[Softpedia] 2014-Bound Intel Broadwell-K CPUs Get 80% Graphics Boost from Iris Pro

[mikk]

edit, I feel like I have worded this badly.
My original argument was (and still stands) that giving Broadwell K 80% more performance in games over HD4600/Haswell K is believable, both from a technical and an economic standpoint. Giving it 80% more performance over HD5200/4770R may be possible with enough resources dedicated, but it doesn't seem like something that is economic.

80% faster than 4770k means Broadwell-K GT3e brings only an 20-25% improvement over Haswell GT3e 55W. Considering that these performance figures are often from 3dmark it could be even less. Broadwell GT3 surely has slightly more EUs. If this is true than Gen8 basically doesn't bring performance improvements. Not sure why Intel claimed they made a huge redesign when there are no improvements.

 

[erunion]

Piroko, I understand that you are skeptical about BDW GT3e being 80% faster than HSW GT3e. You're free to not believe the rumor. But the alternative scenarios you've presented are far less convincing than the rumor; especially since you selectively accept some parts of the rumor as true while rejecting others. (eg " Maybe the 80% figure is true, its just 80% faster than something other than HSW GT3e")

Ninja edit: now that I've made a fool of myself , there is a comment on the CPU world by "gshv" (is that the author?) that says 80% is versus 4770k

 

[mikk]

Even worse it would mean that Broadwell GT3 isn't really faster than Haswell GT3 in the mobile segment. A shrink to 14nm and such a small improvement. That would be a joke from Intel.

 

[PPB]

I think its 80% more EUs. But can be whatever really, not knowing 80% more than what at what metric and which platform is bogus, at best.

Move on, nothing to see here.

 

[Khato]

Umm, the 16 EU Ivy Bridge has 2x the Flops of 12 EU Sandy Bridge.

But how does that count as more than a simple evolution exactly? The EUs are merely one part of the design and for all we know the gains seen by them could have simply been fixing bugs on the Gen6 implementation. (It's not uncommon in the least for some performance features to not work and simply be disabled rather than delaying the launch.)

Now that I think about it a bit more, if I were to guess at why Ivybridge's GPU was labeled Gen7 instead of Gen6.5 it'd be the directx 11 support and changes in partitioning to set themselves up for the future. But again, that's more about features rather than performance.

Also the "big jump" with Broadwell they are claiming they contrast with Gen 3 with Gen 4 jump. Performance-wise, that was horrendous. GMA X3000 was in no way a "big jump" over GMA 950.

Haha, quite correct with respect to performance. But in the context of the driver side (which is what the comment was with respect to) there was little in common between GMA X3000 and GMA 950. Now whether or not having a major shift in architecture that affects Intel's drivers is a good thing or not remains to be seen... and sadly isn't likely to be answered for well over half a year.

 

[Piroko]

80% faster than 4770k means Broadwell-K GT3e brings only an 20-25% improvement over Haswell GT3e 55W. Considering that these performance figures are often from 3dmark it could be even less. Broadwell GT3 surely has slightly more EUs. If this is true than Gen8 basically doesn't bring performance improvements. Not sure why Intel claimed they made a huge redesign when there are no improvements.

Piroko, I understand that you are skeptical about BDW GT3e being 80% faster than HSW GT3e. You're free to not believe the rumor. But the alternative scenarios you've presented are far less convincing than the rumor; especially since you selectively accept some parts of the rumor as true while rejecting others. (eg " Maybe the 80% figure is true, its just 80% faster than something other than HSW GT3e")

Even worse it would mean that Broadwell GT3 isn't really faster than Haswell GT3 in the mobile segment. A shrink to 14nm and such a small improvement. That would be a joke from Intel.

Remember that this news is about Broadwell K specifically, and has no rating whatsoever of different SKUs.

Also, I didn't selectively accept parts of the rumor. There is no additional information to be had in this article. No clockspeeds, no benchmarks, no defined test comparison. I have drawn a conclusion, based on one fairly logical assumption (that Broadwell K is compared to Haswell K) whereas you added lots of assumptions (clock rate assumption to analyze relative performance improvements; comparison to a chip that is in a different market segment; performance derived from a specific benchmark; die shrink is used to improve the performance metric specifically). Those are all assumptions that I have either not rated at all or rated to be hard to follow.

