5Ghz Amd & Intel (speculation topic)

[jigger]

Intels ring bus and current CPU designs have saturated each other. Intel are now at the limit of performance for the config.

Like I said a 5Ghz Ryzen would murder Intel's current designs. That should change in 2021 when I think Intel will have a more Ryzen-esk design.

 

[gavinh87]

Intels ring bus and current CPU designs have saturated each other. Intel are now at the limit of performance for the config.

Like I said a 5Ghz Ryzen would murder Intel's current designs. That should change in 2021 when I think Intel will have a more Ryzen-esk design.

Thats not strictly true though is it, the 8400 without HT is competing with the 1600 and in fact beating the 1800x in tasks like x264 due to the gimped avx support on ryzen.
In gaming the 8400 can often hold its own against the 1800x despite having 2 cores less with no HT.

 

[jigger]

Thats not strictly true though is it, the 8400 without HT is competing with the 1600 and in fact beating the 1800x in tasks like x264 due to the gimped avx support on ryzen.
In gaming the 8400 can often hold its own against the 1800x despite having 2 cores less with no HT.

Not in all situations, but we have to look at what is the bottleneck and how that can be worked around in software. Saying ah but look if you use this specific path to work around this bottle and limit it's impact isn't a fix for the problem because in an ideal scenario would wouldn't have to "gimp" the software's with such a fix.

Like I said this isn't the best place to discuss the ins and outs, but when you're at the point of workarounds to lessen performance impacts in hardware something has gone wrong.

With my system in Sig I can see the limitations of a ring bus with 4 high performance cores under load. Adding more high performance cores only makes that worse.

 

[AndreiD]

Jigger keeps talking about Intel's ring bus being """saturated""" despite them using it in XCC Broadwell Xeons, with up to 12 cores per ring.
Ring is fine and they'll keep using it for their future consumer CPUs, they switched to Mesh for Xeon parts because they actually needed to scale higher than 24 cores in that market (they used dual rings for their higher core count Broadwell Xeons).
For example, the 10 core 6950X used ring and performance was quite good, in a lot of consumer scenarios it's better than the 10 core 7900X (gaming for example).

And 5Ghz Ryzen ideally would be very competitive with 5Ghz Intels, but there's still an IPC deficit of about 5% (ideally). Another issue with Ryzen is that the CCX&Infinity Fabric design might be good for yields, but it also affects performance in scenarios where you have lots of cache thrashing (like... games!), reason why at the same clock Ryzen is somewhere between Ivy Bridge and Haswell in gaming.
If AMD make a true 8 Zen core die that can reach 5Ghz, with beefed up FMAs, it will actually pose a lot greater threat to Intel, but that's fantasy at this point.

 

[jigger]

Jigger keeps talking about Intel's ring bus being """saturated""" despite them using it in XCC Broadwell Xeons, with up to 12 cores per ring.
Ring is fine and they'll keep using it for their future consumer CPUs, they switched to Mesh for Xeon parts because they actually needed to scale higher than 24 cores in that market (they used dual rings for their higher core count Broadwell Xeons).
For example, the 10 core 6950X used ring and performance was quite good, in a lot of consumer scenarios it's better than the 10 core 7900X (gaming for example).

And 5Ghz Ryzen ideally would be very competitive with 5Ghz Intels, but there's still an IPC deficit of about 5% (ideally). Another issue with Ryzen is that the CCX&Infinity Fabric design might be good for yields, but it also affects performance in scenarios where you have lots of cache thrashing (like... games!), reason why at the same clock Ryzen is somewhere between Ivy Bridge and Haswell in gaming.
If AMD make a true 8 Zen core die that can reach 5Ghz, with beefed up FMAs, it will actually pose a lot greater threat to Intel, but that's fantasy at this point.

Wow you really don't get it. A true 8 core would need to communicate even more with fine grain tasks not to mention imit scaling. It would cost a fortune to bring a chip like that to market and developing it would be a nightmare. We've gone down the fine grain road. It's too late to turn back and focus on high performance multi tasking.

Intel's ring bus is not fine. It hasn't been for some time but Intel have to stick with it until 2021.

 

[AndreiD]

Can you provide any more detail on why Intel's ring bus isn't fine? Preferably with any performance benchmark or review to back it up, any would do.

