Ryzen 2 IPC improvements measured

[Taja]

56 minutes ago, mynameisjuan said:

Reactions on LTT

 

Intel: "New processors have 10-15% IPC improvement- FUCK YOU INTEL AND YOUR MONEY GRABBING. GIVE US REAL IMPROVMENTS!

 

AMD: "New chips have 7% better IPC- HOLY SHIT THATS AMAZING! AMD IS THE BEST COMPANY EVER!!"

Well, if intel processors werent THE SAME with a little clock boost... That would be right.

 

But its more like: intel 1-3% IPC gains + clock gains amounting to 10% +-

 

This new ryzen is around 10% in just IPC, we are hoping for better clocks also... And it is not an exciting product, it is average, but for a refresh is not bad - unlike kaby lake was (which is what most people shit intel on)

[Mira Yurizaki]

1 hour ago, SpaceGhostC2C said:

Are you honestly telling me there was a 10-15% increase between generations since Sandy Bridge? In other words, are you telling me that Kaby Lake has between 60% and 100% higher IPC than Sandy Bridge (or 46-75% if we just don't count Broadwell because, well, Broadwell. In which case it would still be 60-100% for Coffee Lake vs. Sandy Bridge)?

 

I think you grossly overestimate it, but I'll be happy to see sources that support that claim.

I also think the 10% IPC claim for Ryzen+, while definitely "pretty good" if true, is not obvious looking at the available information. I could easily see Ryzen+ having a 10% or higher boost in performance, but IPC? Gaming benchmarks aren't the best tool for that.

I went out to find this: https://www.anandtech.com/show/9483/intel-skylake-review-6700k-6600k-ddr4-ddr3-ipc-6th-generation/11. But I'll bite, 15% is overestimating, since as far as the normalized single threaded performance goes, 10% is optimistic across the board. Though the i7-3770K and i7-4770K is an interesting case because they both have the same specs (clock speed, core count, and cache), so that's a perfect generation to compare IPC performance.

 

But I'm reeling in that people seem to be claiming that clock speed is winning the performance gains today, something that attributed to the massive performance gains in the 90s and 2000s that people seem to desire today.

 

Either way, sure, AMD gaining anything seems impressive when Intel isn't doing much, but until AMD is matching Intel, they're playing catch up as far as I'm concerned. Until someone beats Intel outright, I can't really shower them with praise. Maybe it's impressive with the resources they have, but I'm looking at things from an absolute performance point of view.

[Blademaster91]

1 hour ago, The Benjamins said:

I didn't think it was needed, this thread is about IPC improvements.

The  Ryzen+ improvements come from cache and memory latency, working better with RAM is nice to see though I still don't see this as an upgrade if someone already has a Ryzen cpu.

1 hour ago, mynameisjuan said:

Reactions on LTT

Intel: "New processors have 10-15% IPC improvement- FUCK YOU INTEL AND YOUR MONEY GRABBING. GIVE US REAL IMPROVMENTS!

AMD: "New chips have 7% better IPC- HOLY SHIT THATS AMAZING! AMD IS THE BEST COMPANY EVER!!"

A new chip that shows a 7% gain? OMG I'M GOING TO PREORDER!

[Taf the Ghost]

54 minutes ago, Camoxide said:

I'm surprised there was any gains to be honest considering this is still based on 'Zen 1', the really improvements should be next year with Ryzen 3/Zen 2

It's the first revision of a brand new architecture. There's always "easier" places to improve. It's the 3rd generation that's always harder.

[SpaceGhostC2C]

11 minutes ago, M.Yurizaki said:

I went out to find this: https://www.anandtech.com/show/9483/intel-skylake-review-6700k-6600k-ddr4-ddr3-ipc-6th-generation/11. But I'll bite, 15% is overestimating, since as far as the normalized single threaded performance goes, 10% is optimistic across the board. Though the i7-3770K and i7-4770K is an interesting case because they both have the same specs (clock speed, core count, and cache), so that's a perfect generation to compare IPC performance.

That shows at most¹ a ~40% increase in performance between Skylake's 6700K and Sandy Bridge's 2600K, which puts it below 10% per generation (~7%). But performance is not IPC: aside of things that may have changed between platforms, the clocks themselves went from 3.4GHz for the 2600K to 4.0GHz for the 6700K. I emphasize this because the discussion wasn't about performance gains but IPC gains specifically, which compounded with the increase in clocks speeds would imply more drastic boosts to overall performance.