 

[mikk]

Remember that this news is about Broadwell K specifically, and has no rating whatsoever of different SKUs.

Also, I didn't selectively accept parts of the rumor. There is no additional information to be had in this article. No clockspeeds, no benchmarks, no defined test comparison. I have drawn a conclusion, based on one fairly logical assumption (that Broadwell K is compared to Haswell K) whereas you added lots of assumptions (clock rate assumption to analyze relative performance improvements; comparison to a chip that is in a different market segment; performance derived from a specific benchmark; die shrink is used to improve the performance metric specifically). Those are all assumptions that I have either not rated at all or rated to be hard to follow.

I see lots of assumptions from your side. Why do you mention clockspeeds? Broadwell-K isn't a mobile CPU and therefore unlikely that it's crippled by clockpeeds and/or TDP limited. 80% if true must come from an Intel slide. These estimates are usually overly optimistic. It's also safe to say that Broadwell has a new Gen8 GPU and at least 40 EUs for GT3, no I think an EU increase is realistic. 80% over a Gen7.5 GT2 is very very poor for a desktop GT3e Gen8. A TDP unrestricted desktop Haswell GT3e with 1300 Mhz could do the same. If true then Gen8 is a big fail from Intel.

 

[Shivansps]

They say socket 1150, but we know anything about compatibility with current x8x chipsets and mbs?

As for EU GT1 is not gona be less than Cherry Trial, if Cherry Trail is 16EU is probable that the scale is 16(GT1)/32/64, either that or GT1 is 20 EU with 20/40/80 but it seems too much.

 

[IntelUser2000]

If true then Gen8 is a big fail from Intel.

And why do you consider that?

If you consider that GT4 is available which will be that much faster with 2x the EUs, and GT3 in K CPUs is a big advancement itself because Haswell only has GT2. Silicon performance is explicitely TDP limited nowadays. If the GT4 ends up being 2x fast as Iris Pro in Haswell, that would make perfect sense as a shrink would afford a 2x headroom. Were people disappointed when new GPUs went from a "Moores Law" cubed(Nvidia claim) to a gain more in line with process shrink? Yes, but that's the reality.

Many times we have been disappointed by overinflated projections and misunderstood wordings, regardless of whether they were from Intel, Nvidia, AMD, or whether it was graphics cards, CPUs, chipsets.

 

[mikk]

And why do you consider that?

Not faster than Gen7.5 despite the huge redesign.

 

[ShintaiDK]

They say socket 1150, but we know anything about compatibility with current x8x chipsets and mbs?

As for EU GT1 is not gona be less than Cherry Trial, if Cherry Trail is 16EU is probable that the scale is 16(GT1)/32/64, either that or GT1 is 20 EU with 20/40/80 but it seems too much.

Broadwell will not work in current boards, unless the mobo maker had foresight for it. Broadwell-K is a new addition. Originally Broadwell would only be E3 xeons and mobile only. Now desktop got 1-2 chips as well.

Haswell will work in 9 series boards tho.

 

[jpiniero]

How many dies would Intel do? I suppose if the mobile Quad models are only going to be GT3e and GT4e it makes sense. I still find it very strange that Intel is even bothering to do a socket model for only the Broadwell-K and Xeons. The volume can't really justify it?

Any word on the cpu side? Performance, IPC, clocks, power consumption?

No improvement, but Crystalwell might add a few percent.

 

[ShintaiDK]

How many dies would Intel do? I suppose if the mobile Quad models are only going to be GT3e and GT4e it makes sense. I still find it very strange that Intel is even bothering to do a socket model for only the Broadwell-K and Xeons. The volume can't really justify it?



No improvement, but Crystalwell might add a few percent.

If Intel could, they would release Broadwell into everything.

The reason why mobile and xeons get it, is because performance/watt and power consumption matters a great deal. The Broadwell-K is simply a late addition after capacity evaluation.

