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

[mr moose]

5 hours ago, 5x5 said:

Just going to add - the reasons are lengthy and expensive legal issues that bled AMD dry of money as they were fighting Intel in courts over Intel's anti-competitive business practices.

Not going to pretend Intel's anti consumer behavior didn't have it's effects (because it certainly did) but the truth is that AMD's practices at the time played a much larger role in their poor financials than suing Intel did.  In fact their courts costs etc were nothing next to their purchase of ATI in 2006.  And the mass failure of their server product market also had nothing to do with Intel.

[Blademaster91]

An IPC improvement would put Zen 3 over Intel, however plenty of software still likes clockspeed, then again Zen 2 is already close enough to Intel with more efficiency.

But I don't know how people can still take adoredtv as a source, considering they have been at least 50% wrong with CPU and GPU "leaks".

[strajk-]

>According to AdoredTV's sources

Stopped reading there.

[Way2Broke]

28 minutes ago, comander said:

I want it. 

My 3900x is overkill for my current use cases but I want it. 

It'll likely be bought used in a few years and the Zen2 CPU will go into a server somewhere. 


----

I should send AdoredTV a "leak" from my company email stating that the forthcoming X product will be y% better than the last one due to Z technology that's novel and impressive. He'd eat it up. It'd be a lie, but he'd eat it up. 

Scottish accent: Alright, I would like to bring to you a very important announcement that I have from a reliable source. Now, this could change but I'm pretty confident that in 69 days we'll be in for something special. 

Overkill is better than underkill since you spend that $$$ on a cpu. Intel is still king in clockspeed.

[mr moose]

55 minutes ago, Blademaster91 said:

 

But I don't know how people can still take adoredtv as a source, considering they have been at least 50% wrong with CPU and GPU "leaks".

And contrary to what people want to believe, 50% is a guess.  It is exactly the same probability as a guess, which means for any given prediction,  anyone on this forum would have exactly the same accuracy.  I don't even understand why people watch his videos let alone take him seriously. 

 

On topic.  It's nice to see AMD getting back on top of things over the last few years and I am genuinely excited to see what the next generation of products bring.  But where they land performance wise still remains to be seen.  I mean IPC is becoming such an ambiguous metric when different workloads perform so differently.    There is no point in jumping up and down about IPC in multi core process when your program only uses one thread and different CPU's clock very differently.

[leadeater]

2 hours ago, Curufinwe_wins said:

Always is the wrong word, but given a lack of other mitigating factors/refinements, yes.

Well I don't think you need it, if you're comparing different cache types or different fabs caches then all you need is a slap. 32MB L3 cache from Intel on Skylake would be slower than 16MB L3 cache from Intel on Skylake. Larger is always slower.

 

AMD has taken two 16MB slices and made it one 32MB slice, it'll be slower. We're only talking very small ns extra though.

[Deli]

5 hours ago, trag1c said:

Thats exactly where I am at with my 5820k. I was going to look at a 3950x or one of the thread rippers but I might wait now lol.

So many people have 5820K here.

 

My 5820K has served its purpose in my main rig, now sits quietly in my HTPC.

[Rune]

So...wait til zen 4. Okay.

Edit: Oh, AdoredTv. This is all bullshit then.

[Curufinwe_wins]

1 hour ago, Deli said:

So many people have 5820K here.

 

My 5820K has served its purpose in my main rig, now sits quietly in my HTPC.

That's cause it was offered at a great price (barely more than a 4790k on sale) and overclocked to be a great all around chip, minus the whole x99 motherboards being more expensive. But even then, if we shopped the same feature set, it wasn't much more hahaha.

 

Says another proud 5820k owner. 

 

2 hours ago, leadeater said:

Well I don't think you need it, if you're comparing different cache types or different fabs caches then all you need is a slap. 32MB L3 cache from Intel on Skylake would be slower than 16MB L3 cache from Intel on Skylake. Larger is always slower.

 

AMD has taken two 16MB slices and made it one 32MB slice, it'll be slower. We're only talking very small ns extra though.

Almost certainly yes, but again tons of mitigating factors (which are worth saying), you take the cache substructure from a current ARM uArch and compare it to smaller caches from a few years ago and the latency will be noticeably faster at all access sizes.

 

Obviously Intel is very much fighting the margins right now on cache optimization (given the ++++), and I don't expect AMD will find some magic sauce to overcome the size difference... but it is technically possible that it would end up with a radical redesign that is straight up faster. Zen3 isnt going to be on the same process node as Zen2 for example.

[mr moose]

Trying really hard to find some decent articles on all of this,  but all I can find are sources saying "according AMD" this test result is awesome and then "according to wccftech" that new CPU will be a billion times faster and of course most of this goes back to adored just making shit up for clicks so I don't think it's worth even discussing at this point,  we may as well use a ouji board and discuss the results from that.

[leadeater]

1 hour ago, Curufinwe_wins said:

Almost certainly yes, but again tons of mitigating factors (which are worth saying), you take the cache substructure from a current ARM uArch and compare it to smaller caches from a few years ago and the latency will be noticeably faster at all access sizes.

I know but you literally just violated what I just said, *slapping ensues*

 

A banana is not an orange. But a 10cm orange is larger than a 5cm orange thus takes longer to eat.

