Intel claims ‘superior gaming performance’ over AMD, but uses better GPU for comparison

[ianh]

You don't get a 15% IPC increase from Zen 1 to Zen 2 only by shrinking the die, the designs are fairly radically different.

Zen 2 is the same overall chiplet design with the same multi CCX design with the cores being "tweaked" versions of the Zen 1 cores with no "radical" redesign, as you say.

Shrinking the process allowed them to tweak the cores slightly (as I also said) but in no way are these tweaks a "radical" change between Zen 1 and Zen 2, it's just slight changes. If they hadn't shrunk the process to 7nm and instead stayed at 14nm they wouldn't have seen anywhere near a 15% IPC improvement so it's the combination of process shrink AND design tweaks which have made the impressive IPC change manageable.

Zen 3 will be the first "redesign" of the original Zen 1 as it moves to mono-CCX and we'll go through the same thing again with Zen 4 which will take the Zen 3 redesign and shrink the process once more with a few tweaks but no redesigns.

Of course we may just have different views on what a "radical" redesign means. For me it means things like moving from single Die to Chiplet or moving from multi-CCX to mono-CCX or moving to a completely new process like Intel changing from Nehalem to Sandybridge etc, that's what I consider to be "radically different" rather than slight cache tweaks, increased ITA, raising the number of entries in the Physical Entry File from 168 to 180 etc (plus a whole list of other similar tweaks) - all of which add up to IPC increases but aren't what I would consider to be "radical" changes.

 

[humbug]

Zen 2 is the same overall chiplet design with the same multi CCX design with the cores being "tweaked" versions of the Zen 1 cores with no "radical" redesign, as you say.

Shrinking the process allowed them to tweak the cores slightly (as I also said) but in no way are these tweaks a "radical" change between Zen 1 and Zen 2, it's just slight changes. If they hadn't shrunk the process to 7nm and instead stayed at 14nm they wouldn't have seen anywhere near a 15% IPC improvement so it's the combination of process shrink AND design tweaks which have made the impressive IPC change manageable.

Zen 3 will be the first "redesign" of the original Zen 1 as it moves to mono-CCX and we'll go through the same thing again with Zen 4 which will take the Zen 3 redesign and shrink the process once more with a few tweaks but no redesigns.

Of course we may just have different views on what a "radical" redesign means. For me it means things like moving from single Die to Chiplet or moving from multi-CCX to mono-CCX or moving to a completely new process like Intel changing from Nehalem to Sandybridge etc, that's what I consider to be "radically different" rather than slight cache tweaks, increased ITA, raising the number of entries in the Physical Entry File from 168 to 180 etc (plus a whole list of other similar tweaks) - all of which add up to IPC increases but aren't what I would consider to be "radical" changes.

3100 vs 3300X

Both are 4 core 8 thread CPU's with 16MB L3

The 3100 is the same multi CCX as all other Zen 2 chips, the 3300X has its all its cores on a single CCX.

The IPC difference in games, or Rainbow Six at least is 8% higher to the 3300X, in F1 2019 4.4Ghz vs 4.4Ghz its a whopping 20%.

These are two slides from one end of the spectrum to the other, so just by removing the CCX and turning these CPU's something resembling in this case a 7700K the IPC in games is anything from 8 to 20% higher. And if you look at the GN slide it takes a 5.1Ghz 7700K to beat a 4.4Ghz 3300X by 3%, that's a clock speed difference of 16%.

AMD don't have to shrink the node, all they have to do is make them 8 core CCX's and in games they will have 13% higher IPC than a 9900K, about the same level of IPC difference they have in productivity workloads, unsurprisingly enough....

Oh BTW

 

[SupraWez]

Interesting, this confirms what I was thinking in the 3600,3800,3900XT thread regarding disabling cores so only a single CCX is in operation and benchmarking

 

[ianh]

AMD don't have to shrink the node, all they have to do is make them 8 core CCX's and in games they will have 13% higher IPC than a 9900K, about the same level of IPC difference they have in productivity workloads, unsurprisingly enough

Yeap, as I said, as they've redesigned the old Zen 1 based multi-CCX design into the new Zen 3 (which is still 7nm same as Zen 2) by using a mono CCX, they'll get a big boost from that redesign, leaving the next shrink for Zen 4.

Zen 3 will be the first "redesign" of the original Zen 1 as it moves to mono-CCX and we'll go through the same thing again with Zen 4 which will take the Zen 3 redesign and shrink the process once more with a few tweaks but no redesigns.

I'm quite looking forwards to Zen 3/4 myself, but I wonder what happens to whatever "Zen 5" is, because the current Zen chiplet/CCX architecture will need a new "redesign" by that point, rather than just another process shrink to say 3nm etc, and I don't know how much is left in the basic Zen design to change so we may see a completely new chiplet/CCX design come out rather than Zen tweaks.

 

[humbug]

Yeap, as I said, as they've redesigned the old Zen 1 based multi-CCX design into the new Zen 3 (which is still 7nm same as Zen 2) by using a mono CCX, they'll get a big boost from that redesign, leaving the next shrink for Zen 4.



I'm quite looking forwards to Zen 3/4 myself, but I wonder what happens to whatever "Zen 5" is, because the current Zen chiplet/CCX architecture will need a new "redesign" by that point, rather than just another process shrink to say 3nm etc, and I don't know how much is left in the basic Zen design to change so we may see a completely new chiplet/CCX design come out rather than Zen tweaks.

Yeah i think you're right, i think Zen 3 will reach the limits of the architecture.

Interesting, this confirms what I was thinking in the 3600,3800,3900XT thread regarding disabling cores so only a single CCX is in operation and benchmarking

Oh, yes.... i forgot, you can do that....

 

[kitfit1]

Interesting, this confirms what I was thinking in the 3600,3800,3900XT thread regarding disabling cores so only a single CCX is in operation and benchmarking

For what it's worth, i just ran Time Spy by disabling one CCX on my 3950x. As you can see, the Graphics score is about the same but the CPU score is way lower resulting in a much lower score overall.

https://www.3dmark.com/spy/12898911
https://www.3dmark.com/spy/12180487