Broadwell demoed used 30% less power for the same task.

4 new instructions is also added. And the cache is changed besides other minor uarch improvements.

 

[Shivansps]

So, no Broadwell non-K for desktop?

 

[jpiniero]

Broadwell demoed used 30% less power for the same task.

At mobile speeds, yeah. I imagine once you get near 4 ghz the increased heat density at 14 nm will eliminate any advantage.

So, no Broadwell non-K for desktop?

Well we don't know if Intel will be more aggressive about desktop BGA models. And of course there's always Haswell Refresh.

 

[blackened23]

They say socket 1150, but we know anything about compatibility with current x8x chipsets and mbs?

As for EU GT1 is not gona be less than Cherry Trial, if Cherry Trail is 16EU is probable that the scale is 16(GT1)/32/64, either that or GT1 is 20 EU with 20/40/80 but it seems too much.

It isn't compatible with existing LGA 1150.

 

[Piroko]

I see lots of assumptions from your side. Why do you mention clockspeeds? Broadwell-K isn't a mobile CPU and therefore unlikely that it's crippled by clockpeeds and/or TDP limited. 80% if true must come from an Intel slide. These estimates are usually overly optimistic. It's also safe to say that Broadwell has a new Gen8 GPU and at least 40 EUs for GT3, no I think an EU increase is realistic. 80% over a Gen7.5 GT2 is very very poor for a desktop GT3e Gen8. A TDP unrestricted desktop Haswell GT3e with 1300 Mhz could do the same. If true then Gen8 is a big fail from Intel.

Oh boy, reading comprehension.

Again, my argument: HW-K +80% for BW-K would be very realistic. HW-R +80% for BW-K would be possible, but less realistic.
You make yourself a fool by overanalyzing this message. All that talk about Gen8 vs. Gen7.5 is coming from you only.

 

[IntelUser2000]

Not faster than Gen7.5 despite the huge redesign.

Again, I will ask. Why is that a fail? If GT4 Broadwell gets 2x the speed of Haswell GT3, it would be perfectly fine, because new process technology offers 2x thermal and transistor count headroom. If there's anything that failed, you failed to answer rest of my points.

And who said Iris Pro is 80% faster than the HD 4600? Go look at Anandtech's benchmarks again. Iris Pro is only 50-60% faster. 20% faster on the Broadwell GT3 would mean it would be approximately 80% faster than HD 4600. That makes perfect sense.

Who cares about the redesign? I want you to give me proof of a scenario where a radical redesign offered substantial performance improvements.

GMA X3000, Nvidia NV30, Pentium 4, Bulldozer were ALL radical changes that brought minimal to degradation in performance.

Featureset!=Performance

 

[IntelUser2000]

I still find it very strange that Intel is even bothering to do a socket model for only the Broadwell-K and Xeons. The volume can't really justify it?

Haswell has something like 8 dies already, that's excluding the server parts.

The volume doesn't matter, because Broadwell-K die can be same as the Quad Core Mobile parts. They will be shared anyway.

 

[mikk]

Again, my argument: HW-K +80% for BW-K would be very realistic. HW-R +80% for BW-K would be possible, but less realistic.
You make yourself a fool by overanalyzing this message. All that talk about Gen8 vs. Gen7.5 is coming from you only.

You argument isn't convincing and even if it turns out right Intel failed on Broadwell.

Again, I will ask. Why is that a fail?

Because of no performance improvements from Gen8 expect more EUs. Efficiency not impovement which doesn't bode well.

If GT4 Broadwell gets 2x the speed of Haswell GT3

Give me as source to Broadwell GT4. Nothing is confirmed and the driver claims it for desktop only not for mobile. As already mentioned GT3 is a big problem for notebooks if there are no improvements.

And who said Iris Pro is 80% faster than the HD 4600? Go look at Anandtech's benchmarks again. Iris Pro is only 50-60% faster.

Mobile Iris Pro 47/55W is 50-60% faster. A TDP unrestricted desktop 65W+ 1,3 Ghz Iris Pro would be 70% faster or more (mobile is usually slower due to worse DDR3 latencies). Just 10% faster than this is a huge fail. And once again 80% from an Intel claim doesn't bode well, it's often exaggerated.