[Curufinwe_wins]

5 minutes ago, leadeater said:

I know but you literally just violated what I just said, *slapping ensues*

 

A banana is not an orange. But a 10cm orange is larger than a 5cm orange thus takes longer to eat.

Yes but we dont know if Zen3 is an banana or an orange yet, so it's premature to say that it takes longer to eat than Zen2... which was the premise in context, I had problems with saying "always".

 

Hahahahaha. I mean we don't actually disagree on substance

[leadeater]

9 minutes ago, Curufinwe_wins said:

Yes but we dont know if Zen3 is an banana or an orange yet, so it's premature to say that it takes longer to eat than Zen2... which was the premise in context, I had problems with saying "always".

 

Hahahahaha. I mean we don't actually disagree on substance

Caches don't change that much though, it won't in Zen3 and the fab node change doesn't actually mean the caches will be actually fabbed smaller. There is a reason the rule larger is slower exists and it's never not been true.

 

Just look at Broadwell-EP vs Skylake-SP, L2 cache is slower in Sylake-SP despite both fab optimizations and architecture changes.

 

I'll say it again, larger is always slower and it's never not been true.

 

Edit:

Quote

The massive L3 caches do have some disadvantages: latency goes up. The L3 cache of the Xeon E5-2699 v3 (45MB) has a latency between 20 and 32 ns while the 20MB cache of the Xeon E5-2690 hovers between 15 and 20 ns.

https://www.anandtech.com/show/8423/intel-xeon-e5-version-3-up-to-18-haswell-ep-cores-/11

 

22nm vs 14nm.

[LAwLz]

4 hours ago, mr moose said:

I don't even understand why people watch his videos let alone take him seriously. 

Because they want the stuff he says to be true.

I've hated AdoredTV ever since he got started. He always spews a bunch of bullshit. He always cherry picks benchmarks to get whichever result he wants.

I am 90% sure that at least some of his "rumor" videos have been completely made up just so that he will get a bunch of clicks.

 

 

16 hours ago, 5x5 said:

+10-15% IPC over Zen 2 (yes, over the current IPC leading CPUs)

Pretty sure Sunny Cove is the IPC leader right now, and that's without taking into consideration AVX-512 support.

[Bombastinator]

34 minutes ago, Curufinwe_wins said:

Yes but we dont know if Zen3 is an banana or an orange yet, so it's premature to say that it takes longer to eat than Zen2... which was the premise in context, I had problems with saying "always".

 

Hahahahaha. I mean we don't actually disagree on substance

I think i can comfortably predict that a zen3 cpu will be neither a banana nor an orange.  Neither of those fit well in an am4 socket.

[mr moose]

Just now, Bombastinator said:

I think i can comfortably predict that a zen3 cpu will be neither a banana nor an orange.  Neither of those fit well in an am4 socket.

Have you tried? bananas are quite squishy,  I predict they will fit much easier.

[RejZoR]

18 minutes ago, LAwLz said:

Because they want the stuff he says to be true.

I've hated AdoredTV ever since he got started. He always spews a bunch of bullshit. He always cherry picks benchmarks to get whichever result he wants.

I am 90% sure that at least some of his "rumor" videos have been completely made up just so that he will get a bunch of clicks.

 

 

Pretty sure Sunny Cove is the IPC leader right now, and that's without taking into consideration AVX-512 support.

Intel's only saving grace is high clocks that make up for lack of IPC...

[mr moose]

1 minute ago, RejZoR said:

Intel's only saving grace is high clocks that make up for lack of IPC...

So a metric that is independent of total clock speed only matters due to clock speed?

 

IPC is instruction per clock,  so it doesn't matter if the clock is 3Ghz or 33Ghz the frequency of the clock does not change IPC.  What changes IPC is how well the the HT/SMT is working, how well the code is refined and how many cores nd combinations of the three.   Clocks itself do not change IPC.

[Bombastinator]

8 minutes ago, mr moose said:

Have you tried? bananas are quite squishy,  I predict they will fit much easier.

There could be ram impingement problems though.  There are very small bananas of course but they tend to be narrow.

[williamcll]

I hope this gets good single thread speeds to beat intel.

[LAwLz]

17 minutes ago, RejZoR said:

Intel's only saving grace is high clocks that make up for lack of IPC...

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.

[leadeater]

3 minutes 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.

The way I read is Zen3 isn't getting AVX-512 too, reads to me as coming in Zen4.

[5x5]

21 minutes 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.

Those are all marketing slides, though. We have no idea if the comparison was done with a Pentium G4400 and single channel memory. We really need benchmarks of both when they release but 10% IPC gain sound a lot more believable than 100%

[GOTSpectrum]

AMD need to learn to sandbag leaks on purpose. If they know they will get around 12% IPC gain, then 'Leak' to the press that it will be 5-10%, that way they are overachieving. 

 

And remember folks, learn from scotty

 

 

 

[LAwLz]

4 minutes ago, 5x5 said:

Those are all marketing slides, though. We have no idea if the comparison was done with a Pentium G4400 and single channel memory. We really need benchmarks of both when they release but 10% IPC gain sound a lot more believable than 100%

I wouldn't be surprised if Sunny Cove has a ~15% IPC gain over Skylake in general. I mean, it's something like 5 years worth of improvements we're talking about here.

 

Please note that the 100% IPC gain is specifically for AVX-512, and that's not really unbelievably high. But you have to remember that it's for one specific instruction, so overall performance won't be as big of an improvement.