I'm really excited about these. As always, rumors tend to get a bit TOO crazy with expectations, but right now those rumors are saying this is biggest IPC jump in Zen history. That, along with the Strix Point rumor pointing towards a 16 CU RDNA-3.5 iGPU, another big jump over the current 12 CU RDNA-3 one, has got me super hyped for AMD's launches this year.
AMD's slide says a mixture of 4-nm/3-nm for the Zen 5 chips. I'm guessing Granite Ridge and Turin will be 4-nm (an upgrade from current 5-nm), while Strix Point will be monolithic 3-nm (an upgrade from current 4-nm).
I've been waiting for Zen 5 to upgrade my 10.5 year old 4770k system. Thought about Zen 3 but didn't want to be on the tail end of DDR4 like I was with DDR3. Zen 4 was a bit too expensive early on and by the time prices dropped enough it isn't long to Zen 5.
Yes 7nm vs 5nm is .71 smaller, or 1.4 times as much stuff. While 5nm vs 3nm is .6 smaller (and .43 smaller than 7nm, for 2.3X), or 1.6 times as much stuff.
But, when we get to 1 or 2nm, along with the necessary architecture changes (to use so small a node), that's when we'll have not just stacked cache but stacked CPUs (and NPUs) that will be so much faster (combined with the certainty of more than 6096 pins, for PCIe 6 and CXL 3.x).
Worth waiting for PCIe 6, waiting for Zen 7 is a few years away; but the huge cache and NPU will be interesting. 🤔
I hope both of you realize that the node names are marketing names and there's no point doing calculations based on the node number anymore. You have to look at published information about density. Recent nodes have mediocre scaling and SRAM isn't scaling at all at TSMC (for the moment).
TSMC's optimization of N3E, N3P, should have about 1.66x the logic density of N5, with no improvement (1x) for SRAM.
You should keep Zen 4 on the table after Zen 5 comes out, if whatever you want (e.g. 7800X3D) becomes really cheap. There have already been some good bundle deals too (Newegg).
Don't get your hopes up with IPC uplift. The rumor I heard is that it's better than Zen4, but not a substantial gain. There were a lot of performance losses while fixing all the bugs transitioning to a new node.
Its not going on a new node, at least not the desktop discrete parts. This should be an uplift on the level of Intels Comet Lake to Alder Lake, Going to at least +50% wider front end decode. AMD is doing the same thing to the Zen 4 core that Intel did in producing the P cores in AlderLake. Based on that alone, IPC uplift should at least equal their previous best jump since Zen 1, which was Zen 2 to Zen 3's +19%.
Don't expect a huge IPC uplift anymore. 10-20% range is likely.
It's all about bandwidth when it comes to Strix Point. If AMD doesn't want to go above 128-bit for mainstream APUs, they need to add an Infinity Cache equivalent eventually. Strix Halo is a different story but it's expected in 2025 now.
Strix Point should also be 4nm. I think the only 3nm product will be Zen 5c chiplets for Epyc, not sure.
10-20% IPC uplift sounds pretty good to me - Zen4 is already really good. There's not much room for the old "huge uplift just because the previous generation was awful" any more.
It certainly isn't bad, but abnormally high estimates are probably wrong. Mobile will be more interesting than desktop for the Zen 5 generation. The best AMD could do to make desktop exciting is make an 8 Zen5 +16 Zen5c monster with 3D V-Cache.
Power efficiency and AVX-512 will be aspects to watch.
Going from a 4 wide to a 6 or 8 wide decode front end is not going to be a 10% uplift. Im going to be shocked if its less than 19%. They are doing exactly what Apple and Intel did to get their huge IPC gains for M1 and Alder Lake/Golden Cove-- going at least 50% wider in # of front end decodes per clock. Saying it wont be a very good IPC uplift makes no sense. Intel got ~30%+ and Apple currently has the highest IPC of any CPU available.
The split node numbering, eg: 4nm/3nm, has also previously indicated the CPU die vs IO die process nodes of the chiplet setup used in desktop and sever parts.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
23 Comments
Back to Article
ikjadoon - Friday, February 2, 2024 - link
Please do double-check the headline. It states 2025, while the text states 2024.Ryan Smith - Friday, February 2, 2024 - link
I don't know what you're talking about. It has always said "2024", and there is now no evidence to the contrary. =P(Seriously though, thank you. We caught this one right after the article went live, but not quite fast enough for your eagle eyes!)
NextGen_Gamer - Friday, February 2, 2024 - link
I'm really excited about these. As always, rumors tend to get a bit TOO crazy with expectations, but right now those rumors are saying this is biggest IPC jump in Zen history. That, along with the Strix Point rumor pointing towards a 16 CU RDNA-3.5 iGPU, another big jump over the current 12 CU RDNA-3 one, has got me super hyped for AMD's launches this year.AMD's slide says a mixture of 4-nm/3-nm for the Zen 5 chips. I'm guessing Granite Ridge and Turin will be 4-nm (an upgrade from current 5-nm), while Strix Point will be monolithic 3-nm (an upgrade from current 4-nm).
schujj07 - Friday, February 2, 2024 - link
I've been waiting for Zen 5 to upgrade my 10.5 year old 4770k system. Thought about Zen 3 but didn't want to be on the tail end of DDR4 like I was with DDR3. Zen 4 was a bit too expensive early on and by the time prices dropped enough it isn't long to Zen 5.CrystalCowboy - Friday, February 2, 2024 - link
I'm holding out for Zen 7.Rοb - Saturday, February 3, 2024 - link
Yes 7nm vs 5nm is .71 smaller, or 1.4 times as much stuff.While 5nm vs 3nm is .6 smaller (and .43 smaller than 7nm, for 2.3X), or 1.6 times as much stuff.
