Forget Intel Broadwell, Skylake On the Way

From what I saw from Passmark's charts, Clarksfield to Sandy Bridge had almost a doubling of processing score (100% improvement) per product segment, then 10-15% improvement from Sandy Bridge to Ivy Bridge and 5-10% improvement from Ivy Bridge to Haswell.

 

AnandTech Nehalem Review
AnandTech Sandy Bridge Review
AnandTech Haswell Review (bonus: 5 generations compared)

This should give a pretty good picture of the IPC improvements from one arch to the next.

 

Way to reference something besides the notorious PassMark. :thumbsup:

Nehalem to Sandy Bridge is clearly the largest leap, but it's not as massive as many seem to believe when you factor in the clock speed differences. I think Sandy Bridge did much more for mobile than it did for desktops because of how disappointing Clarksfield was, especially the quad-cores. Had the misfortune of owning an i7-740QM at one time. Ick.

I wonder if Nehalem's tick (Westmere) had any meaningful IPC improvements like Ivy Bridge had over Sandy Bridge...

 

IMHO best you can hope for is give a choice. People will believe what they want whether it's right or wrong. Just remember that nobody is perfect, if we were we wouldn't be human.

That's why I see the funny side of your post.

 

is passmark even legit?

 

Ofc not, which is entirely the point.

 

k thats what i thought...i saw that one day and i said i never heard of this and its rating system seems complete BS and stupid

 

Yep, the results in their database are totally whack. PassMark is a poor benchmarking software, there's a reason nobody uses it.

 

Passmark is legit. It's just it's a measure of raw horsepower and doesn't really equate to a real world workload. It's like claiming horsepower figures between engines, although what does that mean for your daily commute?

 

It's a pretty specific "type" of horsepower though -- there is no single variable in computing that translates to performance so cleanly as horsepower in engines.

Regarding IPC: it is important, but not that important. Remember, in the glory days of CPUs when we were getting literally double the performance every couple of years, much of this performance increase came not from IPC, but from dramatic increases in operating frequency. Such increases don't happen anymore (Clarksfield to Sandy Bridge doesn't really count because the former was a server architecture unwisely stuffed into a mobile form factor), but there is a noticeable increase in clock speed from Sandy Bridge to Haswell. As long as this doesn't come with increases in price or power consumption (and it doesn't), it is no less a legitimate method of improving performance than IPC.

If you take all differences into account, the improvement from Sandy Bridge to Haswell in the real-life production application I work on (i.e. on server CPUs, not laptop ones), is around 25% on a per-core basis. This is nowhere near what it was in the glory days, but given that Haswell Xeons also have more cores for the same price, it's not terrible.

 

The way I see it is that newer platforms 'Always' give me better performance (if they didn't; that is what 100% refund/return policies are for).

This is a good thread/post for the whiners about the small jumps they think newer Intel architectures bring.

 

Well see one could argue it's precisely because of the slowdown in clockspeed increase that IPC improvements have now become more important than ever.

Also I think you really ought to clarify that only laptop CPUs experienced a somewhat significant speed increase from Sandy Bridge to Haswell; desktop chips had a very minor bump (+100MHz) for the most part (negating 4790K which really is factory overclocked 4770K) if you compare "like for like" chips.

 

SB has a 5-5.5GHz top clock and haswell has a 4.5-4.7GHz top clock so there was a significant drop for the architecture. so haswell beat or equals SB depending on the chip

are we counting factory clocks or max clocks because factory clocks are irrelevant in my eyes.

 

Thanks yeah I didn't want to bring OC headroom into the discussion since it's too variable and most people don't OC. BUT yes you raise a very good point in that any IPC advantage Haswell has over SB vanishes when you take into account the much lower OC headroom as well as consistency on Haswell chips.

 

haswell has a nice consistent range though and uses a heck of a ot less power IIRC

 

If you are talking about IPC as in Instructions Per Clock then frequency is irrelevant.

 

What I'm trying to get at is that with IPC you can do more for less ie 4.4GHz on my Haswell is equivalent to 4.5GHz on my Ivy Bridge.

However, if IPC is better but absolute useable clockspeed drops, then at some point the IPC advantage vanishes because you could simply clock the chip with less IPC higher to compensate for it.

 

LOL no it isn't. That's why nobody uses it. It's not on HWBot or in any major hardware reviewer's benchmark suite as far as I know.

The reason their charts are all messed up is that they average out all the user submitted scores for a particular piece of hardware, regardless of the test platform, overclocking, PassMark version, and outliers. There's too many variables they don't account for. No standardization at all means unreliable results.

If you take a look at the top of this chart, there's just so many things wrong with it. Their graphics test is an utter joke. No multi-GPU support and has about the technical complexity of 3DMark2001.

 

Actually Passmark is on HWBot, just doesn't count for points and under a different name -- Performance Test.

