WYP
News Guru
Higher IPC, increased clock speeds and a new process node.
Read more on the future of AMD's Ryzen architecture.
Read more on the future of AMD's Ryzen architecture.
AMD discusses their Zen 2 architecture
- Thread starter WYP
- Start date
Great to hear they're not stopping here. Even a 20% overall single thread performance improvement would be way more than Intel can offer, but the numbers from GlobalFoundries sound really impressive!
It's also great to know that as an early AM4 adopter, I likely won't be left out after 2 generations and will almost definitely have access to Zen 3 and maybe even further without a motherboard upgrade.
It's also great to know that as an early AM4 adopter, I likely won't be left out after 2 generations and will almost definitely have access to Zen 3 and maybe even further without a motherboard upgrade.
WYP
News Guru
The thing about Foundry notes is that the numbers do not automatically apply to every product. A node jump will offer a big leap, but it remains to be seen how much 7nm will help things.
Yeah, it is great that AM4 will have some upgrade headroom. Unlike AM3+, AM4 will likely keep this promise, I remember using my Phenom II 1090T on my AM3+ board and waiting for Steamroller on desktop (It never releases on desktop...).
AngryGoldfish
Old N Gold
It's good to see the future of Zen still being hyped. It should help keep a few from investing in Coffee Lake if they know that Ryzen has bigger plans to come.
Excalabur50
Well-known member
I just got my refund from Intel and was thinking of a 1700x I wonder if I should just keep using my laptop and wait for Zen2 instead
AngryGoldfish
Old N Gold
Or do what I might do: Buy a 1600X now and then Ryzen 2 8c when it comes out.
Excalabur50
Well-known member
Truth be told I don't know that I can wait I'm not a patient person, I'm already struggling waiting for the motherboard that I want to release mid September and I'm already chomping at the bit LOL
WYP
News Guru
I am now remembering the leaked roadmaps that pointed towards a future EPYC CPU using Zen 2 with 48 cores. Makes me wonder if core counts will increase per die or if AMD will create Zen 2 dies with more cores in each die.
If the 48-cores rumour is true, AMD could use the same 4-die threadripper design and have 12-cores each or a new 6-die design with 8 cores each like Ryzen. Interesting prospect.
Will AMD push to have even more cores per die with Ryzen 2 or the same core count on a smaller die? Some food for thought for Ryzen 2. 12-core per die would allow 12-core CPUs on AM4!
It is far too early to know what AMD is actually doing, so this is all just me talking to the wall. Even so, a 4-die solution seems best given AMD's memory controller configuration, as a 6x 8-core die solution would leave two dies with no memory interconnects if AMD uses the same socket and 8-channel memory setup.
Again, I am just talking hypotheticals here. AMD has not released any real info on Zen 2 yet.
Thelosouvlakia
Classy-Fried-Chicken
Yeah if AMD has a socket and makes a promice that says "hey guys, we will be on AM4 for the next 2 generations to come, you can have the latest cpu even on an x370 board on a b350 board, you'll be fine but some features will be unavailable because of the chipset" That would mean good old AMD when you could have an am3 cpu on an am2 motherboard and am3+ cpus on am3 motherboards and you could have a phenom II on a crosshair IV and then upgrade to an fx8150 without buying a new board
AngryGoldfish
Old N Gold
Is AM4 set up for 12 cores though? Will current motherboards be able to handle them? I think 8 cores is enough for AM4 personally. They just need to get the prices down, clocks up (and IPC), and stabilise the platform fully. An 8 core CPU that clocks to 4.4Ghz with a 65W TDP, IPC on par with Skylake/Kaby Lake, all for €300. That would be drool worthy.
WYP
News Guru
It really depends on how AMD wants to design their future CPUs, it is anyone's guess what the company will do.
As for 12 cores on AM4, it is possible as 7nm should save the die space, then the question is power, which is partially fixed by 7nm but would likely need some power saving architectural changes to work.
I personally would prefer to have 8 faster cores than 12 slower cores, it would also make Threadripper have an even more limited use case. Again it depends on how AMD designs their next generation, as the same dies should be used for AM4, TR4 and SP3 (EPYC). (How many cores will be in each Zen 2 die and will they use the same dies for each product stack?)
The big questions for Zen 2 are exact architectural changes are planned and how good (and how well utilised) 7nm is.
