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

[RejZoR]

9 hours ago, LAwLz said:

Well I am talking about IPC here, so clock speed doesn't matter.

I am pretty sure Sunny Cove has higher IPC than Zen2.

 

Sunny Cove is on supposedly on average around 18%* faster than Skylake at the same clock speed, and up to 40%* in some specific cases (for example Cinebench).

Again, not factoring in AVX-512. If you take that into account it's somewhere between 100%* and 150%* faster than Whiskey Lake.

 

*Based on Intel's measurements.

 

 

I think it would be more accurate to say Zen2 is the current desktop IPC leader. And if we want to be more pedantic, when AVX-512 is not used.

You don't get it either. You can't make conclusions on IPC basis when AMD can't do 5GHz with ANY of their products. Even if intel has worse IPC, they make up that difference with higher clock.

[MageTank]

1 minute ago, RejZoR said:

Erm, no. Can any AMD do 5GHz at any point? No. Intel can. So, how are you going to compare IPC clock to clock when AMD can't do 5GHz at all? By downclocking Intel to 4 point something? Yeah, no. Sure you get raw IPC metric that way, but you don't get realistic performance numbers. And that's the point I was making. Intel might not have class leading IPC, but they make up for that with clock. For now. They can't just increase clocks to infinity though as 5GHz is already stupid high for a clock...

Great, now people are going to invoke the FX 9590 and we will all be forced to remember what it did to that poor MSI 970 Krait. 

[5x5]

23 minutes ago, MageTank said:

Great, now people are going to invoke the FX 9590 and we will all be forced to remember what it did to that poor MSI 970 Krait. 

Hello, darkness, my old friend. I've come to light you on fire again...

 

[5x5]

26 minutes ago, RejZoR said:

You don't get it either. You can't make conclusions on IPC basis when AMD can't do 5GHz with ANY of their products. Even if intel has worse IPC, they make up that difference with higher clock.

Yes you can and yes they can. Remember, people got Zen 2 to 5GHz on LN2 - it was notably faster in single-core than Intel's CPUs. So at both 4GHz AND 5GHz - Intel have lower IPC

[leadeater]

25 minutes ago, MageTank said:

Great, now people are going to invoke the FX 9590 and we will all be forced to remember what it did to that poor MSI 970 Krait. 

Well since you bring it up........

Spoiler

Nothing to see here move along

 

[RejZoR]

10 minutes ago, 5x5 said:

Yes you can and yes they can. Remember, people got Zen 2 to 5GHz on LN2 - it was notably faster in single-core than Intel's CPUs. So at both 4GHz AND 5GHz - Intel have lower IPC

Because LN2 is totally a practical solution for 24/4 right? Guess what, those 9900K's do run at 5GHz 24/7 on most common AiO's, so nothing expensive or clumsy. Yeah, your example just doesn't work.

[leadeater]

19 minutes ago, 5x5 said:

Yes you can and yes they can. Remember, people got Zen 2 to 5GHz on LN2 - it was notably faster in single-core than Intel's CPUs. So at both 4GHz AND 5GHz - Intel have lower IPC

Well I hate to nit pick but IPC is actually independent of clock speed, Zen 2 would be higher even at 1MHz.

 

Thing is it's actually far more complex than just IPC or clock rate or the function of just those two, that is an incomplete performance picture. We may even be misidentifying the performance differences as IPC, there are actually other possibilities and ones that may be more technically correct.

 

 

Quote

The Performance Equation
The performance equation analyzes execution time as a product of three factors that are relatively independent of each other.

 

This equation remains valid if the time units are changed on both sides of the equation. The left-hand side and the factors on the right-hand side are discussed in the following sections.

 

The three factors are, in order, known as the instruction count (IC), clocks per instruction (CPI), and clock time (CT). CPI is computed as an effective value.

 

The Performance Equation
The performance equation analyzes execution time as a product of three factors that are relatively independent of each other.

 

Quote

Instruction Count

Computer architects can reduce the instruction count by adding more powerful instructions to the instruction set. However, this can increase either CPI or clock time, or both.

Clocks Per Instruction

Computer architects can reduce CPI by exploiting more instruction-level parallelism. If they add more complex instructions it often increases CPI.

Clock Time

Clock time depends on transistor speed and the complexity of the work done in a single clock. Clock time can be reduced when transistor sizes decrease. However, power consumption increases when clock time is reduced. This increase the amount of heat generated.

Instruction Count

Instruction (IC) count is a dynamic measure: the total number of instruction executions involved in a program. It is dominated by repetitive operations such as loops and recursions.

