Question Zen 6 Speculation Thread

[adroc_thurston]

N5X doesn't exist in TSMC's official roadmaps as far as I know, but a perf-oriented "N5HPC" was publicly mentioned years ago

yeah they rebranded them.
It was N7HPC and N5HPC until it wasn't.

 

[Doug S]

N5X doesn't exist in TSMC's official roadmaps as far as I know, but a perf-oriented "N5HPC" was publicly mentioned years ago, so I guess it's basically the same thing?

Sure it does, its just that it is called N4X.

 

[OneEng2]

Sure it does, its just that it is called N4X.

Interesting that everyone seems to now accept that an enhancement to an existing process node is a new process node.

 

[Markfw]

Interesting that everyone seems to now accept that an enhancement to an existing process node is a new process node.

Not me. Until TSMC or Intel or other chips mfg has a new node, I don't.

 

[adroc_thurston]

Interesting that everyone seems to now accept that an enhancement to an existing process node is a new process node.

Nodelets are a thing and they're whatever.
X-stuff isn't pure enhancement per se, they're the nuclear option.

 

[igor_kavinski]

X-stuff isn't pure enhancement per se, they're the nuclear option.

What do they lose? Are yields lower for the X nodes?

 

[branch_suggestion]

What do they lose? Are yields lower for the X nodes?

Leakage.

 

[adroc_thurston]

What do they lose?

leakage

 

[LightningZ71]

Remember, it's a triangle. You get efficiency on one point, performance on the second, and density on the third. You can tweak the node to push out any vertex at the expense of one or both of the other two. The X nodes just push the performance vertex at the expense of density AND efficiency. This is a vast oversimplification, but it gets the point across.

 

[Win2012R2]

Full rate AVX512 should work very nicely with so much more bandwidth afforded by GDDR7, would be shame if they cut it.

 

[Hulk]

Remember, it's a triangle. You get efficiency on one point, performance on the second, and density on the third. You can tweak the node to push out any vertex at the expense of one or both of the other two. The X nodes just push the performance vertex at the expense of density AND efficiency. This is a vast oversimplification, but it gets the point across.

So like a three-legged stool that you need to be level (balanced).

 

[maddie]

So like a three-legged stool that you need to be level (balanced).

No. It just means that all 3 can't be maximized simultaneously.

 

[poke01]

Interesting

 

[fastandfurious6]

View attachment 121296
Interesting

makes sense

 

[poke01]

I could definitely see the steam deck 2 making use of these cores

 

[Kaffeekenan]

Sooooo, if Zen 6 really uses N2X as a node...how much better is that compared to Zen 5 purely regarding the nodeadvantage? Anyone here who can answer that? If possible an answer with a rough percentage and not an answer like "waaaaaaay better my friend". Thanks guys. 🙂

 

[igor_kavinski]

Zen 6%

And if you like riding the hype train, Zen 60% (there's gotta be at least ONE multicore benchmark where it manages to do that).

 

[RTX]

Sooooo, if Zen 6 really uses N2X as a node...how much better is that compared to Zen 5 purely regarding the nodeadvantage? Anyone here who can answer that? If possible an answer with a rough percentage and not an answer like "waaaaaaay better my friend". Thanks guys. 🙂

It may use less power than N2P if N4P is anything to go by.

Page 381 - Discussion - Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)

Page 381 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

 

[fastandfurious6]

Sooooo, if Zen 6 really uses N2X as a node...how much better is that compared to Zen 5 purely regarding the nodeadvantage? 🙂

looking at TSMC announcements, can't find exact N4P vs N2X (zen5 and zen6 accordingly) but based on below two tables:






N5 -> N4P (z5)
-22% power
+11% perf
+6% density

N5 -> N3E
-34% power
+18% perf
+30% density

N3E -> N2P
-30% / -40% power
+15% / +20% perf
+15% density

* (for N2P -> N2X let's assume around +5%)

** (we get approximate N4P base by subtracting N5 -> N4P from N5 -> N3E percentages)

so based on above tables just the node improvement from zen5 to zen6 will offer up to:

-57% power
+32% perf
+44% density

 

[Kaffeekenan]

looking at TSMC announcements, can't find exact N4P vs N2X (zen5 and zen6 accordingly) but based on below two tables:

View attachment 121389

View attachment 121390


N5 -> N4P (z5)
-22% power
+11% perf
+6% density

N5 -> N3E
-34% power
+18% perf
+30% density

N3E -> N2P
-30% / -40% power
+15% / +20% perf
+15% density

* (for N2P -> N2X let's assume around +5%)

** (we get approximate N4P base by subtracting N5 -> N4P from N5 -> N3E percentages)

so based on above tables just the node improvement from zen5 to zen6 will offer up to:

-57% power
+32% perf
+44% density

Thank you, great answer! So Zen 6 could be an absolute banger, if they also manage a nice architecture-boost. Nice!!

 

[LightningZ71]

Those improvements should have "or" between them. If you push maximum density, you sacrifice performance and get higher leakage. If you push maximum performance, you have to relax density and suffer excess power draw. If you optimize for power, density and performance suffer.

Overall, N2X will offer considerably better performance while delivering improved performance per watt numbers while still allowing you to bloat the cores by a good 50% or more as compared to Zen5 on N3P. It's not magic, but it is a bigger jump than we've seen in a while.

 

[Meteor Late]

10% more IPC, 13-14% more max frequency, I expect something like that. So 6.5GHz 10950X with 25% higher single core performance seems reasonable.

 

[Thunder 57]

10% more IPC, 13-14% more max frequency, I expect something like that. So 6.5GHz 10950X with 25% higher single core performance seems reasonable.

Careful with that kind of thinking. You might be setting yourself up to be disappointed.

 

[fastandfurious6]

Those improvements should have "or" between them. If you push maximum density, you sacrifice performance and get higher leakage. If you push maximum performance, you have to relax density and suffer excess power draw. If you optimize for power, density and performance suffer.

Overall, N2X will offer considerably better performance while delivering improved performance per watt numbers while still allowing you to bloat the cores by a good 50% or more as compared to Zen5 on N3P. It's not magic, but it is a bigger jump than we've seen in a while.

my guess total node improvements will be around half compared to max possible:

-30% power
+20% perf
density allowing 12core CCDs just enough

 

[eek2121]

10% more IPC, 13-14% more max frequency, I expect something like that. So 6.5GHz 10950X with 25% higher single core performance seems reasonable.

11950X*