The main issue with Ring is scaling, but for low-medium core parts (<12c) it's fine. Or are you going to disregard Broadwell-E and 6 core Coffee Lake performance? We have 6 core Coffee Lakes trade blows with 8 core Ryzens in multithreaded workloads while having massively better single thread performance (>25% clock + ~5% IPC), yet "ring bus is not fine"
If anything it's AMD's infinity fabric that's the issue here, and they'll need to tackle the latency issue with future iterations otherwise they're just going to be leapfrogged again by Intel.

 

[humbug]

Of course it's going to be faster than it's current form, it has a higher clockspeed.
Both cpus at the same clockspeed and intel would still win due to the ipc. The stilt has done Many tests and had shown that kaby is 10% ahead on average using a multitude of tests. Add the clockspeed advantage and your looking at 15-20% faster in some scenarios.
For some that's a gpu upgrade.
How you come to the conclusion that ryzen murders intel at the same clockspeed has me intrigued. Feel free to pm explaining what you can not go into here, you're almost making me regret getting rid of the 1700!

Everytime you write how much faster Intel are you add another 5% to it, its gone from 5% to 10% to 15% to 20%.... its a few %, on what is AMD's worst benchmark, Cinebench.

Yes Intel are a few % ahead in single core IPC, but Ryzen's SMT is about 10% better than Intel's so in MT the best Intel can hope for is equal performance, thats assuming Ryzen+ has 0 IPC increase.

 

[jigger]

Can you provide any more detail on why Intel's ring bus isn't fine? Preferably with any performance benchmark or review to back it up, any would do.

The main issue with Ring is scaling, but for low-medium core parts (<12c) it's fine. Or are you going to disregard Broadwell-E and 6 core Coffee Lake performance? We have 6 core Coffee Lakes trade blows with 8 core Ryzens in multithreaded workloads while having massively better single thread performance (>25% clock + ~5% IPC), yet "ring bus is not fine"
If anything it's AMD's infinity fabric that's the issue here, and they'll need to tackle the latency issue with future iterations otherwise they're just going to be leapfrogged again by Intel.

Not sure it's worth it because you have to look at the situation objectively and that's something you struggle with to be fair. Plenty of white papers have been produced that go into fine detail. Go fill your boots if you're genuinely interested and not just looking to argue how great/Crap Intel are compared to AMD.

 

[jigger]

Everytime you write how much faster Intel are you add another 5% to it, its gone from 5% to 10% to 15% to 20%.... its a few %, on what is AMD's worst benchmark, Cinebench.

Yes Intel are a few % ahead in single core IPC, but Ryzen's SMT is about 10% better than Intel's so in MT the best Intel can hope for is equal performance, thats assuming Ryzen+ has 0 IPC increase.

AMD's 10% advantage would increase with clock speed and scale over cores. I'd think the combined gains would improve performance pretty dramatically just from the SMT advantage.

 

[AndreiD]

on what is AMD's worst benchmark, Cinebench.

?? Cinebench is actually Ryzen's best case scenario, why do you think AMD keeps using Cinebench for all of their marketing, because it's their "worst benchmark"?
Cinebench is ideal for Ryzen, barely any cross CCX communication so the Infinity Fabric latency doesn't have any impact and it also doesn't use FMA/AVX, only SSE.

 

[gavinh87]

Everytime you write how much faster Intel are you add another 5% to it, its gone from 5% to 10% to 15% to 20%.... its a few %, on what is AMD's worst benchmark, Cinebench.

Yes Intel are a few % ahead in single core IPC, but Ryzen's SMT is about 10% better than Intel's so in MT the best Intel can hope for is equal performance, thats assuming Ryzen+ has 0 IPC increase.

They have a 5-10% IPC increase based on the stilts findings, add the clock speed advantage into that and you are looking at 20% in certain tests.
Intels 6 core keeping up and beating amd's 8 core speaks enough for itself.

 

[jigger]

?? Cinebench is actually Ryzen's best case scenario, why do you think AMD keeps using Cinebench for all of their marketing, because it's their "worst benchmark"?
Cinebench is ideal for Ryzen, barely any cross CCX communication so the Infinity Fabric latency doesn't have any impact and it also doesn't use FMA/AVX, only SSE.