 

[Techhog]

Hopefully this applies to laptops too, and it won't only be on high-end CPUs. 660M-like power on an IGP for ~$1000? Yes please.

 

[Techhog]

Broadwell will not work in current boards, unless the mobo maker had foresight for it. Broadwell-K is a new addition. Originally Broadwell would only be E3 xeons and mobile only. Now desktop got 1-2 chips as well.

Haswell will work in 9 series boards tho.

So, Haswell is a DOA platform? *sigh*

 

[ShintaiDK]

So, Haswell is a DOA platform? *sigh*

Only if you buy Broadwell-K or Xeon E3v4. Originally there was no desktop Broadwell CPUs.

But again, why would you upgrade from Haswell to Broadwell?

 

[IntelUser2000]

A TDP unrestricted desktop 65W+ 1,3 Ghz Iris Pro would be 70% faster or more

In fantasy land.

Apple's new iMac shows its even slower than the 45W reference platform, let alone that one at 55W.

Because of no performance improvements from Gen8 expect more EUs. Efficiency not impovement which doesn't bode well.

You don't get it. Efficiency is an OVERALL thing. New process gives 2x headroom for power, assuming the design is done right, and if Broadwell GT4 does that, then they would have met their goal. I'd hardly call 2x a fail. Perhaps you should see what AMD GPUs and discrete GPUs are doing.

It does not matter if the shaders are more "efficient" if its power limited and goes back to levels same as the last generation.
-Haswell and Bay Trail shows that Intel's 22nm efficiency was at low power, and at high end the gains were very small
-Ivy Bridge had design to allow 2x the Flops... for a mere 40% gain in performance. Of course the 2x "efficiency"(actually a very generic term, a correct one is perf/watt) is at the low power level, to enable chips like Y series.

I do not have an issue with claiming that the highest end Broadwell iGPU config with GT4 might end up faster than 2x. I do have a problem with you saying 2x gain for GT4 vs GT3 is a disappointment(therefore indicating Broadwell GT3 is only 20-30% faster than Haswell GT3). 1.8x for GT3 vs GT3 would indicate somewhere in the range of 3x for GT4! Extraordinary claims require extraordinary proof.

Nothing is confirmed and the driver claims it for desktop only not for mobile. As already mentioned GT3 is a big problem for notebooks if there are no improvements.

Zero issue.

Notebooks will have GT4, and Desktops, which never really had GT3, would get GT3, for a big improvement over current GT2. BTW, drivers can claim what it wants, but actual company plans usually turns out different. Fast iGPU is much more useless on a Desktop hence no need for the fastest one.

 

[mikk]

In fantasy land.

Apple's new iMac shows its even slower than the 45W reference platform, let alone that one at 55W.

iMac is running with a 1,15 Ghz Iris Pro i5 and 2MB less LLC. I referred to a 1,3 Ghz Iris Pro. Different drivers could be the reason as well in some games. Afaik even the R-models using SO-DIMM which makes CL11 likely. The iMAC itself isn't really a desktop, thermical characteristics unlikely to improve. A real Desktop GT3e on Broadwell could use the common CL9 with DDR3-1600 assuming DDR3-1600 max official stays. This alone can bring 1-5% over the Haswell Iris Pro.

New process gives 2x headroom for power

Not really. Let's say 1.5x-1.6x is a more realistic improvement.

, assuming the design is done right, and if Broadwell GT4 does that, then they would have met their goal. I'd hardly call 2x a fail.

Once again, your source for GT4?

-Ivy Bridge had design to allow 2x the Flops... for a mere 40% gain in performance.

Ivy Bridge at the same frequency improved ~50%, with latest drivers it's bigger. Intel claimed that Gen8 is a bigger step.

I do have a problem with you saying 2x gain for GT4 vs GT3 is a disappointment

I haven't said this.

Notebooks will have GT4

And here once again, your source?

80% over GT2 is a fail. Get over it. And I doubt Intel failed so much with Gen8.