Intel New CPU Architecture 2019

[TechguyMaxC 516]

https://techreport.com/review/34319/intel-talks-about-its-architectural-vision-for-the-future/2

 

Intel is finally focusing on increasing single thread performance once again.  I expect to see performance increases of AT LEAST 25% across the board as a result of these architectural changes.  It may require a code re-compile, but not a re-write.  And that's a very important distinction, a re-compile is trivial where a re-write may not be.  

[TechguyMaxC 516]

I was on the fence about Coffee Lake Refresh (9900k), now I've made up my mind to wait for Sunny Cove.  Forget 8 cores with IPC that's been available since 2015, I want more IPC!

[Guest]

The biggest issue facing Intel is "can they deliver in 2019"?  

But take note of some of the changes:

1.  Large L1/L2 Cache is good, but at what frequency?  As I found out with the 9900K, increasing the cache frequency was a huge benefit, so much so that I reduce CPU frequency so that I could increase cache frequency for better overall performance.  So there is definitely something to be gained there and that will not require a recompile.

2.  AVX-512 instruction set extensions ... this is where a recompile will be needed provided the compilers (i.e Visual Studio get updates).  Unfortunately AVX is not really used much by any current day flight simulators.  And, it's hard to OC and retain AVX at the same frequency, produces a lot of heat ... so I would imagine AVX-512 to produce even more heat so the downstep in AVX frequency might be even higher of an offset.

So that projected 25% single core improvement in practical terms for flight simulators is probably going to be more like 10-15%.  But every little bit helps.  10% of 28 FPS (+2.8 FPS) hits that magical 30 FPS smoothness.

AMD's next gen CPU may still have the upper hand and will probably beat Intel on delivery dates.  Lets see what really happens on the pricing front, my gut is fixed pricing to avoid price war and ensure profits (even if illegal to openly do so).

So look forward to see what 2019 brings, AMD or Intel.

Cheers, Rob.

[TechguyMaxC 516]

Rob, the 25% figure is derived purely from architectural changes, enhancements to the front-end of the CPU which impacts virtually all workloads one can concoct.  AVX 512 was literally not a factor in my estimate.  This is supported by my earlier statement that a re-compile may be necessary to extract the most performance from existing code, but not a re-write.  Intel has made execution resources 25% wider generally, and throughput on specific instructions has increased by 50%, 100%, and even 638% in the case of integer division.

Here is a bit more of a deep dive on Sunny Cove: https://www.anandtech.com/show/13699/intel-architecture-day-2018-core-future-hybrid-x86/2

Commentary on much of the technology Intel previewed yesterday from David Kanter, a well-regarded CPU analyst: 

 

 

In the above video if you pay close attention, David Kanter mentions that he does not expect clock speeds to decrease, though Intel was vague on this matter (citing a 10-15% transistor performance increase, but not clarifying whether this translates into power savings or clock speed increase for the products in question).  

As for AMD, their next-gen CPUs are slated to have a "13% IPC improvement" in the average case.  They may be winning the core count race but as we simmers know well, core count is not the most important factor of CPU performance for flight simulator. 

There's enough evidence here to say that we may be looking at more than a 25% performance improvement, I just don't want to go out on a limb and start hyping people up when we're still 6-9 months away from launch.  

 

[TechguyMaxC 516]

More evidence that Intel's next-gen architecture will immediately bring performance uplifts to all software from the man responsible for the new core architecture:

This is really shaping up to be a winner folks.  I for one would not recommend the purchase of a new CPU today unless you absolutely have to (i.e.due to component failure)

[Guest]

56 minutes ago, TechguyMaxC said:

I for one would not recommend the purchase of a new CPU today

Not sure I agree with you there, Intel have a problem of delivery and nasty habit of missing timelines not by just a few months but in some cases year(s).  Although I do like what I see on the performance front for next gen Intel ... for someone like me, I swap out CPU's/MBs pretty regularly so even if Intel hit their delivery date (and actually have sufficient stock) we're still looking at sometime summer of 2019 ... a lot can happen in 6 months on the technology front.

Cheers, Rob.

[Luke 782]

On 12/12/2018 at 3:26 PM, TechguyMaxC said:

Rob, the 25% figure is derived purely from architectural changes, enhancements to the front-end of the CPU which impacts virtually all workloads one can concoct.  AVX 512 was literally not a factor in my estimate.  This is supported by my earlier statement that a re-compile may be necessary to extract the most performance from existing code, but not a re-write.  Intel has made execution resources 25% wider generally, and throughput on specific instructions has increased by 50%, 100%, and even 638% in the case of integer division.