 

Unfortunately, I'm not Ananadtech and I cannot run all those tests  So, as a back of the envelope calculation, I will just assume performance = instructions, and therefore Performance = IPC*frequency (is what everyone does when testing two CPUs at the same clock speeds for "IPC tests" anyway). The increase in clocks was 4.0/3.4-1=17.65%. Therefore, the implied average IPC between generations would be

 

(1.4-0.1765)^1/5-1 = 4%.

 

Of course, that average reflects some instances of no gains and some instances of larger gains.

 

11 minutes ago, M.Yurizaki said:

 

But I'm reeling in that people seem to be claiming that clock speed is winning the performance gains today, something that attributed to the massive performance gains in the 90s and 2000s that people seem to desire today.

Well, it has been the case for the couple last releases (except Ryzen, because it was flipping FX upside down). Still, to me, ultimately it is performance that matters, or at most performance per watt, it doesn't really matter if they achieve it running 10GHz chips, 1GHz super IPC chips, moving the memory controller in, taking cache out, or whatever.

 

11 minutes ago, M.Yurizaki said:

Either way, sure, AMD gaining anything seems impressive when Intel isn't doing much, but until AMD is matching Intel, they're playing catch up as far as I'm concerned. Until someone beats Intel outright, I can't really shower them with praise. Maybe it's impressive with the resources they have, but I'm looking at things from an absolute performance point of view.

I frankly don't see the point in showering them with praise, or criticism, or anything. I just want to understand which products are out there, so everything that translates as faster or cheaper is good news to me. I mean, it's not like Intel or AMD are eagerly waiting for SpaceGhostC2C's pat on the back or snarky remarks

 

--

¹The results vary by test, lowest I saw was 16%, but most were int eh 36-42% range.

[ravenshrike]

2 hours ago, SpaceGhostC2C said:

Of course, that average reflects some instances of no gains and some instances of larger gains.

It's funnier than that. In gaming there's been effectively no increase in IPC for Intel since Haswell. It can all pretty much be attributed to memory improvements and clock speed.

https://www.anandtech.com/show/9483/intel-skylake-review-6700k-6600k-ddr4-ddr3-ipc-6th-generation/10

 

 

[Sakkura]

3 hours ago, M.Yurizaki said:

I went out to find this: https://www.anandtech.com/show/9483/intel-skylake-review-6700k-6600k-ddr4-ddr3-ipc-6th-generation/11. But I'll bite, 15% is overestimating, since as far as the normalized single threaded performance goes, 10% is optimistic across the board. Though the i7-3770K and i7-4770K is an interesting case because they both have the same specs (clock speed, core count, and cache), so that's a perfect generation to compare IPC performance.

 

But I'm reeling in that people seem to be claiming that clock speed is winning the performance gains today, something that attributed to the massive performance gains in the 90s and 2000s that people seem to desire today.

 

Either way, sure, AMD gaining anything seems impressive when Intel isn't doing much, but until AMD is matching Intel, they're playing catch up as far as I'm concerned. Until someone beats Intel outright, I can't really shower them with praise. Maybe it's impressive with the resources they have, but I'm looking at things from an absolute performance point of view.

Your link is a bit off - a couple pages back there is an IPC comparison with the CPUs set to the same clocks. Haswell to Skylake was a 5.7% IPC increase, and bear in mind that's two steps - both a tick and a tock. Sandy to Ivy Bridge, a tick, was a very similar IPC increase, which was quite good for a tick (also why Intel called it a tick+ at the time). Ivy Bridge to Haswell is clearly the biggest IPC jump of course, but that was a tock.

 

What AMD is getting here, with a tick, seems to be at least as good as Intel's tick+ on Ivy Bridge, and possibly even better. So that's why people are a bit surprised.

[vanished]

5 hours ago, M.Yurizaki said:

How is 10% "pretty nice"?

Because (at least personally) I was expecting more like 0, but even Intel has only managed about half that gen over gen according to the results I've looked at:

Specifically, observe the "IPC Data" section.  To be fair, this analysis could potentially capture improvements due to memory as well so it, if anything, is an overestimate.  I read ahead and saw you claiming 10 - 15% on average and I have no idea where you got that number.  Their overall performance usually doesn't even increase that much from one gen to the next, never mind IPC alone.

