Intel Skylake / Kaby Lake

[IntelUser2000]

I don't know why you've become so Intel-cynical.

Two of the greatest Intel defenders on this forum, III-V and Arachnotronic have turned away from Intel. Any Intel fans left, or has the fanbase just become less polarized ?

Count me as one of them too.

The basis for my reasoning was that their core was strong. And that was manufacturing and process. They don't really have that anymore, and the little they have is seemingly in jeopardy of being eclipsed.* Even if the lead was there, they care too much about margins to make full use of it. I find it very funny they compare density with foundry guys when they make CPU designs that far outweigh any density advantages their process offers. Atom was always larger than ARM cores at same performance for example.

And more importantly, while we might view them as a company/group, they are still run by humans, and individuals. People tend to forget that. With lead architects of the core projects leaving, I can't help but imagine they have lot more serious issues that are yet to surface. The future of this company is not good, to put it nicely.

This explains what I am trying to say: http://www.anandtech.com/show/4122/intel-settles-with-nvidia-more-money-fewer-problems-no-x86

Page 2 notes: NVIDIA Doesn't Get x86 License, Including Rights To Make an x86 Emulator

The $1.5 billion they received from Intel in exchange is a piddly amount compared to what they lost. For the size of the company Nvidia's execution and technological firepower is phenomenal. Nvidia was the one that was able to match/exceed Intel's memory controller performance with NForce chipsets, and they had very good IGP when Intel's barely handled 2D. Whatever happened made Intel not hand out x86 patents to a potentially formidable competitor. Thanks to that they get to keep x86 market practically to themselves and keep its fat margins.

*Apple with TSMC collaboration is yielding quite the amazing results. Packaging technologies like InFO, and absolutely cutting edge memory like LPDDR4-4266 is mind blowing compared to what Intel is doing. THEY(Intel) used to be the one that spearheaded technological developments in computers, in all areas. Not anymore.

 

[witeken]

Count me as one of them too.

The basis for my reasoning was that their core was strong. And that was manufacturing and process. They don't really have that anymore, and the little they have is seemingly in jeopardy of being eclipsed.*

*Apple with TSMC collaboration is yielding quite the amazing results. Packaging technologies like InFO, and absolutely cutting edge memory like LPDDR4-4266 is mind blowing compared to what Intel is doing. THEY(Intel) used to be the one that spearheaded technological developments in computers, in all areas. Not anymore.

You're right, many of Intel's amazing products have been plagues by delays. Silicon photonics is awesome, but was delayed by about a year, but now it's out there. Knights Landing is now also out there, but a year too late, but I don't know if there are any systems yet (in TOP500), and together with it Omni Path Architecture was also delayed. I don't know how HBM is doing. Broadwell-E and Broadwell Xeon also didn't make 2015. The 3D NAND with breakthrough cost is only just showing up, and 3D XPoint won't make 2016 as promised, and I don't know if it will make Kaby Lake S. The 14nm story we know. The mobile story as well.

By dealying products, you are sucking the excitement out of them.

But in any case, Intel's manufacturing advantage is still standing strong. Each foundry has their own things they focus on. For Intel it's the transistor, both performance and density, they have air gaps, they have 2.5D I think. (For Samsung it's DRAM, for Micron it used to be NAND, but they showed up late for 3D NAND.)

 

[IntelUser2000]

You're right, many of Intel's amazing products have been plagues by delays. Silicon photonics is awesome, but was delayed by about a year,....

I will tell you what looks good. Silicon Photonics and Knights Landing. They are both cutting edge. The possibility of bringing optical performance to mass scale systems are quite exciting. For Knights Landing, its finally something that was promised by them years ago. You'd think taking a small core, and adding a vector unit and putting many of them to make it faster is easy. No its not. Knights Corner basically needed lot of optimizations to make it faster than Xeons. Ok, that's what you have them for, but they were slower for even multi-threaded applications until optimizations came. With KNL, you get faster than 1P Xeon out of the box, and you can make them even faster if you optimize, you don't need a separate CPU wasting power/cost, and you can add more Xeon Phis or even Tesla GPUs if you want. And market is responding.

KNL on original schedule would have been dominating. 20% faster in real-world + 20% lower power + 1 year ago.