If it relies on new instructions and a recompile by software authors, it's not going to go anywhere. This reminds me of the days with new instructions in the 80286 and 80386 that we simply couldn't take advantage of without multiple code paths and versions, and all of the attending support nightmares. Unless you're in a very specific niche where the need for performance is almost unlimited and the user base asymptotically approaches 1, this is a non-starter.

Where I see significant possibilities is in managed code. If you're running in MSIL or on a JVM, you're not actually compiling your app to machine code, just an intermediate version that can be executed on the client machine using whatever capabilities the local processor has. I could see a .NET VM taking advantage of these paths on a particular processor family, but again it's a question on Microsoft's end whether they want to go down that road and what the cost/benefit ratio would be.

Let's face it, single-thread performance is a niche, getting smaller and smaller every day, intersecting with the equally small and declining niche of compiling code directly to machine instructions.

Cheers!

[TechguyMaxC 516]

4 hours ago, Luke said:

If it relies on new instructions and a recompile by software authors, it's not going to go anywhere. This reminds me of the days with new instructions in the 80286 and 80386 that we simply couldn't take advantage of without multiple code paths and versions, and all of the attending support nightmares. Unless you're in a very specific niche where the need for performance is almost unlimited and the user base asymptotically approaches 1, this is a non-starter.

Where I see significant possibilities is in managed code. If you're running in MSIL or on a JVM, you're not actually compiling your app to machine code, just an intermediate version that can be executed on the client machine using whatever capabilities the local processor has. I could see a .NET VM taking advantage of these paths on a particular processor family, but again it's a question on Microsoft's end whether they want to go down that road and what the cost/benefit ratio would be.

Let's face it, single-thread performance is a niche, getting smaller and smaller every day, intersecting with the equally small and declining niche of compiling code directly to machine instructions.

Cheers!

The performance potential of Sunny Cove is not reliant upon the execution of new instructions (the new extensions introduced are geared towards acceleration of specific workloads, such as cryptography and machine learning).  Neither is a re-compile required to realize these performance gains.  I only mentioned re-compilation because any time a drastic change is made to a CPU front-end, it is inevitable that compilers will receive updates to take advantage of these changes in order to produce code that runs best on these machines.  A re-compile for Sunny Cove may result in say 30% more performance over unoptimized code which will still run something like 25% faster.  No one knows for sure the exact percentage yet.

Single thread performance is not "a niche". The vast majority of applications that are user interactive have a single heavyweight thread.  Do you know why that is?  You can parallelize a "wait for input" command until you overflow virtual address space but it won't complete any faster.  Besides, if you have multi-thread code, a new CPU architecture will accelerate EVERY THREAD, not just the first one.  Not sure how you can overlook that in your assessment...  

Single thread performance raises all boats.  Multi-thread performance is heavily dependent on both the workload and the efficiency of the code.  

[TechguyMaxC 516]

5 hours ago, Rob Ainscough said:

Not sure I agree with you there, Intel have a problem of delivery and nasty habit of missing timelines not by just a few months but in some cases year(s).  Although I do like what I see on the performance front for next gen Intel ... for someone like me, I swap out CPU's/MBs pretty regularly so even if Intel hit their delivery date (and actually have sufficient stock) we're still looking at sometime summer of 2019 ... a lot can happen in 6 months on the technology front.

Cheers, Rob.

You sure you're not thinking of AMD, Rob?  Historically, they're the tech company that promises the moon and always falls short.  Just look at Vega.  Or the FX CPUs (Bulldozer).  Or Phenom.  Or Radeon HD 2900.  Or any CPU prior to the K7 Athlon. 

Aside from the delay of 10nm, when was the last time Intel failed to deliver on their product roadmap?  The fact that Sunny Cove is going to be built on 10nm and will launch next year tells us that Intel is confident these troubles are in the past.  

I know you just bought a 9900k and spent a lot of time and money getting it running at a clock speed with which you are happy, but in ~9 months it's going to get beat soundly by Sunny Cove.  That's progress, and I for one am happy for it.  

[Guest]

5 hours ago, TechguyMaxC said:

but in ~9 months it's going to get beat soundly by Sunny Cove.

And if that's the case, I'll be buying one ... unless AMD bring game.  I built my PC with the understanding I was going to be swapping out components, hence why I used so many quick disconnects in the tubing.

Cheers, Rob.

[Luke 782]

15 hours ago, TechguyMaxC said:

A re-compile for Sunny Cove may result in say 30% more performance over unoptimized code which will still run something like 25% faster.  No one knows for sure the exact percentage yet.

If it's more than 10%, I'll be shocked. There's a world of difference between early marketing materials and then what makes it out into production. Intel has been caught out several times by challenges with chip architecture and fabrication issues, so I'll believe it when I see it.

