Skylake S SiSoft Benchmark: Confusion and 30% IPC Improvement Over Haswell?[UPDATE]

[Deletive]

These threads turn into essays by the 3rd page

[patrickjp93]

These threads turn into essays by the 3rd page

That's what spirited debate often requires. @Opcode I would consider passionate if a bit sensationalist and overtly Pro-AMD, but he(she?)'s worth having these conversations with.

[ViruZ_HUN]

Is there still chance that it will have ddr4 support?

I personally think DDR4 will be only for the extreme CPUs for a while. DDR3 is still mainstream and most popular ram. DDR4 will take atleast 2-3 years to get mainstream

[patrickjp93]

Yeah but...

...TL;DR ;P

Sometimes reading is worth it. I'm a student buzzing around all the time and I still have the time and patience to craft and write my own responses as well as fully evaluate others'. This is why gamers and tech nerds generally look stupid. They only care about cliff notes, but then they miss the whole point.

[patrickjp93]

I personally think DDR4 will be only for the extreme CPUs for a while. DDR3 is still mainstream and most popular ram. DDR4 will take atleast 2-3 years to get mainstream

Nah. DDR3 prices have been steadily rising lately, and DDR4 has fallen to half its launch price already. By the launch of Skylake in June/July, it'll be at a fair consumer price. Besides, AMD's Zen will have to use it or it'll look like a dead-end platform, and that's going to launch early 2016 as well.

[Opcode]

These threads turn into essays by the 3rd page

As it should, one long post that makes use of touching multiple points rather than a bunch of single posts touching base here and there.

 

That's what spirited debate often requires. @Opcode I would consider passionate if a bit sensationalist and overtly Pro-AMD, but he(she?)'s worth having these conversations with.

Not so much Pro-AMD but more over intrigued in the technology involved. Nvidia shares quite a few of these technologies but doesn't compete in the #APUMasterRace.

[patrickjp93]

I did some digging about IPC since people have been complaining about Intel's small gains of late. The following assumes perfect scaling, which we know is never the case, though we would expect better scaling than Haswell, so the real results should be closer to this theoretical approximation than a similar calculation from IB to HW.

 

This is based on SiSoft's existing 4770k scores.

 

The hyperthreaded Haswell quad-core here (4770K) runs at 3.7 GHz with turbo boost and attains a score of 114.58 GFLOPs.

 

This Skylake quad-core presumably running at 2.3 GHz attains a score of 91.8 GFLOPs

Haswell's clock speed (3.7 GHz) divided by Skylake's (2.3 GHz) = 1.6087 times higher clock speed

Skylake's arithmetic score (91.8) multiplied by 1.6087 = 147.68 GFLOPs

147.68 GFLOPs (Skylake's theorized score at 3.7 GHz) divided by the equivalent Haswell score at the same clock speed (114.58) = 1.29x higher IPC, again based only on these assumptions of accuracy and scaling. I know SiSoft is far from a perfect source of specs and conditions, but for now it's all we have.

 

Also, the same calculations of the iGPU scores against the 4770k igpu suggests a 56% throughput increase per core assuming perfect scaling with the given 300MHz clock. We don't know the RAM speeds unfortunately.

[vandidant]

Finally time to upgrade from our i5 750's and i7 920's.

[Opcode]

Finally time to upgrade from our i5 750's and i7 920's.

Good job necroing a thread from last year...

[deviant88]

OP somethings wrong? you make the thread sounds like its CPU IPC when its actually GPU ICP performance ~30%,i cant see anything else in the post or pdf about CPU all i can see is GPU improvements,is this what is has come to at Intel? all we can talk about is feeding more good transistors into a useless iGPU instead of giving us 6 core i5's and 8 core i7 for the same price?

 

Im totally against APU's(except for mobile) idk why is this even an excitement,when i will have to make an upgrade i will forced again to buy an APU i dont need?seems like its heading that way,hopefully amd will make pure CPU for desktops again with Zen i cant stand Intel's shit anymore,all that space and transistors wasted on i3/5/7 for GPU could have been extra core and cache i still cant believe how intel is ripping off and no one seems to care.

[LukaP]

@LukaP, I don't think I've forgotten how to do basic math. Can you make 184 Stream Processors make sense in all this given what we know of Intel's graphics architecture?

no, but i can make sense of it if they changed it up (which we all know they arent afraid of doing with their GPU arch) 

 

But with a improved 14nm TriGate (or possibly already an Omega-Gate) process, i wouldnt be surprised in seeing them do a 46*8SIMD=184SIMD design*

 

*this is of course purely hypothetical.