By dealying products, you are sucking the excitement out of them.

Worse. You get worse products than if it wasn't delayed. Why is it delayed in the first place? Mismanaged expectations to put it simply. You try to do more then your team is capable of doing. And in the process they find out they can't. Delays = Missed performance target + late delivery

For Intel it's the transistor, both performance and density, they have air gaps, they have 2.5D I think. (For Samsung it's DRAM, for Micron it used to be NAND, but they showed up late for 3D NAND.)

You have to explain to me how the theory of having leadership process works when you can't compare to anyone. Intel isolated themselves with the gate called "x86". They have to stand on their own. And based on that it sucks. The CPUs that came out after 2011 are on a standstill. I've always stood for real-world products speaking for themselves. And Intel used to make products that lived up to that. Speaking of transistors, I knew for many years they had absolute leadership process because despite what others claimed(IBM/AMD/TSMC), the real world data showed otherwise. Now Intel is saying they have a 2+ year lead, they are the first to FinFET, and they have a 2nd gen now, and density is great. Who cares? Your CPUs are 5% faster, it still costs the same, and you got kicked out of the mobile market.

 

[Nothingness]

Ok, that's what you have them for, but they were slower for even multi-threaded applications until optimizations came. With KNL, you get faster than 1P Xeon out of the box

KNL is still a pain to program and out of the box the performance is bad. It requires a lot of work to get the most out of it.

 

[Arachnotronic]

KNL is still a pain to program and out of the box the performance is bad. It requires a lot of work to get the most out of it.

I love this quote from Murthy:

IDGNS: Nvidia is approaching the automotive markets aggressively with its GPUs, how will you compete?

I have a great deal of respect for Nvidia. But every time I think of Nvidia, I think about Californian wine where they can make great wine but it contains only one grape—great Cabernet Sauvignon or a great Chardonnay. I love French wines and French wines are blends where you need to be great at growing Cabernet, great at growing Merlot, great at growing Cabernet Franc. The art is in the mixture. That’s the benefit Intel has. We have GPU, we have CPU, we have custom silicon, we have embedded storage, we have FPGA. Nvidia’s going to basically say “I’ve got GPUs and I’ve got GPUs and I’ve got GPUs.” Great strategy, but it doesn’t give anywhere near the extensibility, flexibility and scalability that Intel is able to offer.

Total PR spin to deflect from the fact that KNL is not really competitive with GP100.

 

[norseamd]

Count me as one of them too.

The basis for my reasoning was that their core was strong. And that was manufacturing and process. They don't really have that anymore, and the little they have is seemingly in jeopardy of being eclipsed.* Even if the lead was there, they care too much about margins to make full use of it. I find it very funny they compare density with foundry guys when they make CPU designs that far outweigh any density advantages their process offers. Atom was always larger than ARM cores at same performance for example.

And more importantly, while we might view them as a company/group, they are still run by humans, and individuals. People tend to forget that. With lead architects of the core projects leaving, I can't help but imagine they have lot more serious issues that are yet to surface. The future of this company is not good, to put it nicely.

This explains what I am trying to say: http://www.anandtech.com/show/4122/intel-settles-with-nvidia-more-money-fewer-problems-no-x86

Page 2 notes: NVIDIA Doesn't Get x86 License, Including Rights To Make an x86 Emulator

The $1.5 billion they received from Intel in exchange is a piddly amount compared to what they lost. For the size of the company Nvidia's execution and technological firepower is phenomenal. Nvidia was the one that was able to match/exceed Intel's memory controller performance with NForce chipsets, and they had very good IGP when Intel's barely handled 2D. Whatever happened made Intel not hand out x86 patents to a potentially formidable competitor. Thanks to that they get to keep x86 market practically to themselves and keep its fat margins.

*Apple with TSMC collaboration is yielding quite the amazing results. Packaging technologies like InFO, and absolutely cutting edge memory like LPDDR4-4266 is mind blowing compared to what Intel is doing. THEY(Intel) used to be the one that spearheaded technological developments in computers, in all areas. Not anymore.

Dear god, as much as I hate Nvidia, I would have loved for Nvidia to start producing x86 CPUs, which would have notably increased the competition in the CPU marketspace.