15 hours ago, TechguyMaxC said:

Single thread performance is not "a niche". The vast majority of applications that are user interactive have a single heavyweight thread.  Do you know why that is?  You can parallelize a "wait for input" command until you overflow virtual address space but it won't complete any faster.

If you're waiting for input, then you're clocked down to 800Mhz or whatever the chip will scale down to. LOL

Apps are single-threaded because for 99% of the use cases, performance is already "fast enough" and what you need is simplicity in the code base over an optimization that only the programmer cares about. The vast majority of times, your apps are waiting for user input, I/O or some other long-blocking tasks, and the compute-intensive portions run for maybe 10-15ms at a time. Shaving off a millisecond doesn't matter.

I'll stand by what I said earlier - anything valuable that was performance sensitive has by and large been ported to multi-core and scales with core count, not single-thread speed. The reason why most stuff is still dependent on a single thread is because it's a simpler coding mechanism and it's already fast enough.

Cheers!

 

[TechguyMaxC 516]

12 hours ago, Rob Ainscough said:

And if that's the case, I'll be buying one ... unless AMD bring game.  I built my PC with the understanding I was going to be swapping out components, hence why I used so many quick disconnects in the tubing.

Cheers, Rob.

Fair enough.  I've played that game for a long time now but have been holding steady with my 7700k @ 5.2GHz for almost 2 years now (the longest I've kept any single CPU in a mainstream desktop build used primarily for flight sim).  I was tempted to go with Coffee Lake, but held off.  Then Coffee Lake Refresh came out and I was tempted again, but held off.  In both cases I believe that was the right decision given the announcement of Sunny Cove.  

I wonder for others though who might be tempted to upgrade today but will not be in the position to upgrade again in 9 months.  Would you recommend they upgrade today, or hold off?

[TechguyMaxC 516]

1 hour ago, Luke said:

If it's more than 10%, I'll be shocked. There's a world of difference between early marketing materials and then what makes it out into production. Intel has been caught out several times by challenges with chip architecture and fabrication issues, so I'll believe it when I see it.

If you're waiting for input, then you're clocked down to 800Mhz or whatever the chip will scale down to. LOL

Apps are single-threaded because for 99% of the use cases, performance is already "fast enough" and what you need is simplicity in the code base over an optimization that only the programmer cares about. The vast majority of times, your apps are waiting for user input, I/O or some other long-blocking tasks, and the compute-intensive portions run for maybe 10-15ms at a time. Shaving off a millisecond doesn't matter.

I'll stand by what I said earlier - anything valuable that was performance sensitive has by and large been ported to multi-core and scales with core count, not single-thread speed. The reason why most stuff is still dependent on a single thread is because it's a simpler coding mechanism and it's already fast enough.

Cheers!

 

I'll be shocked if it's less than 15%.  Intel has had the time and the resources to develop a competitive answer to Zen/Zen 2.  They've been in a similar position in the past with Netburst vs. K7/K8 and have learned their lesson.  Some may view Intel's products of recent years as underwhelming, I'm inclined to think of them as evidence of sandbagging, if anything.  

As for the importance of single-thread performance in many compute workloads (or in the specific case of 3d flight simulation), I think you and I are coming from different places.  I don't know the specifics of your background nor the extent of your knowledge on the subject, but I can state with certainty that you are wrong when it comes to the importance of single-thread performance for simulators (or more broadly, 3d games). 

Not all problems can be parallelized.  The number of problems that can be parallelized easily is fewer.  The smartest minds in the history of computer science have yet to crack this egg. You can setup 1000 parallel assembly lines to build as many cars as you like, and those assembly lines can churn out cars at a fantastic volume.  However, increasing the number of assembly lines will not cause a single car to be built any faster than if there was only 1 assembly line.

[Luke 782]

3 hours ago, TechguyMaxC said:

I can state with certainty that you are wrong when it comes to the importance of single-thread performance for simulators (or more broadly, 3d games).

I think we both agree on the importance of single-thread performance to current simulations, but my point is that simulators are an extreme niche that in no way are driving processor design. The center of gravity is towards lower power consumption and more cores, or some combination of the two.

Cheers!

[jabloomf1230 2,445]

5 hours ago, Luke said:

If it's more than 10%, I'll be shocked.

I just took the Passmark single thread CPU table and adjusted the top 80 entries for a clock of 4.4GHz, which is probably a reasonable  OC on H2O for any processor with a base clock over 3.5 GHz. Here's a summary of increases in score.

9900 - baseline

9700 - 2%

8700 -10%

7700 - 23%

6700 - 37%

Looks like Luke (sounds like a Dr. Seuss rhyme, sorry ) is more or less correct, except that the 9900 can hit higher OCs over 5GHz. But that's not the standard way that most people would run their systems.

Edited December 15, 2018 by jabloomf1230
grr spellchecker