 (100%)

source: http://browser.primatelabs.com/geekb...803921/2068301

http://www.cpu-world.com/CPUs/Nano_X2/VIA-Nano X2 U4025.html

so that's not really definitive, but if i had to evaluate:

1. its BIG (even for 40nm)
2. it is another class when it comes to power consumption
3.it doesn't have a GPU on die
4. it doesn't have its norther/south bridge on die

 

[shady28]

http://www.cpu-world.com/CPUs/Nano_X2/VIA-Nano X2 U4025.html

so that's not really definitive, but if i had to evaluate:

1. its BIG (even for 40nm)
2. it is another class when it comes to power consumption
3.it doesn't have a GPU on die
4. it doesn't have its norther/south bridge on die

That Nano die is 66mm^2. A z2760 is 65mm^2.

Other points dead spot on. Nano isn't suitable for a tablet.

 

[itsmydamnation]

That Nano die is 66mm^2. A z2760 is 65mm^2.

Other points dead spot on. Nano isn't suitable for a tablet.

but thats just cores right? no north bridge, no GPU. jaguars cores for example are what 3.1mm sq.