So obviously Intel failed in the mobile space, and ARM is starting to take over more and more of the microprocessor marketplace, but how is Intel doing in IOT and embedded right now?

 

[witeken]

Total PR spin to deflect from the fact that KNL is not really competitive with GP100.

How is pricing and TDP? Well, I'll answer that myself with another little piece of research. @Arachnotronic @Nothingness @Idontcare

Note: All FLOPS are FP64.

Knights Landing: 3.46 TFLOPS @1.5GHz at 245W at $6250 1K unit list price
Pascal GP100: 4.78 TFLOPS @1.33GHz at 300W at $9,428 1 unit list price (11% more TFLOPS available with boost clock; 5.3TFLOPS)

KL: $1806 per TFLOPS
GP100: $1779 per TFLOPS (calculated with 5.3TFLOPS)

Pascal has about 13% higher performance per watt and slightly higher performance per dollar. However, these are the top end SKUs. Nvidia also has a 4.7TFLOPS one with 12GB RAM at $1.255 per TFLOPS. Intel, on the other hand, goes down to $916 per TFLOPS with their 7210 which does 2.66TFLOPS at only $2.440 (and has 16GB).

Die info.

Knights Landing: 8B x'tors at ~683mm² --- for 85.4mm²/1B --- for 0.288 TFLOPS / (1GHz * 1B)
Pascal GP100: 15.3B x'tors at 610mm² --- for 39.9mm²/1B --- for 0.235 TFLOPS / (1GHz * 1B)

[One thing that you do have to note here is that a 683mm² die for $2440 is actually quite competitive, while Nvidia has much higher prices for a smaller die. This shows that probably TSMC yields aren't spectacular either.]

So Pascal x'tors count per mm² is more than twice that of Knights Landing, which is insane density. However, if we normalize for clock speed and for transistor count, Knights Landing, very suprisingly, does better than Pascal. In other words, KL has more performance per transistor.

Now, what would happen if Intel decide to up their game with those short and sweet cells that Apple and Nvidia know and love to use? Only assumption I make here is that Intel can achieve 1.4x x'tor density of Nvidia (so density becomes 28.5mm² per 1B x'tors), which is of course controversial to say the least.

Knights Landing 2.0:

0.288 TFLOPS / (1GHz * 1Bx'tors) * (1B x'tors / 28.5mm²) * 683mm² * 1.5GHz = 10.35 TFLOPS (the thing I love about math is that units cancel)

Now, that would be a beast and an awesome prove of transistor lead.

[Now for a brief intermezzo. If we take this 28.5mm² number, and multiply it by 3.2, we come up with a 91.2mm² area for the Apple A10. This is indeed almost 1.4x smaller area than TSMC Apple A10. So Apple could really push down area if they go with Intel. This also shows that Apple A10 and Nvidia GP100 have comparable x'tor density. I do believe though that Apple A10 has higher density because it uses 16FFC, so an Intel 14nm+ Apple A10 wouldn't really be 91mm², but close to 100mm².]

I don't understand why Intel doesn't optimize for transistor count. If you're dealing with GPUs or many-cores, then you know that x'tors density is crucial. Especially with the 14nm yield problems you would expect Intel to do everything to optimize wafer area utilization.

Conclusions Pascal vs. Knights Landing

• Performance per dollar at high-end is about the same, but Nvidia has advantage because their price is for 1 unit, while Intel is 1K units
• Intel is the clear, absolute performance per dollar winner at low-end
• In other words, they don't have the performance crown, but they are like AMD: competitive pricing, so they have a compelling product
• Intel has poor transistor per area; Nvidia has more than double the x'tor density
• However, Intel actually has 23% better performance per transistor (at same clock speed)
• Nvidia has a 13% performance per watt advantage (based on TDP)

Sources

https://images.nvidia.com/content/pdf/tesla/whitepaper/pascal-architecture-whitepaper.pdf
http://www.nextplatform.com/2016/06/20/intel-knights-landing-yields-big-bang-buck-jump/
https://www.microway.com/hpc-tech-tips/nvidia-tesla-p100-price-analysis/
http://www.anandtech.com/show/9802/...s-knights-landing-xeon-phi-silicon-on-display

 

[Arachnotronic]

How is pricing and TDP?