Personally I'd say their CPU charts are usable if you just want a general idea how one chip stacks against another and ignore the absolute numbers. I think the real value of Passmark is for comparing mid to low end chips for budget conscious people, since for those chips there are few if any reviews out there at all. Also good if you're sitting on a crappy old chip and want to know just how much it sucks. I still laught at just how pathetically weak the ML-30 I had in my Compaq laptop was, and I was still using that in 2011!
It's not exactly a popular benchmark, so there's less skew introducted from extreme benchers, and averaged over 1000 data points the trend is still there.

And FWIW they do actually exclude results if they are too far outside the norm. I remember running Passmark with RAPID on back in the day and getting ridiculous numbers for my storage score. Then 2 days later my results get "excluded due to anomalies".

 

I just don't see much value in PassMark when resources such as these are readily available:

AnandTech | Bench - CPU

Performance Charts - Tom's Hardware

 

For CPUs where other resources are available, sure I agree. Passmark has by far the most comprehensive coverage though, and sometimes it's the only thing available especially if you want to compare some rather esoteric chips.

 

Anand and Tom's are more comprehensive as far as benching different categories of hardware, but I agree regarding the more obscure chips. Still don't like PassMark though.

 

Exactly. Passmark is the only benchmark that directly compares old and new, mobile and desktop processors. Its value is in its database of past results and the ability to directly compare processors that were never reviewed or compared with other benchmarks or reviews. It may not be comprehensive, but it's a great comparison tool.

 

I should clarify comprehensive coverage in terms of chips, not that the benchmarking is comprehensive. It's hard to find another database that has pretty much every variation on every chip that Intel and AMD have released to date.

 

notebookcheck.net has a huge database...is it much larger then theres? NBC also is very reputable and shows you review break downs of each hardware

 

Yes, Passmark's database is bigger than Notebook Check's. Notebook Check's database is mainly devoted to mobile graphics, and does an exceptional job with that. I also agree that their reviews are superb - possibly the best review site on the internet. What Passmark is especially good at is providing a way to compare a wide variety of processors. It's all about the right tools for the job. If I'm comparing graphics cards, I go to Notebook Check. If I'm comparing processors, I go to Passmark's CPUBenchmark.net.

 

Bit of a grey area as IPC improvements may be made to some instructions but not others.

For real IPC improvements see 4. Instruction tables: Lists of instruction latencies, throughputs and micro-operation breakdowns for Intel, AMD and VIA CPU.

What you will see in software will be very dependent on what instructions are being used in the software and how well they have been optimized or not.

Do you have some figures? Linpack at 3.5GHz on an undervolted 4700MQ will draw over 70W. Of course it's using AVX2, be interesting to see what Skylake does with AVX3.

Yes, it's nice to see. Where do they get their benchmarks from, is it just their own testing?

 

you go from GPU to CPU...make up your mind. For CPU they have tested majority of all CPUs even desktops to a decent extent...minus AMD but amd desktops suck and other sources exist.

its there own testing over the years

The fact its a smaller process is for one and its well know haswell is more efficient. I can rock out much better freqs with my IB over other SBs. You can also google around there is this huge thread on OCing SB and haswell on overclock.net or hardocp. I forget which though. Its been posted a ways back IIRC

 

Wouldn't this mean manufacturers would start shipping until after "2H" 2015, so consumers won't get their hands on any laptops with skylake inside until 1 year from now, if they decide to release during back-to-school or maybe it gets pushed back to christmas? I don't see how Intel can delay broadwell this long then release skylake 6 months later, or maybe I'm just bitter because I need a new laptop soon and it looks like its going to be a tick, again.

 

Okay, thought you might have known a good source of power measurements off the top of your head. Not so many power statistics out there.

Did find this one 2760QM shown running prime95 at 2.98GHz 1.226V 70W

If I run prime95 on my 4700MQ at default voltage of 1.010V using 30x multi I get 69W and worse temperatures.

While I would agree with your statement the Haswell is more efficient I'm not so sure about your previous statement.

Looks like Haswell doesn't use less power, heats up more but does get more work done. Perhaps what you meant was Haswell uses less power for producing the same work. Of course this is just going by that one post so maybe not reliable and Haswell does have the advantage that it is easier to undervolt to produce better results.

 

a 4.5GHz haswell is like a 5.2GHz SB IIRC so a 4.5GHz haswell is more efficient and uses less power than the SB CPU. Also haswell OC range is from 4.5-4.7 GHz while SB is 5-5.5 GHz. Haswell has a more consistent over clock. Thats what i was getting at. Basically Haswell and SB are equal performance from what i read when you count max clock....assuming there is no major difference due to instructions. You'll be able to see peoples TDP ranges in those OC threads so you'll be able to roughly calc how much more efficient it is.


Also whats the point of that link? It doesn't show anything from what i understand. I can make my 920xm run at 70W at 2 GHz if i force excessive voltage.

 

Preliminary Skylake mobile SKU lineup

 

So finally i5 quads on mobile now after, what, 5 years? If there is an unlocked part, lack of HT might actually help when overclocking.