 

also with Haswells 0.83 IPC that i remember you claiming (havent checked it) this would make Skylake have an IPC of 0.83*1.288=1.069 which as far as theory of simple instructions goes is impossible. Unless of course they are using SIMD/MIMD models in SiSoft

[vandidant]

Good job necroing a thread from last year...

lol I wanted to discuss Skylake. What am I to do, start a new thread? What's your problem with posting in old existing threads that have relevant information already there?

 

I searched for Skylake IPC and this was a top result.

[vandidant]

OP somethings wrong? you make the thread sounds like its CPU IPC when its actually GPU ICP performance ~30%,i cant see anything else in the post or pdf about CPU all i can see is GPU improvements,is this what is has come to at Intel? all we can talk about is feeding more good transistors into a useless iGPU instead of giving us 6 core i5's and 8 core i7 for the same price?

 

Im totally against APU's(except for mobile) idk why is this even an excitement,when i will have to make an upgrade i will forced again to buy an APU i dont need?seems like its heading that way,hopefully amd will make pure CPU for desktops again with Zen i cant stand Intel's shit anymore,all that space and transistors wasted on i3/5/7 for GPU could have been extra core and cache i still cant believe how intel is ripping off and no one seems to care.

I agree that the focus on gpu is rather stupid, they are wasting resources reinventing the wheel that ATI and Nvidia had already mastered as far as I'm concerned. And as far as I'm concerned, power savings don't matter. If you're buying a chip with gpu that you won't use, you're throwing your money away. That's all there is to it. I hope Skylake will have Xeon's without gpu again.

 

I don't see how this IPC improvement is only for gpu, the gflop performance numbers imply cpu performance. If it were gpu, then I'm sure the rest of this thread would have a very different tone towards Intel.

[patrickjp93]

I agree that the focus on gpu is rather stupid, they are wasting resources reinventing the wheel that ATI and Nvidia had already mastered as far as I'm concerned. And as far as I'm concerned, power savings don't matter. If you're buying a chip with gpu that you won't use, you're throwing your money away. That's all there is to it. I hope Skylake will have Xeon's without gpu again.

I don't see how this IPC improvement is only for gpu, the gflop performance numbers imply cpu performance. If it were gpu, then I'm sure the rest of this thread would have a very different tone towards Intel.

iGPUs are extremely useful for heterogeneous acceleration without the delay of sending stuff off to the big iron dGPU. All Xeons will have iGPU this time around. Intel's doing it to use OpenMP and OpenACC to counter HSA.

[vandidant]

iGPUs are extremely useful for heterogeneous acceleration without the delay of sending stuff off to the big iron dGPU. All Xeons will have iGPU this time around. Intel's doing it to use OpenMP and OpenACC to counter HSA.

Does that mean that Skylake will be able to use integrated and discrete simultaneously, not just switching between them?

[LukaP]

Does that mean that Skylake will be able to use integrated and discrete simultaneously, not just switching between them?

not like you are thinking about it. both can be viewed as coprocessors, but as PCIe is slow and riddled with latency, its not suitable to host a coprocessor of the same calibre as an iGPU is. so you will probably use the iGPU for some lighter tasks that need minimal latency, and dGPU for big tasks that are usually what GPUs do

[patrickjp93]

Does that mean that Skylake will be able to use integrated and discrete simultaneously, not just switching between them?

You've always been able to with OpenMP and OpenACC. The problem since 2006 has been software developers, not CPUs.

 

 

not like you are thinking about it. both can be viewed as coprocessors, but as PCIe is slow and riddled with latency, its not suitable to host a coprocessor of the same calibre as an iGPU is. so you will probably use the iGPU for some lighter tasks that need minimal latency, and dGPU for big tasks that are usually what GPUs do

 

Very true, although Intel has some bigger plans when it comes to the Skylake Xeons. If on the E5/E7 line they host 144 EUs as rumored that's 1192 Stream processors. At 1 GHz that's 1192*(2+1)*1*10^9 = 3.576 TFlops SP, 1.788 TFlops DP in just iGPU alone.

 

This paired up with their KNL chips means a Hell of a boost in supercomputing power. At some point heterogeneous acceleration will trickle down to consumer software. I'll be one of those to make sure of that.

[Opcode]

Does that mean that Skylake will be able to use integrated and discrete simultaneously, not just switching between them?

This has been possible for a while. The stretch is eventually all modern consumer microprocessors will come loaded with an iGPU. When that becomes a reoccurring trend for both markets then you will start seeing games leveraging the iGPU for game logic meanwhile the discrete card does all the rendering. There are a few AMD tech demo's that's been around for a little while now demonstrating the iGPU being leveraged for physics meanwhile the dGPU handles the rendering resulting in a performance advantage.