 

Instruction count is affected by the power of the instruction set. Different instruction sets may do different amounts of work in a single instruction. CISC processor instructions can often accomplish as much as two or three RISC processor instructions. Some CISC processor instructions have built-in looping so that they can accomplish as much as several hundred RISC instruction executions.

 

For predicting the effects of incremental changes, architects use execution traces of benchmark programs to get instruction counts. If the incremental change does not change the instruction set then the instruction count normally does not change. If there are small changes in the instruction set then trace information can be used to estimate the change in the instruction count.

 

For comparison purposes, two machines with different instruction sets can be compared based on compilations of the same high-level language code on the two machines.

Clocks Per Instruction

Clocks per instruction (CPI) is an effective average. It is averaged over all of the instruction executions in a program.

 

CPI is affected by instruction-level parallelism and by instruction complexity. Without instruction-level parallelism, simple instructions usually take 4 or more cycles to execute. Instructions that execute loops take at least one clock per loop iteration. Pipelining (overlapping execution of instructions) can bring the average for simple instructions down to near 1 clock per instruction. Superscalar pipelining (issuing multiple instructions per cycle) can bring the average down to a fraction of a clock per instruction.

 

For computing clocks per instruction as an effective average, the cases are categories of instructions, such as branches, loads, and stores. Frequencies for the categories can be extracted from execution traces. Knowledge of how the architecture handles each category yields the clocks per instruction for that category.

Clock Time

Clock time (CT) is the period of the clock that synchronizes the circuits in a processor. It is the reciprocal of the clock frequency.

 

For example, a 1 GHz processor has a cycle time of 1.0 ns and a 4 GHz processor has a cycle time of 0.25 ns.

 

Clock time is affected by circuit technology and the complexity of the work done in a single clock. Logic gates do not operate instantly. A gate has a propagation delay that depends on the number of inputs to the gate (fan in) and the number of other inputs connected to the gate's output (fan out). Increasing either the fan in or the fan out slows down the propagation time. Cycle time is set to be the worst-case total propagation time through gates that produce a signal required in the next cycle. The worst-case total propagation time occurs along one or more signal paths through the circuitry. These paths are called critical paths.

 

For the past 35 years, integrated circuit technology has been greatly affected by a scaling equation that tells how individual transistor dimensions should be altered as the overall dimensions are decreased. The scaling equations predict an increase in speed and a decrease in power consumption per transistor with decreasing size. Technology has improved so that about every 3 years, linear dimensions have decreased by a factor of 2. Transistor power consumption has decreased by a similar factor. Speed increased by a similar factor until about 2005. At that time, power consumption reached the point where air cooling was not sufficient to keep processors cool if the ran at the highest possible clock speed.

https://www.d.umn.edu/~gshute/arch/performance-equation.xhtml

 

Differences in instructions sets and compiler optimizations could result in the same task requiring less Instructions (IC) which would appear as higher IPC (CPI) without digging deeper.

[RejZoR]

We know what IPC is. It's work done per 1Hz. One cycle. Not 1GHz, not 4GHz, not 5GHz. Just running it all at certain GHz is just an equalization to get results easier. But some don't seem to understand that IPC is not an absolute defining factor of performance. You can have 50% better IPC than competitor, but if your clocks are only hitting 2.5GHz and competitor is hitting 5GHz, you're essentially having the same final performance. But it sound awesome on paper, 50% BETTER IPC THAN COMPETITION!!!!!1111oneoneone

I like to make fun of Intel's incompetence at the moment, but their 5GHz is 5GHz that AMD cannot reach at all under any practical circumstances even on same core count CPU's if one might argue that more cores are more taxing to power delivery and OC. And those 5GHz is what's saving Intel. They don't have to invest ANY R&D into IPC if they are confident they can push clocks even higher. Which is probably easier than redesigning entire compute logic of the CPU. So, yeah...

[5x5]

2 minutes ago, RejZoR said:

We know what IPC is. It's work done per 1Hz. One cycle. Not 1GHz, not 4GHz, not 5GHz. Just running it all at certain GHz is just an equalization to get results easier. But some don't seem to understand that IPC is not an absolute defining factor of performance. You can have 50% better IPC than competitor, but if your clocks are only hitting 2.5GHz and competitor is hitting 5GHz, you're essentially having the same final performance. But it sound awesome on paper, 50% BETTER IPC THAN COMPETITION!!!!!1111oneoneone

I like to make fun of Intel's incompetence at the moment, but their 5GHz is 5GHz that AMD cannot reach at all under any practical circumstances even on same core count CPU's if one might argue that more cores are more taxing to power delivery and OC. And those 5GHz is what's saving Intel. They don't have to invest ANY R&D into IPC if they are confident they can push clocks even higher. Which is probably easier than redesigning entire compute logic of the CPU. So, yeah...