But, when we get to 1 or 2nm, along with the necessary architecture changes (to use so small a node), that's when we'll have not just stacked cache but stacked CPUs (and NPUs) that will be so much faster (combined with the certainty of more than 6096 pins, for PCIe 6 and CXL 3.x).
Worth waiting for PCIe 6, waiting for Zen 7 is a few years away; but the huge cache and NPU will be interesting. 🤔
Terry_Craig - Sunday, February 4, 2024 - link
On chips with a lot of cache, 3nm offers just over 20% compared to 4nm.CrystalCowboy - Monday, February 5, 2024 - link
"While 5nm vs 3nm is .6 smaller (and .43 smaller than 7nm, for 2.3X), or 1.6 times as much stuff."That's in one dimension. Semiconductors are two-dimensional. So 3nm vs 5nm should fit up to 2.77 times more stuff.
nandnandnand - Monday, February 5, 2024 - link
I hope both of you realize that the node names are marketing names and there's no point doing calculations based on the node number anymore. You have to look at published information about density. Recent nodes have mediocre scaling and SRAM isn't scaling at all at TSMC (for the moment).TSMC's optimization of N3E, N3P, should have about 1.66x the logic density of N5, with no improvement (1x) for SRAM.
nandnandnand - Saturday, February 3, 2024 - link
You should keep Zen 4 on the table after Zen 5 comes out, if whatever you want (e.g. 7800X3D) becomes really cheap. There have already been some good bundle deals too (Newegg).GeoffreyA - Friday, February 2, 2024 - link
It's going to be quite interesting to see the microarchitecture of Zen 5 and what they've done.meacupla - Friday, February 2, 2024 - link
Don't get your hopes up with IPC uplift.The rumor I heard is that it's better than Zen4, but not a substantial gain. There were a lot of performance losses while fixing all the bugs transitioning to a new node.
Santoval - Sunday, February 4, 2024 - link
I've heard anything from +10% to +30% higher IPC over Zen 4.Most likely Zen 5 will end up with an 15 - 18% IPC edge.
Josh128 - Thursday, February 8, 2024 - link
Its not going on a new node, at least not the desktop discrete parts. This should be an uplift on the level of Intels Comet Lake to Alder Lake, Going to at least +50% wider front end decode. AMD is doing the same thing to the Zen 4 core that Intel did in producing the P cores in AlderLake. Based on that alone, IPC uplift should at least equal their previous best jump since Zen 1, which was Zen 2 to Zen 3's +19%.nandnandnand - Saturday, February 3, 2024 - link
https://youtu.be/A0s0Usfg__g?t=1317Don't expect a huge IPC uplift anymore. 10-20% range is likely.
It's all about bandwidth when it comes to Strix Point. If AMD doesn't want to go above 128-bit for mainstream APUs, they need to add an Infinity Cache equivalent eventually. Strix Halo is a different story but it's expected in 2025 now.
Strix Point should also be 4nm. I think the only 3nm product will be Zen 5c chiplets for Epyc, not sure.
boozed - Sunday, February 4, 2024 - link
10-20% IPC uplift sounds pretty good to me - Zen4 is already really good. There's not much room for the old "huge uplift just because the previous generation was awful" any more.nandnandnand - Monday, February 5, 2024 - link
It certainly isn't bad, but abnormally high estimates are probably wrong. Mobile will be more interesting than desktop for the Zen 5 generation. The best AMD could do to make desktop exciting is make an 8 Zen5 +16 Zen5c monster with 3D V-Cache.Power efficiency and AVX-512 will be aspects to watch.
Josh128 - Thursday, February 8, 2024 - link
Going from a 4 wide to a 6 or 8 wide decode front end is not going to be a 10% uplift. Im going to be shocked if its less than 19%. They are doing exactly what Apple and Intel did to get their huge IPC gains for M1 and Alder Lake/Golden Cove-- going at least 50% wider in # of front end decodes per clock. Saying it wont be a very good IPC uplift makes no sense. Intel got ~30%+ and Apple currently has the highest IPC of any CPU available.evanh - Friday, February 2, 2024 - link
The split node numbering, eg: 4nm/3nm, has also previously indicated the CPU die vs IO die process nodes of the chiplet setup used in desktop and sever parts.Koenig168 - Saturday, February 3, 2024 - link
A little birdie said Granite Ridge will launch in April. Rumors, so take with a dollop of salt.Josh128 - Thursday, February 8, 2024 - link
Zero chance. August at the earliest.quaz0r - Saturday, February 3, 2024 - link
I can't wait til zen5 comes out so I can start waiting for zen6nandnandnand - Sunday, February 4, 2024 - link
If you can't wait, start waiting for Zen6 now.