[Taf the Ghost]

56 minutes ago, Ryan_Vickers said:

Because (at least personally) I was expecting more like 0, but even Intel has only managed about half that gen over gen according to the results I've looked at:

Specifically, observe the "IPC Data" section.  To be fair, this analysis could potentially capture improvements due to memory as well so it, if anything, is an overestimate.  I read ahead and saw you claiming 10 - 15% on average and I have no idea where you got that number.  Their overall performance usually doesn't even increase that much from one gen to the next, never mind IPC alone.

Since Sandy Bridge, Intel has mostly been in the process that we've seen in Car manufacturing. The top speeds haven't really improved much in years, but the way you get to those speeds (both in time & efficiency) has changed quite a lot. That's why the actual IPC gain has been fairly minimal, but lots of little things are significantly faster. Though because Intel was on Quad Cores for about 9 years, much of that benefit really didn't show up for the consumer, but it paid off in huge returns in the Server space.

[SpaceGhostC2C]

1 hour ago, ravenshrike said:

It's funnier than that. In gaming there's been effectively no increase in IPC for Intel since Haswell. It can all pretty much be attributed to memory improvements and clock speed.

The thing with games, and why I don't like them as benchmarks for anything except for playing those games itself, is that you are always testing the game together with the hardware, as the vast heterogeneity across games show. You run 3 games and you will obtain different numbers. Sometimes you won't even get the same sign. In some games, feeding the GPU is all the CPU does, while in others there's a lot to do beyond graphics...

(it's also interest their discussion about apparent performance regression and the potential explanations)

 

I don't know, I guess I can get philosophical about what the technical capabilities of a CPU are vs. what a particular program is able to achieve using a given hardware. I tend to prefer understanding, on one hand, how well does a CPU perform a number of generic CPU functions, and on the other hand, how a particular software uses the different functions, in which combination, etc. But it's also harder  

 

Still, I didn't notice they had the "same clock tests" in a different page. If I go to the non-gaming tests, Anandetch's IPC calculations are:

Quote

Sandy Bridge to Ivy Bridge: Average ~5.8% Up
Ivy Bridge to Haswell: Average ~11.2% Up
Haswell to Broadwell: Average ~3.3% Up
Broadwell to Skylake (DDR3): Average ~2.4% Up
Broadwell to Skylake (DDR4): Average ~2.7% Up

which implies an average 4.5% increase. Comparing this with @Ryan_Vickers, I see my back-of-the-envelope wasn't too far off 

In the end, it's easy to get confused by small numbers and the magic of exponential growth. No one's gonna freak out over a 3% difference between two things, but a 3-point difference in steady growth rates over long periods of time can compound to extremely large differences over time.

 

 

I think the bottom line is that IPC isn't a matter of engineers showing up at work every day for a year. Major changes on how a CPU works depend on research, and truly groundbreaking research outcomes happen sparsely and unpredictably. You can count on some refinements to take place more or less routinely, but other than that... It's not about being lazy or stupid, as we would hear people say with Intel's nth iteration or AMD's Bulldozer. I mean, we don't ask drive manufacturers every year why our HDDs aren't more than 10% faster than last year, or why they "sat on their asses" after the jump from HDD to SSD...

[vanished]

40 minutes ago, Taf the Ghost said:

Since Sandy Bridge, Intel has mostly been in the process that we've seen in Car manufacturing. The top speeds haven't really improved much in years, but the way you get to those speeds (both in time & efficiency) has changed quite a lot. That's why the actual IPC gain has been fairly minimal, but lots of little things are significantly faster. Though because Intel was on Quad Cores for about 9 years, much of that benefit really didn't show up for the consumer, but it paid off in huge returns in the Server space.

Yeah, IPC gains have been marginal, but (stock) clockspeed has been going up a lot.  However, the max reachable clockspeed assuming a good OC has also seen only marginal improvements - roughly 4.6 back in the sandy bridge days to around 5.2 now on coffee lake, if I'm not mistaken.

[AnonymousGuy]

12 hours ago, Taf the Ghost said:

Coffee Lake's 14nm++ is actually a pretty big jump in efficiency over Kaby Lake's 14nm+ process. That's why the Thin/Light & Laptop spaces are seeing a big jump in raw performance with the 8th Gen Intel parts. I'm still not sure how Intel's engineers found that much of a lower-clock efficiency improvement within a node improvement, but it's actually quite impressive in that space. In the Desktop, the addition of 50% more cores was always going to be a little limiting, but the top SKU parts clock well at good efficiency. 

 

Now, Skylake-X is a different issue.