Dude chill - the reason Intel can't sustain the clockspeed chace is because they already use much more power than AMD CPUs for the same performance. Hence the comparison with FX and P4

[RejZoR]

Why would I need to "chill"? It's people writing dumb shit and pissing on Intel on some stupid metrics. We can all laugh at Intel, even I do, but not at expense of logic and realistic performance metrics (like the above "but muh LN2" example to claim Ryzen can do 5GHz, it's like reversed Intel's chiller fiasco). Hence, stop praising IPC as the only performance defining factor.

[leadeater]

20 minutes ago, RejZoR said:

We know what IPC is. It's work done per 1Hz. One cycle. Not 1GHz, not 4GHz, not 5GHz. Just running it all at certain GHz is just an equalization to get results easier. But some don't seem to understand that IPC is not an absolute defining factor of performance. You can have 50% better IPC than competitor, but if your clocks are only hitting 2.5GHz and competitor is hitting 5GHz, you're essentially having the same final performance. But it sound awesome on paper, 50% BETTER IPC THAN COMPETITION!!!!!1111oneoneone

I like to make fun of Intel's incompetence at the moment, but their 5GHz is 5GHz that AMD cannot reach at all under any practical circumstances even on same core count CPU's if one might argue that more cores are more taxing to power delivery and OC. And those 5GHz is what's saving Intel. They don't have to invest ANY R&D into IPC if they are confident they can push clocks even higher. Which is probably easier than redesigning entire compute logic of the CPU. So, yeah...

Really it makes no difference that Intel can clock to 5GHz and AMD can't, that is literally just an arbitrary number. There is no negative to AMD not being able to do 5GHz at all, work done is work done.

 

Thing is for Intel it's superior to increase IPC than it is to increase clocks because we are not making gains like we were from Pentium to Pentium 2 to Pentium 3 to Pentium 4 anymore. If IPC is 1 and clock rate is 5 you could increase either to increase performance, but right now it would actually be simpler to increase IPC than it would clock rate. Increase IPC by 10% to 1.1 at the same clock and the effective equivalent to do the opposite is a 0.5 (500MHz) increase. Do you really think Intel can do that? Then how much higher is it across multiple cores and not just one or a few.

 

IPC increases have no conditions like all core boost thus it is a superior way of increasing performance compared to clocks today when we are fab limited and package power constrained. Increasing IPC doesn't additionally increase power leakage. Increasing IPC however does increase power which could necessitate lower clocks i.e AVX offset (not exactly the same thing but illustrates the point).

 

Intel is good at generally improving both IPC and clocks and they try and balance that as much as possible, clocks raise the tide for all where IPC increase may be case specific.

[MageTank]

2 minutes ago, RejZoR said:

Why would I need to "chill"? It's people writing dumb shit and pissing on Intel on some stupid metrics. We can all laugh at Intel, even I do, but not at expense of logic and realistic performance metrics (like the above "but muh LN2" example to claim Ryzen can do 5GHz, it's like reversed Intel's chiller fiasco). Hence, stop praising IPC as the only performance defining factor.

Mostly because you get easily worked up in many posts when the subject matter is not all that serious. Even now, you are arguing a point that nobody is disagreeing with because you are not completely reading in to what everyone is saying. Nobody is contesting that Intel's performance advantage comes from their superior clock speeds. I'll even go as far to say that their refined IMC is playing a key part in that as well, but there is no denying that AMD's core architecture is superior if we are adjusting for a specific power envelope or equalized clock speeds.

 

To boil it down, AMD at 4ghz is currently faster than Intel at 4ghz. Intel at 5ghz is currently faster than AMD at 4ghz. These two statements do not contradict each other in any way, and both are true. You are not wrong, nor are the people you are crusading against. Even those that are taking the time to offer insight on the misconception of IPC are not out to refute your claims, but to merely provide context to the situation so that the rest of us can better understand both sides.

 

I too think the fanboy gloating and company trash talking is getting old, but I make it a point not to feed into it when I can. At the end of the day, I just want the best bang for my buck, and it's up to the companies to compete against each other to earn my dollar.