Note: All FLOPS are FP64.

Knights Landing: 3.46 TFLOPS @1.5GHz at 245W at $6250 1K unit list price
Pascal GP100: 4.78 TFLOPS @1.33GHz at 300W at $9,428 1 unit list price (11% more TFLOPS available with boost clock; 5.3TFLOPS)

KL: $1806 per TFLOPS
GP100: $1779 per TFLOPS (calculated with 5.3TFLOPS)

Pascal has about 13% higher performance per watt and slightly higher performance per dollar. However, these are the top end SKUs. Nvidia also has a 4.7TFLOPS one with 12GB RAM at $1.255 per TFLOPS. Intel, on the other hand, goes down to $916 per TFLOPS with their 7210 which does 2.66TFLOPS at only $2.440 (and has 16GB).

Die info.

Knights Landing: 8B x'tors at ~683mm² --- for 85.4mm²/1B --- for 0.288 TFLOPS / (1GHz * 1B)
Pascal GP100: 15.3B x'tors at 610mm² --- for 39.9mm²/1B --- for 0.235 TFLOPS / (1GHz * 1B)

So Pascal x'tors count per mm² is more than twice that of Knights Landing, which is insane density. However, if we normalize for clock speed and for transistor count, Knights Landing, very suprisingly, does better than Pascal. KL has more performance per transistor.

Now, what would happen if Intel decide to up their game with those short and sweet cells that Apple and Nvidia know and love to use? Only assumption I make here is that Intel can achieve 1.4x x'tor density of Nvidia (so density becomes 28.5mm² per 1B x'tors), which is of course controversial to say the least.

Knights Landing 2.0:

0.288 TFLOPS / (1GHz * 1Bx'tors) * (1B x'tors / 28.5mm²) * 683mm² * 1.5GHz = 10.35 TFLOPS

Now, that would be a beast and an awesome prove of transistor lead.

I don't understand why Intel doesn't optimize for transistor count. If you're dealing with GPUs or many-cores, then you know that x'tors density is crucial. Especially with the 14nm yield problems you would expect Intel to do everything to optimize wafer area utilization.

Conclusions Pascal vs. Knights Landing

• Performance per dollar at high-end is about the same, but Nvidia has advantage because their price is for 1 unit, while Intel is 1K units
• Intel is the clear, absolute performance per dollar winner at low-end
• Intel has poor transistor per area of less than halve the amount of x'tors per area
• However, Intel actually has 23% better performance per transistor
• Nvidia has a 13% performance per watt advantage (based on TDP)

Sources

https://images.nvidia.com/content/pdf/tesla/whitepaper/pascal-architecture-whitepaper.pdf
http://www.nextplatform.com/2016/06/20/intel-knights-landing-yields-big-bang-buck-jump/
https://www.microway.com/hpc-tech-tips/nvidia-tesla-p100-price-analysis/
http://www.anandtech.com/show/9802/...s-knights-landing-xeon-phi-silicon-on-display

careful with the performance/$ comparisons. The list prices from NVIDIA and Intel are not what any serious customers looking to buy in quantity will pay.

Anyway, I don't know why you are so obsessed with these product level xtor density compares. All that matters is delivered performance and power.

 

[jpiniero]

The Tesla PCIe model is 250 W, it's the mezzanine version that's 300 W and 5.3 TF. Also remember that all of these are cut dies from 60 SM to 56.

The funny thing is that GP100 isn't all that die efficient either; Volta is likely more like how AMD handles DP.

 

[witeken]

careful with the performance/$ comparisons. The list prices from NVIDIA and Intel are not what any serious customers looking to buy in quantity will pay.

That's why I didn't put all that much emphasis on it. My reasoning is that if X buys a $100M supercomputer from Intel, they will get a certain discount, competitive offer for their volume. If Y buys a $100M supercomputer from Nvidia, they will also get a certain discount, and those things will just cancel. Based on those list prices, I don't believe pricing of the chips themselves will be the decisive factor for a customer to go with Intel or Nvidia.

Anyway, I don't know why you are so obsessed with these product level xtor density compares. All that matters is delivered performance and power.

