The + is Dead - Long Live The 3! - Zen 3 information suggests big IPC and cache improvements

[porina]

Just to complicate matters a bit more, there's different levels of looking at IPC. I think what we generally refer to is some idealised "peak" IPC, how fast can the processor work if not limited elsewhere, such as lack of ram or other bandwidths. This itself can be split further, with and without SMT. At a simple level, I think this is a useful measure to understand how an architecture performs in an abstract way. If you put different instruction mixes in, you can again get very different results out.

 

Now, what if we do introduce those real world limits? You can get a different result again. How big the various layers of cache are, what is the ram configuration? As these factors start to play in the real world, the scaling starts to slip away from the peak levels predicted.

 

I had in the past (wrongly) predicted we were at a plateau in single thread performance, that what could be optimised largely already had been, and that bigger gains would be from going wide. It will be very interesting to see what Zen 3 actually brings to the table when it arrives, likewise whatever the first post-Skylake derived mainstream desktop architecture from Intel will, as I'm not looking to get it on mobile any time soon.

[mr moose]

11 hours ago, porina said:

Just to complicate matters a bit more, there's different levels of looking at IPC. I think what we generally refer to is some idealised "peak" IPC, how fast can the processor work if not limited elsewhere, such as lack of ram or other bandwidths. This itself can be split further, with and without SMT. At a simple level, I think this is a useful measure to understand how an architecture performs in an abstract way. If you put different instruction mixes in, you can again get very different results out.

 

Now, what if we do introduce those real world limits? You can get a different result again. How big the various layers of cache are, what is the ram configuration? As these factors start to play in the real world, the scaling starts to slip away from the peak levels predicted.

 

I had in the past (wrongly) predicted we were at a plateau in single thread performance, that what could be optimised largely already had been, and that bigger gains would be from going wide. It will be very interesting to see what Zen 3 actually brings to the table when it arrives, likewise whatever the first post-Skylake derived mainstream desktop architecture from Intel will, as I'm not looking to get it on mobile any time soon.

And don't forget software refinements, some software scales a lot better with SMT or more cores than others do. 

[descendency]

Regardless of what AMD is doing, their foundry is going to give them 10-15% performance gain in terms of clock speed.  Part of the reason Intel is staying on their 14nm process is because they can't get their 10nm process to produce chips that can run fast enough yet. Process refinement when jumping from node to node is important. 

[StDragon]

IPC and clock speed are different things, but are close cousins of each-other in regards to advancement in chip design. The smaller the node, the more transistors you can cram on a chip. More transistors generally equals more complexity (L1 and L2 cache excluded) which = better IPC metrics. Typical of CISC design (modern chips are really hybrid RISK/CISC designs these days). But the added bonus with smaller nodes is that you can crank up the clock speed too.

There is overlap however; specifically when comparing old CPUs to new ones. A newer slower clocked CPU can meet or exceed the performance of a faster clocked old CPU.

 

TBH, I'm rather amazed that Intel can still refine more performance out of the same 14nm process. While Intel gets many things wrong, they also get many things right!

 

 

[thechinchinsong]

On 4/8/2020 at 4:15 AM, LAwLz said:

All I did was point out that OP might be incorrect when saying Zen2 has the highest IPC on the market right now.

Is Sunny Cove on the market right now? I wasn't you could purchase Sunny Cove based products right now? Is one of the Intel 10nm laptop processors based on Sunny Cove? I'm not aware of an arch that's on the market with higher IPC than Zen2.

[leadeater]

2 hours ago, StDragon said:

But the added bonus with smaller nodes is that you can crank up the clock speed too.

Not anymore with the now extremely small nodes, not without changing fab technology methods to things like UEV or materials. The die trace resistivity goes up when you go smaller which reduces the achievable clocks, that's what Intel as been fighting with 10nm DUV.

[leadeater]

1 hour ago, thechinchinsong said:

Is Sunny Cove on the market right now? I wasn't you could purchase Sunny Cove based products right now? Is one of the Intel 10nm laptop processors based on Sunny Cove? I'm not aware of an arch that's on the market with higher IPC than Zen2.

Intel has two active current generation mobile CPUs, one is 10nm and the other is 14nm. All the Ice Lake mobile chips are 10nm.

[LAwLz]

33 minutes ago, leadeater said:

Intel has two active current generation mobile CPUs, one is 10nm and the other is 14nm. All the Ice Lake mobile chips are 10nm.

And for those wondering, ice lake uses sunny cove cores. 