[leadeater]

1 minute ago, MageTank said:

At the end of the day, I just want the best bang for my buck, and it's up to the companies to compete against each other to earn my dollar.

Funny thing is what's stopping me from buy a new CPU is GPU pricing, I'm not doing one without the other at this point and right now pricing is a joke so nobody has gotten my money yet.

[GDRRiley]

15 minutes ago, leadeater said:

Funny thing is what's stopping me from buy a new CPU is GPU pricing, I'm not doing one without the other at this point and right now pricing is a joke so nobody has gotten my money yet.

as someone who went CPU now GPU later. CPU upgrades can be pretty nice if you do work that can use it. video editing is soo much nicer with a 2700 over a 6600K given both run for me a 4.0ghz (i know I can get more out of the intel chip I just have possibly the worst z170 board made)

[MageTank]

14 minutes ago, leadeater said:

Funny thing is what's stopping me from buy a new CPU is GPU pricing, I'm not doing one without the other at this point and right now pricing is a joke so nobody has gotten my money yet.

Yeah... I am in a similar boat. I have a failing motherboard that has been too abused by a rabid memory overclocker, a 2080 Ti that doesn't support the HDMI 2.1 I need to run my OLED TV at 4k 120hz and nothing to really upgrade to at the moment.

 

When an HDMI 2.1 GPU launches, I plan on upgrading my GPU, but not without upgrading platforms as well, but neither side is really winning me over. If I go with Intel, I am getting more of the same thing (without any real benefits, because my 8700k can already do 5.2ghz all-core 24/7) and if I go AMD, I get additional cores that I do not need for my intended workloads and likely a compromise on gaming performance in my older, single threaded titles.

 

I really want DDR5 to show up, but I know for a fact it's going to be another year minimum before we even see them start to trickle down to consumer platforms. If AMD made overclocking a little more fun, I'd probably jump ship just to tinker for a while, but after my experience with tweaking the 3990X and getting my rear end handed to me by PBO doing a better job, I don't feel inclined to make that switch just yet.

[LAwLz]

2 hours ago, RejZoR said:

You don't get it either. You can't make conclusions on IPC basis when AMD can't do 5GHz with ANY of their products. Even if intel has worse IPC, they make up that difference with higher clock.

Why are you mad at me? All I did was point out that Zen2 isn't the architecture with the highest IPC right now.

I was not talking about absolute performance of a chip. I was just talking about IPC. Frequency is irrelevant when talking about IPC. Performance is kind of irrelevant as well because of the reasons you have pointed out.

 

We are not arguing about the same thing here.

I was talking about IPC.

You accuse me of "not getting it" because you think IPC isn't a good measurement of performance (which I agree, it isn't).

[RejZoR]

IPC alone tells you nothing. Clock alone tells you nothing. Clock and IPC together tell you something. You don't gain 15-20% IPC gains by slight tweaks. You need to do quite radical modifications in the compute pipeline to reach that. Is that really financially sustainable? Intel apparently thought it's not, which is why they opted for the clock game. They maybe made slight tweaks to caches and they gained a 1% here and 1% there and some with new specific instructions, but their main performance gains came from increased clocks.

[5x5]

1 minute ago, RejZoR said:

IPC alone tells you nothing. Clock alone tells you nothing. Clock and IPC together tell you something. You don't gain 15-20% IPC gains by slight tweaks. You need to do quite radical modifications in the compute pipeline to reach that. Is that really financially sustainable? Intel apparently thought it's not, which is why they opted for the clock game. They maybe made slight tweaks to caches and they gained a 1% here and 1% there and some with new specific instructions, but their main performance gains came from increased clocks.

And that's why they are currently rehashing 5-year-old tech, pushing large amounts of power, heat and clocks all in the effort to catch up with a modern, efficient design. Again, if Intel hadn't stagnated and focused on clocks, they wouldn't be backed into a corner now with Skylake+++++

[porina]

1 hour ago, RejZoR said:

IPC alone tells you nothing. Clock alone tells you nothing. Clock and IPC together tell you something. You don't gain 15-20% IPC gains by slight tweaks. You need to do quite radical modifications in the compute pipeline to reach that. Is that really financially sustainable? Intel apparently thought it's not, which is why they opted for the clock game. 

You write as it Sunny Cove doesn't exist. Well, if you only care about desktop, it doesn't, but Intel do have a "much higher IPC than Skylake" architecture. Just that you have to get Ice Lake mobile if you want to get your hands on it right now. For mainstream desktop it is looking like we're into 2021 before they catch up with Zen 3/Zen 4, but I'm still kinda hopeful we get something really new on server/HEDT to play with towards the end of this year.