What? I'm literally just analyzing the technology like sites as AnandTech or ExtremeTech or NextPlatform also love to do. I'm not doing any crazy job listing paranoia or using click baity titles.

 

[Arachnotronic]

That's why I didn't put all that much emphasis on it. My reasoning is that if X buys a $100M supercomputer from Intel, they will get a certain discount, competitive offer for their volume. If Y buys a $100M supercomputer from Nvidia, they will also get a certain discount, and those things will just cancel. Based on those list prices, I don't believe pricing of the chips themselves will be the decisive factor for a customer to go with Intel or Nvidia.

Intel's advantage is that it can bundle in these Xeon Phis with large Xeon E5/E7 orders. Controlling the entire platform is a valuable strategic advantage, something that can often override other competitive disadvantages.

 

[witeken]

Intel's advantage is that it can bundle in these Xeon Phis with large Xeon E5/E7 orders.

And you know that not only power and performance matters. You do the same over at Fool, calculating how much $$$ Apples saved on the 25mm² smaller area because of the 16FFC.

Things like die size, wafer cost, yield. You saw my hypothetical KL2.0: if Intel can squeeze their x'tors closer together with their manufacturing advantage -- and I don't know why they can't -- then they would be able to offer a 5.3 TFLOPS 1.5GHz Knights Landing at just 350mm². As investor, you would really want to know why Intel does not do that because if they did, they would literally be able to get 75% more dies from the same wafer, and in reality that number will be even higher because the yields would be higher and their would be less cutoff (because a wafer is round, bigger dies have more waste area), so they would produce on the order of 2x as many Knights Landing at the same cost, or a single KL would be half as cheap to make.

And never mind if they did this (Apple A10-like) high density for all their products (also Xeons). Intel must really care a lot about high performance because I don't see another reason.

Maybe Intel should do the same as Apple: invest many $$$ upfront to save a lot of manufacturing cost. So what I mean: they should invest $$$$$ to make high density Core / Atom to put in lower frequency Core / Xeon / Xeon Phi.

 

[frozentundra123456]

You're right, many of Intel's amazing products have been plagues by delays. Silicon photonics is awesome, but was delayed by about a year, but now it's out there. Knights Landing is now also out there, but a year too late, but I don't know if there are any systems yet (in TOP500), and together with it Omni Path Architecture was also delayed. I don't know how HBM is doing. Broadwell-E and Broadwell Xeon also didn't make 2015. The 3D NAND with breakthrough cost is only just showing up, and 3D XPoint won't make 2016 as promised, and I don't know if it will make Kaby Lake S. The 14nm story we know. The mobile story as well.

By dealying products, you are sucking the excitement out of them.

But in any case, Intel's manufacturing advantage is still standing strong. Each foundry has their own things they focus on. For Intel it's the transistor, both performance and density, they have air gaps, they have 2.5D I think. (For Samsung it's DRAM, for Micron it used to be NAND, but they showed up late for 3D NAND.)

You seriously believe that slide? That intel still has a 3.5 year advantage?? And going forward, other foundaries have much more optimistic roadmaps than intel. Granted, it remains to be seen if they will be able to execute on them. But it is hard to imagine they will run into any more problems than intel did with 14nm.

14nm has been an absolute disaster, just when they needed it most to be a compelling process. All we hear about is delay after delay. And now we have a revised 14nm?? Does anyone really know when and in what products we will see 10nm? In any case the "process lead" means little in and of itself. What counts is performance improvements (and product innovation actually), either in absolute performance or performance per watt, and we are seeing minimal improvements and lack of real world availability of a lot of promised products (looking at you mobile chips and also iris pro igpus). Now I dont mean this to bash Intel. I would have liked nothing better than for them to bring out a seriously competitive chip to compete with ARM in the mobile market, and to make a powerful ultrabook with iris pro graphics that doesnt throttle. But they just seem to be going nowhere both from the technical *and* product innovation perspectives.

 

[Arachnotronic]

And you know that not only power and performance matters.

It really depends on the market segment. In the server market, total cost of ownership is the metric people care about, and upfront acquisition cost is only a small part of that equation. For such products, it makes sense to trade off density for better performance/power consumption. In smartphone processors, where your bill of materials is severely constrained (even in a premium phone), being able to pack a lot of stuff into a given area is very important or you blow out your budget.