[Bombastinator]

3 hours ago, StDragon said:

IPC and clock speed are different things, but are close cousins of each-other in regards to advancement in chip design. The smaller the node, the more transistors you can cram on a chip. More transistors generally equals more complexity (L1 and L2 cache excluded) which = better IPC metrics. Typical of CISC design (modern chips are really hybrid RISK/CISC designs these days). But the added bonus with smaller nodes is that you can crank up the clock speed too.

There is overlap however; specifically when comparing old CPUs to new ones. A newer slower clocked CPU can meet or exceed the performance of a faster clocked old CPU.

 

TBH, I'm rather amazed that Intel can still refine more performance out of the same 14nm process. While Intel gets many things wrong, they also get many things right!

 

 

The m-16 was a famously crap gun when it came out in the 1960’s but they put 30 years of development and billions of dollars into optimizing it.  No real surprise it works good now.  Think how much better ryzen might be with the same treatment.  What would a ryzen cpu looki like if it could clock to 5.1?

[bcredeur97]

does this mean I have to get off of sandy bridge-E?

[StDragon]

23 minutes ago, Bombastinator said:

The m-16 was a famously crap gun when it came out in the 1960’s but they put 30 years of development and billions of dollars into optimizing it.  No real surprise it works good now.  Think how much better ryzen might be with the same treatment.  What would a ryzen cpu looki like if it could clock to 5.1?

x86 is that M-16 in CPU land. But at some point you just have to step back, reassess, and start over with a new design. Intel attempted that with the Itanium which flopped due to lack of market adoption. Hence it was named the "Itanic" shortly thereafter. The momentum for x86 is so strong that it will be here for...forever actually, even if it has to be emulated!

I don't believe ARM will be the future even though Apple plans on making ARM the core of its new CPU to handle the future of a single OS that combines both branches of iOS and OSX. No doubt in my mind that Apple will unify everything with a single CPU and OS combo.

That said, if the PC is to migrate away from X86 (and let's be honest, it won't unless there's a "brick wall" in the way in terms of additional development), the only other possible architecture I could image would be RISK-V with the massive backing it had already received. But backing and adoption are two entirely separate things. I don't expect the PC to embrace RISK-V until after mobile devices do first. The current outlook so far looks to be pure fantasy now, but it's the closest fantasy the future of PC hardware will ever get to change away from X86.

[Bombastinator]

16 minutes ago, StDragon said:

x86 is that M-16 in CPU land. But at some point you just have to step back, reassess, and start over with a new design. Intel attempted that with the Itanium which flopped due to lack of market adoption. Hence it was named the "Itanic" shortly thereafter. The momentum for x86 is so strong that it will be here for...forever actually, even if it has to be emulated!

I don't believe ARM will be the future even though Apple plans on making ARM the core of its new CPU to handle the future of a single OS that combines both branches of iOS and OSX. No doubt in my mind that Apple will unify everything with a single CPU and OS combo.

That said, if the PC is to migrate away from X86 (and let's be honest, it won't unless there's a "brick wall" in the way in terms of additional development), the only other possible architecture I could image would be RISK-V with the massive backing it had already received. But backing and adoption are two entirely separate things. I don't expect the PC to embrace RISK-V until after mobile devices do first. The current outlook so far looks to be pure fantasy now, but it's the closest fantasy the future of PC hardware will ever get to change away from X86.

There was a point that x86 ran faster emulated then on metal. Wasn’t a very long period, but it happened.  The witcheriii recently came out on the switch and is doin well.  Isn’t that ARM?

[StDragon]

13 minutes ago, Bombastinator said:

The witcheriii recently came out on the switch and is doin well.  Isn’t that ARM?

The Nintendo Switch is based on the nVidia SOC (System On a Chip) called Tegra X1. It basically combines an ARM CPU with an nVidia GPU.

 

I'm sure that's why Apple is looking to design their own ARM based CPU so the code can be cross compatible with an SOC (A-Series) equivalent for mobile devices.

[mr moose]

15 minutes ago, StDragon said:

 

 

I'm sure that's why Apple is looking to design their own ARM based CPU so the code can be cross compatible with an SOC (A-Series) equivalent for mobile devices.

Apple do design all their own SoC, they have done since the beginning. 

 

EDIT: or do you mean ARM based desktop CPU for mac?

[StDragon]

1 minute ago, mr moose said:

 

EDIT: or do you mean ARM based desktop CPU for mac?

ARM based for Mac.