 

Intel's current woes are not in design, it's in manufacturing.

[mr moose]

5 hours ago, RejZoR said:

Erm, no. Can any AMD do 5GHz at any point? No. Intel can. So, how are you going to compare IPC clock to clock when AMD can't do 5GHz at all? By downclocking Intel to 4 point something? Yeah, no. Sure you get raw IPC metric that way, but you don't get realistic performance numbers. And that's the point I was making. Intel might not have class leading IPC, but they make up for that with clock. For now. They can't just increase clocks to infinity though as 5GHz is already stupid high for a clock...

 

You are missing the point, (especially when you tried to claim it was work done at 1 hz)  IPC is not changed by total clock speed.  Overall performance is not dictated by IPC alone nor is it dictated by clock speed alone, IPC is just one metric of many.  Putting two CPU's side by side at a set speed does not change the performance one can get out of the box if one operates faster.

 

Again IPC is Instructions Per Cycle,  that means if you increase the number of cycle yous increase the instructions performed in that time frame (the whole point of overclocking).  If Intel can cycle 300Hz faster out of the box then it's IPC goes up by that factor.   Thus IPC is not the equalizing metric everyone seems to think it is.  It is just another metric that needs to be properly understood before it means anything.

 

 

 

[Bombastinator]

3 hours ago, LAwLz said:

Why are you mad at me? All I did was point out that Zen2 isn't the architecture with the highest IPC right now.

I was not talking about absolute performance of a chip. I was just talking about IPC. Frequency is irrelevant when talking about IPC. Performance is kind of irrelevant as well because of the reasons you have pointed out.

 

We are not arguing about the same thing here.

I was talking about IPC.

You accuse me of "not getting it" because you think IPC isn't a good measurement of performance (which I agree, it isn't).

I would say it’s incomplete.  Slightly different than “not good”. Computer performance can be difficult to rate.  It’s why there are so many different benchmark systems.

[Bombastinator]

3 hours ago, RejZoR said:

IPC alone tells you nothing. Clock alone tells you nothing. Clock and IPC together tell you something. You don't gain 15-20% IPC gains by slight tweaks. You need to do quite radical modifications in the compute pipeline to reach that. Is that really financially sustainable? Intel apparently thought it's not, which is why they opted for the clock game. They maybe made slight tweaks to caches and they gained a 1% here and 1% there and some with new specific instructions, but their main performance gains came from increased clocks.

Which is why 10 year old variable clock CPUs like mine can continue to be even vaguely relevant.

[StDragon]

I loath the X570 chipset for needing a fan. I really hope the next one after will be passively cooled.

[LAwLz]

12 hours ago, RejZoR said:

IPC alone tells you nothing. Clock alone tells you nothing.

I am not sure why you are telling me this because I have never argued otherwise.

 

Again, this is how the conversation went:

OP: Zen2 is the IPC king.

Me: Pretty sure Sunny Cove has higher IPC.

You: Intel has high frequency and a lack of IPC.

Me: Pretty sure Sunny Cove has higher IPC than Zen2.

You: You don't get it! Even if Intel has worse IPC they make up for it with higher frequency!

 

Do you understand what I mean now? You're arguing something completely different. All I did was point out that OP might be incorrect when saying Zen2 has the highest IPC on the market right now. That was it. I never made any comment about performance or frequency. Do you understand?

You come off as someone who just learned that IPC is and eager to tell everyone about it to sound smart. I know what IPC is so you don't have to tell me.

 

All I did was point out that Intel isn't behind AMD when it comes to IPC, which some people seem to think. It entirely depends on which CPUs you're comparing. If you want to be mad at someone, be mad at the people trying to derive real world performance by looking at just IPC and disregarding frequency. But guess what, I am not one of those people. I have made zero claims about performance all throughout this thread. All I have said is that Intel has an architecture that seems to have higher IPC than AMD's architecture. Nothing more, and nothing less.

If I wanted to talk about performance, which I haven't done, then I would have showed or referenced some benchmarks.

[mr moose]

1 hour ago, LAwLz said:

I am not sure why you are telling me this because I have never argued otherwise.

 

 

It seems to me he initially had no idea how IPC worked as a metric (to be honest I am not sure many do beyond a very rudimentary and course position). I say this because 1Hz has nno bearing on one clock cycle nor does it define IPC in any meaningful way.