I don't think Zen2 refers to the 2nd generation of AM4 CPUs. The slide says Zen1 will launch on 14nm and 14nm+, and 7nm probably won't be mature enough for large area die till the end of 2018 at the earliest. If there's this much headroom for improvement I feel we'll see a Kaby Lake-esque "optimisation" step between Zen1 and Zen2 around mid-2018 on 14nm+ with bumped clocks, more overclocking headroom, a few small features/ISA updates/bug fixes, and finer-tuned power curves. The improvements in software support for Ryzen arch combined with these changes should show a reasonable uplift in performance even if IPC only changes by a few percent.
Whether or not these will be labelled 1X50/1X05 CPUs or 2X00 CPUs basically comes down to the whim of the marketing department. Going from AMDs APUs naming schemes I'd assume the former, but Intel has latched onto the latter and that's who they're competing with at the end of the day.
On 12 core/3CCX Zen2 CPUs, the 40% "performance improvement" GF talks of is more or less bang on for the efficiency improvement required to pull it off(You need a 50% improvement to add 50% more cores, but a portion of the die and its power consumption is not taken up by the cores, so 40% should be doable). Combined with the die area improvements this should be doable while theoretically reducing overall die size(Though again, not all of the die is made up of CPU cores, and conversely not all parts of the die will scale as well as this). It's also worth noting a 3CCX die would have significantly better yields on lower end parts, allowing AMD to push more cores to lower price points/lower price points for the same core count(As Intel will also be looking to do). As well as this of course it is a necessity for AMD to push higher core counts to their TR4/server sockets if they want to continue using the same die across their whole product line(Which has proved very valuable to them), and they've already revealed they plan to continue to push higher cores on servers. This is the only way they can do this without trying to cram 6 dies into a TR4 package.
We now have <£200 gaming processors(also AMDs most popular processors) with 12 threads, and I'm sure AMD is pushing devs to make use of these. I think by the time Zen2 launches there will be at least some software on the market that will see an uplift from 12 full cores.
Whether or not these will be labelled 1X50/1X05 CPUs or 2X00 CPUs basically comes down to the whim of the marketing department. Going from AMDs APUs naming schemes I'd assume the former, but Intel has latched onto the latter and that's who they're competing with at the end of the day.
On 12 core/3CCX Zen2 CPUs, the 40% "performance improvement" GF talks of is more or less bang on for the efficiency improvement required to pull it off(You need a 50% improvement to add 50% more cores, but a portion of the die and its power consumption is not taken up by the cores, so 40% should be doable). Combined with the die area improvements this should be doable while theoretically reducing overall die size(Though again, not all of the die is made up of CPU cores, and conversely not all parts of the die will scale as well as this). It's also worth noting a 3CCX die would have significantly better yields on lower end parts, allowing AMD to push more cores to lower price points/lower price points for the same core count(As Intel will also be looking to do). As well as this of course it is a necessity for AMD to push higher core counts to their TR4/server sockets if they want to continue using the same die across their whole product line(Which has proved very valuable to them), and they've already revealed they plan to continue to push higher cores on servers. This is the only way they can do this without trying to cram 6 dies into a TR4 package.
We now have <£200 gaming processors(also AMDs most popular processors) with 12 threads, and I'm sure AMD is pushing devs to make use of these. I think by the time Zen2 launches there will be at least some software on the market that will see an uplift from 12 full cores.
Last edited:
ImprovizoR
Well-known member
I think that I made the right choice waiting for Zen 2. My i5-4590 still does the job, but I'm gonna need more threads very soon.
Last edited:
NeverBackDown
AMD Enthusiast
Honestly if they could just make it so Zen can clock to close to 5ghz, many of there problems would go away. That increase in clock speed would seriously offset most of there current issues. Also makes Intel's advantage completely moot.
Of course, architectural improvements are definitely needed to offset that increase in power consumption.
Dawelio
Active member
I've never really gotten what the difference is between clocks and IPC? I've also thought that they were the same thing? :huh:
AlienALX
Well-known member
IPC = instructions per clock. Or, how many instructions a CPU can perform at a certain clock speed. IPC is not clock speed, but it does increase with clock speed, if that makes any sense.
g0ggles1994
Active member
This, but to build on it a bit. Where IPC focus' on how many instructions the CPU can perform per clock. The clock speed is how many clocks the CPU can run at per second.
So to simplify it, if a CPU has an IPC of 10 and a clock speed of 4GHz (Which basically means 4 billion clocks per second). That CPU will be able to perform 40 billion instructions per second. So whenever the IPC is increased, the new CPU can perform more instructions per second when running at the same speed as the old CPU.
Similar threads