Could you prove that Cinebench is AMD's best case scenario?

 

[gavinh87]

Could you prove that Cinebench is AMD's best case scenario?

Well given how the my 1700 beats my 8700k in cinebench yet gets mullered in everything else I've tried says enough.

 

[jigger]

They have a 5-10% IPC increase based on the stilts findings, add the clock speed advantage into that and you are looking at 20% in certain tests.
Intels 6 core keeping up and beating amd's 8 core speaks enough for itself.

OK, so scale the single core performance with clock speed. Seem to be around 10 points then add improved SMT performance let's say 15-20% for the hell of it and then add another 5-10% for IF and apply that to the nunber of cores. Thats without taking into account changes made to the design to best leverage the new higher clock speed.

 

[humbug]

?? Cinebench is actually Ryzen's best case scenario, why do you think AMD keeps using Cinebench for all of their marketing, because it's their "worst benchmark"?
Cinebench is ideal for Ryzen, barely any cross CCX communication so the Infinity Fabric latency doesn't have any impact and it also doesn't use FMA/AVX, only SSE.

Because its Intel's Favorite benchmarking tool, AMD have, up until Ryzen at least and since Phenom always had lower Floating Point performance, Cinebench is Floating Point which is why its a favourite with Intel, because its a favourite with Intel and AMD are much stronger in it with Ryzen than they have been in the past AMD now use it so that no one can say "AMD are using benchmarks that favour them" which would traditionally be things like Handbreak and 7zip de-compression, those are integer which is where AMD have always been strong in relation to Intel, it is still their strongest aspect, they just ain't weak at all in Floating Point anymore.

 

[jigger]

Well given how the my 1700 beats my 8700k in cinebench yet gets mullered in everything else I've tried says enough.

It's an assumption though. Can it be proved that software takes full advantage of the hardware? So as I said it's hard to measure. In fact it's impossible without a deep understand of the code and hardware.

 

[gavinh87]

https://forums.anandtech.com/threads/ryzen-strictly-technical.2500572/

These are the results of kaby lake 3.5ghz vs zen 3.5ghz.
These tests were;
3DPM
7ZIP
BLACKHOLES
BLENDER
LIBULLET
CINEBENCH 10 ST
CINEBENCH 11.5 ST
CINEBENCH 15 ST
EULER 3D
C-RAY ST
EMBREE
GCC
HIMENO
GMPBENCH
LINPACK
MCRT
NAMD
NBODY
OPENSSL
NQUEEN
STOCKFISH
VAMPIRE NUMBERS
WINRAR
X264
X265

Has there been a more comprehensive test that this?

 

[jigger]

Maybe but that is relative to the FX chip. Bar charts don't apply. We have to look at the data not how someone presented it.

Like I said it's pointless to discuss this hyperteical point then argue the ins and outs past a 5Ghz Ryzen> Intel's current designs.

 

[gavinh87]

Maybe but that is relative to the FX chip.

It could be relative to a 1990s celeron but it still shows that kabylake has a better IPC than zen.

 

[humbug]

https://forums.anandtech.com/threads/ryzen-strictly-technical.2500572/

These are the results of kaby lake 3.5ghz vs zen 3.5ghz.
These tests were;
3DPM
7ZIP
BLACKHOLES
BLENDER
LIBULLET
CINEBENCH 10 ST
CINEBENCH 11.5 ST
CINEBENCH 15 ST
EULER 3D
C-RAY ST
EMBREE
GCC
HIMENO
GMPBENCH
LINPACK
MCRT
NAMD
NBODY
OPENSSL
NQUEEN
STOCKFISH
VAMPIRE NUMBERS
WINRAR
X264
X265

Has there been a more comprehensive test that this?

Couple of things with that, it was publish March 2, so very much pre BIOS Fixes which gave a small but none the less performance hike, so add that to it, with those early BIOS the RAM way also only capable of running at 2400Mhz which was a huge performance deficit.

I think for what is that, 15% on a Beta BIOS as bad as the one Ryzen launched with and Ram speed that was at least 10 or 15% down on performance from 3000Mhz or 3200Mhz thats very good performance with Intel today, at least on par.

Edit, why do people benchmark compression instead of de-compression, do they not know compression its IO bound? its a measure of your hardware drive speed not the CPU.