This is why Intel (and other foundries) offer different metal stacks and xtor types for different products.

Things like die size, wafer cost, yield. You saw my hypothetical KL2.0: if Intel can squeeze their x'tors closer together with their manufacturing advantage -- and I don't know why they can't -- then they would be able to offer a 5.3 TFLOPS 1.5GHz Knights Landing at just 350mm².

Not necessarily! If they could achieve this, I think they would do so.

As investor, you would really want to know why Intel does not do that because if they did, they would literally be able to get 75% more dies from the same wafer, and in reality that number will be even higher because the yields would be bigger and their would be less cutoff (because a wafer is round, bigger dies have more waste area), so they would produce on the order of 2x as many Knights Landing at the same cost, or a single KL would be half as cheap to make.

Sure, but that 8 billion xtor number...what exactly is it measuring? How are they counting the xtors? For example, if the number is a schematic xtor count, in the real world what is one transistor on paper could wind up being implemented with two transistors.

Also, when we talk about FinFETs, what is a "transistor" for the purposes of these marketing numbers? If I am using an xtor with, say, three fins in order to get better drive current, is that three transistors? Or is it one transistor? In that case, the density measurement is kind of worthless because if I end up improving my process so that I can get the same drive current from a transistor with two silicon fins rather than three silicon fins, even if I don't improve things like the contacted gate pitch/metal pitches, then have I really improved the intrinsic density of the process?

Additionally, going back to the metal stack thing I mentioned a while back, take a look at this slide from AMD:

Just by going from a tapered metal stack (thick metal layers at the top with decreasing pitch as you go lower) to a metal stack with denser/more uniform metal layers can buy you a lot of density (although in engineering nothing is free, there are trade offs).

Anyway, my point is that doing these compares when you are looking at different products from different teams is tough. That's why if you care about intrinsic process density, look at the key pitches, look at SRAM cell sizes, etc.

 

[witeken]

post

In my opinion, with a process tech with better pitches, a 2x lower x'tors density has to be explained. And now they can't come up with the composition argument, because both chips have the exact same goal of delivering a ton of FP64 FLOPS. (Sure, KL is based on Silvermont, but I thought SVM had a high density process.)

Although I would be really curious to see those composition and density graphs because obviously if Intel were to normalize for composition, then it would go from an aweful 2x density disadvantage to a good 1.4x advantage in theory. In any case, seeing how KL is no match for GP100 in terms of absolute performance, then I would argue that Intel chose the wrong composition.

I forgot to say that Nvidia also gets credits for their higher performance per watt despite process disadvantage.

 

[witeken]

You seriously believe that slide? That intel still has a 3.5 year advantage??

I believe that slide for what it says, not for what you think it says.

Intel delivered strained silicon 3 years earlier than others.
Intel delivered HKMG 3-3.5 years earlier than others.
Intel delivered Tri-gate 3 years earlier than others.
Intel will deliver III-V >3 years earlier than others.

 

[Arachnotronic]

You seriously believe that slide? That intel still has a 3.5 year advantage?? And going forward, other foundaries have much more optimistic roadmaps than intel. Granted, it remains to be seen if they will be able to execute on them. But it is hard to imagine they will run into any more problems than intel did with 14nm.

14nm has been an absolute disaster, just when they needed it most to be a compelling process. All we hear about is delay after delay. And now we have a revised 14nm?? Does anyone really know when and in what products we will see 10nm? In any case the "process lead" means little in and of itself. What counts is performance improvements (and product innovation actually), either in absolute performance or performance per watt, and we are seeing minimal improvements and lack of real world availability of a lot of promised products (looking at you mobile chips and also iris pro igpus). Now I dont mean this to bash Intel. I would have liked nothing better than for them to bring out a seriously competitive chip to compete with ARM in the mobile market, and to make a powerful ultrabook with iris pro graphics that doesnt throttle. But they just seem to be going nowhere both from the technical *and* product innovation perspectives.

14nm was mainly a problem early on because manufacturing yields simply weren't good. This led to supply issues and obviously a number of products got delayed (and the 14nm delay had a knock on effect on 10nm).