 

https://www.macworld.com/article/3534793/apples-planning-an-aggressive-shift-from-intel-to-arm-processors-oft-correct-analyst-claims.html

[Bombastinator]

42 minutes ago, StDragon said:

The Nintendo Switch is based on the nVidia SOC (System On a Chip) called Tegra X1. It basically combines an ARM CPU with an nVidia GPU.

 

I'm sure that's why Apple is looking to design their own ARM based CPU so the code can be cross compatible with an SOC (A-Series) equivalent for mobile devices.

Apple doesn’t play nice with Nvidia though.  Are you saying you think there will be a change in that?

[StDragon]

1 hour ago, Bombastinator said:

Apple doesn’t play nice with Nvidia though.  Are you saying you think there will be a change in that?

So the Apple A10 was the last SoC that used PowerVR for the GPU portion of the chip. A11 and current are still ARM based, but the GPU is Apple's own design. The API to access it is Metal; their equivalent of Microsoft's DirectX or the OpenGL platform.

 

Not sure if Apple licensed any IP to build their own GPU or not.

[williamcll]

According to digitimes, Zen 3 will arrive in September, bearing codenamed Vermeer

[TechyBen]

On 4/10/2020 at 11:24 PM, Bombastinator said:

There was a point that x86 ran faster emulated then on metal. Wasn’t a very long period, but it happened.  The witcheriii recently came out on the switch and is doin well.  Isn’t that ARM?

Um, it's pretty lower specced. Apples and oranges.

 

A better example would be pi calculations or one of the video encoding benchmarks (which shows ARM can get close at times). Though as others have said, I'd expect RISCV in some form (proprietary spin off later adopted by the entire industry?) to be the real starting ground of a spin off. ARM seems more low power focused, and thus is a specialist, not jack of all trades we ask from general compute/CPU/Server/PCs.

[imreloadin]

On 4/6/2020 at 10:05 AM, 5x5 said:

According to AdoredTV's sources

Aaaaaaaand you lost me there...

[ravenshrike]

On 4/10/2020 at 4:53 PM, Bombastinator said:

The m-16 was a famously crap gun when it came out in the 1960’s but they put 30 years of development and billions of dollars into optimizing it.  No real surprise it works good now.  Think how much better ryzen might be with the same treatment.  What would a ryzen cpu looki like if it could clock to 5.1?

No, just... no. The issues with the M16A1 were twofold. They were not issued cleaning kits, which combined with the lack of chrome lined barrels in the extremely humid jungle environments they were fighting in led to corroded barrels, and the ordinance department bought the wrong ammo, which caused severe failure to eject issues. The gun itself was fine.

[Bombastinator]

54 minutes ago, ravenshrike said:

No, just... no. The issues with the M16A1 were twofold. They were not issued cleaning kits, which combined with the lack of chrome lined barrels in the extremely humid jungle environments they were fighting in led to corroded barrels, and the ordinance department bought the wrong ammo, which caused severe failure to eject issues. The gun itself was fine.

Heh.  “The gun it’s self”. A weapon system is a weapon system.  The cho cho was very similar.  “The gun itself” was fine there too.  Still famous as being exceptionally bad though, because of which guns were tested with which round.
 

What you just described were problems with the weapon system.    How much did it cost and how much time did it take to find those issues?  These are exactly the kinds of problems I am talking about.  Money and time.

 

there were other issues which required modifying the weapon as well.  It’s a very different weapon than it was in the 1960’s.

 

I also note you specifically said m16a1, whereas I specifically said m16.  The a1 is important.  A1 was a revision.  How much time and money did that cost?

 

So “Yes.  Just yes” if you want to go with binary  statements.  I’m not fond of them myself.
 

Did it turn out that a big problem with the m16 was with the bullets. (Or rounds, cartridges, etc.. people often call em bullets though).  Fixing that required a years long alteration to both military inventory world wide and changes to the entire procurement process. You seem to be trying to differentiate that out.  I’m not.

[ravenshrike]

1 hour ago, Bombastinator said:

What you just described were problems with the weapon system.    How much did it cost and how much time did it take to find those issues?  These are exactly the kinds of problems I am talking about.  Money and time.

Except you were talking about the x86 instruction set itself. Which would be the rifle itself. Expecting a modern x86 CPU to do well with no on-die cache or with a insufficient VRM is moronic.

[Bombastinator]

16 minutes ago, ravenshrike said:

Except you were talking about the x86 instruction set itself. Which would be the rifle itself. Expecting a modern x86 CPU to do well with no on-die cache or with a insufficient VRM is moronic.

This is hilarious.  I used an example, and you attacked the concept based on the example.  I pointed out that your attack on the example was a failure and you go back to the original point claiming I made claims I did not make.