ChipWiki has a description, off the top of my head 14nm++ loosened the feature sizes up to take more silicon space but get better performance.  The yields on 14nm are so high now that there's less need to pack as many die on the wafer as possible.

[Taf the Ghost]

4 minutes ago, Ryan_Vickers said:

Yeah, IPC gains have been marginal, but (stock) clockspeed has been going up a lot.  However, the max reachable clockspeed assuming a good OC has also seen only marginal improvements - roughly 4.6 back in the sandy bridge days to around 5.2 now on coffee lake, if I'm not mistaken.

The big changes have as much to do with DDR4 and the mainstreaming of SSDs over the time frame, as well.

https://www.anandtech.com/show/9483/intel-skylake-review-6700k-6600k-ddr4-ddr3-ipc-6th-generation/10

 

Though it's probably worth looking at this quote:

 

Quote

There’s no easy way to write this.

Discrete graphics card performance decreases on Skylake over Haswell.

This doesn’t particularly make much sense at first glance. Here we have a processor with a higher IPC than Haswell but it performs worse in both DDR3 and DDR4 modes. The amount by which it performs worse is actually relatively minor, usually -3% with the odd benchmark (GRID on R7 240) going as low as -5%. Why does this happen at all?

Interesting statement, for certain.

 

Though, I honestly wonder if what we're really talking about, most of the time, is Nvidia's Driver implementation more than anything else.

 

[Taf the Ghost]

@Ryan_Vickers

 

https://www.techspot.com/review/1490-ryzen-vs-core-i7-vega-64-geforce-1080/

 

I forgot how nutty the results in this piece by Steve ended up being.

[Canada EH]

41 minutes ago, Ryan_Vickers said:

5.2 now on coffee lake

and what for Ryzen?

[vanished]

8 minutes ago, Canada EH said:

and what for Ryzen?

~4

[porina]

14 hours ago, SpaceGhostC2C said:

That shows at most¹ a ~40% increase in performance between Skylake's 6700K and Sandy Bridge's 2600K, which puts it below 10% per generation (~7%). But performance is not IPC: aside of things that may have changed between platforms, the clocks themselves went from 3.4GHz for the 2600K to 4.0GHz for the 6700K. I emphasize this because the discussion wasn't about performance gains but IPC gains specifically, which compounded with the increase in clocks speeds would imply more drastic boosts to overall performance.

There's one more factor to consider. The addition of new instructions enables certain tasks to be performed faster. I am cherry picking my niche, but in prime number finding using Prime95 and similar software (LLR, PFGW), there has been a significant effective peak IPC improvement over the generations. Sandy Bridge with AVX bought about 2x peak IPC increase over previous generations. Haswell added 50% on that from FMA3. Skylake another 14% through optimisation. Now we're on the cusp of another possible 2x factor with AVX-512. In case you wonder what I mean by peak IPC, that is the case where other factors are not significantly limiting. In particular, ram bandwidth can be a significant limiting factor depending on the specific ram and CPU configuration.

[Misanthrope]

I thought the 12nm would be mostly BS but it seems like there's something there. I just might get a 2600x

[Sharkyx1]

23 hours ago, Princess Cadence said:

Not at all, it still gives us a fair idea of raw gains one can expect... Certainly not enough to justify someone with a Ryzen 7 or an i5/i7 14nm Intel CPU to upgrade just yet but this refresh so far has been proving to be more meaningful than the Radeon RX 500 series was which is a gain to us all.

Right now benches show unless you have something more than a 1060/1070 and even then you have little reason to upgrade over any core i5 after first gen especially not ryzen which while cool trades punches with a 2500k in games

[RagnarokDel]

23 hours ago, Princess Cadence said:

Not at all, it still gives us a fair idea of raw gains one can expect... Certainly not enough to justify someone with a Ryzen 7 or an i5/i7 14nm Intel CPU to upgrade just yet but this refresh so far has been proving to be more meaningful than the Radeon RX 500 series was which is a gain to us all.

that's actually debatable. Because Zen+ should perform better with overclocked ram and for many people, it's still an ongoing issue. My PC isnt stable if I put my ram at it's rated 3200mhz. It's close(I'll get a freeze maybe once a week) but not quite so I keep it at 2933mhz.

[Doobeedoo]

That actually look very good. Cache latency and memory improvements. 

[TacoSenpai]

Definitely good but I think in my situation I'll save the extra cash for the next round of GPUs that come out.