Pretty bummed that everything is soldered across the board.

 

at the highway robber prices that mobile CPUs go for you figure they would at least offer unlocked i7 chips across the board.

 

Dude this is a corporation that has a virtual monopoly in the high-end CPU market, you really can't expect too much. :/

 

i am just comparing the desktop vs mobile prices and features. If i pay 500 dollars for a 300 dollar chip and the 300 dollar chip is unlocked and mine isn't you at least figure the 500 dollar one would be unlocked too. Obviously they like raping us but still

 

10% of the justification comes from the fact the mobile chips need to be binned much higher than desktop chips since they need to be very efficient.

The other 90% is because they can.

 

Actually, you're paying $1100 for a $340 unlocked i7, a level of extortion that's only matched by mobile GPU's.

 

everyone says they require high binning but i have never seen any proof of that from intel or a credible source.

The CPU is the exact same as the desktops with no difference. I highly doubt any tangible binning is required and that still doesn't mean for 200 extra bucks they can't unlock it.

500 vs 300 bucks i bet the 200 bucks covers any extra binning needed and they could simply allow unlocks. Hell on top of that laptop CPUs come wit virtually no warranty.

@octiceps...i am referring to the non XM chips. You can buy a 500 dollar HQ chip in those nice ultra slim gaming laptops like the X7 pro or whatever but you are forced to crap single clocks because its a locked chip that can't be replaced.

Desktop version is 300 vs 500 and is also unlocked. They could as least unlock it. the 200 bucks isn't an issue to me if it was unlocked.

 

Yeah but the 4940MX is $1100 and 4790K/4770K is $340. Those are the true equivalents.

 

yea but there are unlocked R desktop chips that are the exact same as the laptops but less money and unlocked

 

You mean the 4770R? It looks completely different. BGA, lower TDP, clock speed, locked multiplier, less cache, Iris Pro.

 

thats the same as the HQ processors...thats what i was referencing about the ultra thin laptops that are in the 970/980m thread. Also that should be unlocked because i remember reading a review that was playing around with TDP and CPU speeds.

 

Like I said, higher binning, if true, is only a very small justification. Mostly Intel does it because they can, and there are people willing to pay those prices.

 

I don't know what your definition of unlocked means, but the HQ CPU's don't fit my definition, which are ones with a fully unlocked multiplier and power limits like the -XM/-MX, -K, and -X parts.

 

Also just to be fair, up until Haswell-E Intel pulled the same shenanigans on desktop chips as well. I mean going from 3930K to 3960X/3970X you're basically paying $400 more for 3MB of extra L3 cache and essentially unlimited multipliers (3930K only goes up to 57x, 3960X/3970X goes up to 255x). Same idea with 4930K to 4960X, extra $400 for more L3 cache and 255x multi, except in this case it's even more pointless as 4930K's multi maxes at 63x, and for the average enthusiast they'd be hard pressed to hit even 50x without some luck.

At least with Haswell-E, 5960X offers 2 more cores for the extra $400.

 

well i remember reading that one of the R CPUs was fully unlocked....if thats not the case thats dumb.

 

The link showed power, voltage and temps for SNB at ~3GHz. AFAIK the processor was operating at it's default VID unless you know of some SNB QM processors that allow adjustments to vcore, a complaint that has been long for people wishing to undervolt. The specification TDP for that SNB QM is 45W, for 4700MQ it's 47W. So using either within specification uses just about the same power while the more efficient HSW will run hotter Watt for Watt while generally giving better performance clock for clock.

If you look at the HWBOT records with sub ambient cooling the SNB 2500K/2700K is ~6GHz while 3770K/4700K is 7.2GHz. Heat is a problem and IIRC I also mentioned this to you before but as a reminder, each shrinkage creates higher power densities which means higher thermals Watt for Watt. Either TDP can be lowered to compensate or expansion of cores / iGD to keep the package at the standard TDP of around 45W for mobile quads. Personally I would prefer to improve core performance by keeping TDP high rather than having it reduced but thermals may not permit that.

Not so sure about that, would have to check the voltage scaling.

Can't help wonder if those neutered mobile graphics chips are defective desktop dies. Ouch, that would add salt to the wound.

Only joking, I hope!.

 

I dunno if IPC advantage can be compared this way, but FWIW in a single threaded benchmark like SuperPI Haswell appears to have +200MHz over SB if you compare the SuperPI results for 4770K vs 2700K.

The gap widens to 400-500MHz in a multithreaded benchmark like CineBench R11.5 if you look at some of the numbers for 4770K and the 2700K.

Again, these are very crude comparisons but they suffice to give you an idea of the IPC improvements from Sandy Bridge to Haswell.

 

IPC is a hard one, Haswell cores are more powerful (with the 2 extra execution ports) but the cache is actually slower than Sandy or Ivy so the difference isn't as great as it should be.