I wouldn't call it a disaster, though. For high performance products (server, laptop, desktop), I doubt you'll find anything better in the market. 14nm+ should also be a strong process, as we're seeing with Kaby Lake.

The over exuberance with respect to Intel's process lead is a bit much. At the same time, the pessimism is overdone too.

Intel is good at what it does, but the claims that it's a full generation or 3 years or whatever ahead of TSMC are just overboard and not in line with reality.

 

[Sweepr]

Great showing from Skylake @ Microsoft's latest DirectX 12 title, Forza Horizon 3:

http://gamegpu.com/racing-simulators-/-гонки/forza-horizon-3-test-gpu

Took some time but GameGPU is finally updating the CPUs they test. Dual-core Core i3-6100 still packs a punch in 2016.

 

[jpiniero]

Threading kind of sucks though, the 2600K isn't much faster than the 2500K.

 

[frozentundra123456]

Great showing from Skylake @ Microsoft's latest DirectX 12 title, Forza Horizon 3:

http://gamegpu.com/racing-simulators-/-гонки/forza-horizon-3-test-gpu

Took some time but GameGPU is finally updating the CPUs they test. Dual-core Core i3-6100 still packs a punch in 2016.

Yea, but dont know why they didnt use 6700k, especially since they are using the absolute "top of the line" 9590.

 

[Sweepr]

NBC tested the power consumption of similarly specced laptops with HD 520, Iris 550 or dGPU (940MX) from Lenovo and Acer.

http://www.notebookcheck.net/Lenovo-IdeaPad-510S-14ISK-80TK003KGE-Notebook-Review.175710.0.html

Looking at the twin Lenovo systems, the 940MX laptop draws 45% more power than the Iris 550 counterpart on load average (up to 72% @ load maximum).

Comparing the 15W (HD 520) and 28W (Iris 550) Skylake-U systems from Acer, load average reflects their TDP gap very closely.

Yea, but dont know why they didnt use 6700k, especially since they are using the absolute "top of the line" 9590.

Agreed. i7-4770K is 3 years old by now, and most gamers choose/chose i7-6700K instead of i7-6700. Hopefully i5-7600K/i7-7700K, Broadwell-E and Summit Ridge will be part of their systems in a few months.

 

[Sweepr]

Turns out Forza Horizon 3 is very CPU intensive - if you want to play at 60 FPS:

This can help to address the more sustained drops to frame-rate, but the fact is that the game still has severe streaming issues that requires sheer CPU brute force to overcome. Our Core i7 6700K overclocked to 4.6GHz and paired with 3000MHz DDR4 appears to power past most of the issues, but swapping out the i7 for a Core i5 6500 3.2GHz - similar in performance terms to a fully overclocked Core i5 2500K based on our previous tests - shows the stuttering issue in full effect.

www.eurogamer.net/articles/digitalfoundry-2016-what-does-it-take-to-run-forza-horizon-3-at-1080p60

Did anyone test it on Skylake yet? Curious about Core i5 vs Core i7 and stock vs overclocked here.


Edit: Quantum Break 'Steam' Results

 

[Sweepr]

NotebookCheck is reviewing the Skull Canyon NUC6i7KYK, hopefully we get a better idea of how Skylake GT4e compares to Broadwell GT3e when it's up. Here's a video:

3DMark 11 score is ~3500, about the same as A12-9800 at lower TDP (45W vs 65W). If past comparisons are any indication Iris Pro 580 should stack up favourably in actual games.


Also worth noting, Iris 540 vs HD 520 vs GT940M results from PurePC:

PurePC - Dell XPS 13 9350 Review

 

[Phynaz]

MSI has released BIOS's for their 100 series chipset boards allowing them to run Kaby Lake:

http://techreport.com/news/30732/msi-100-series-bios-updates-show-kaby-lake-drops-into-lga-1151

https://www.msi.com/news/detail/3rq...vRyDMrhph6n2M579HZ1bJXcBBD9iKoOzcRT-zhIckUg~~

 

[SAAA]

Finally a couple 7700K scores on geekbench:

http://browser.primatelabs.com/v4/cpu/583064
http://browser.primatelabs.com/v4/cpu/583275

Over 6100 single... 4.5GHz turbo shows!