Rumor: Next gen AMD Epyc to get 64 cores?

[cj09beira]

1 minute ago, Liltrekkie said:

Is that 256 physical cores (with 512 threads?) or 256 threads (with 128 physical cores)?

probably threads as i dont think an Os can distinguish between them 

[Dylanc1500]

12 minutes ago, leadeater said:

Both, they limit sockets based on edition and they limit the number cores the OS will use whether they are virtual or physical.

 

Edit:

Actual cores btw not threads, meaning you get hit a bit harder when using VMs.

Ya that's what I had thought, it had been awhile since I had gone through all of the windows server 2016 white sheets. I luckily get the privilege of not having to deal with them all to often lol.

[ScratchCat]

27 minutes ago, leadeater said:

With current technology more cores = lower frequency and that is globally true, in reference to the same architecture when doing this comparison.

Not always. More cores on a single die results in a lower clockspeed, if you 'glue together' multiple dies you can keep the same clocks with more cores just like Threadripper (yes this partly has to do with the fact that they are using binned dies). MCM can work magic by making each die smaller therefore having a lower chance of having a misfeature which results in a lower maximum clock speed.

 

[cj09beira]

2 minutes ago, ScratchCat said:

Not always. More cores on a single die results in a lower clockspeed, if you 'glue together' multiple dies you can keep the same clocks with more cores just like Threadripper (yes this partly has to do with the fact that they are using binned dies). MCM can work magic by making each die smaller therefore having a lower chance of having a misfeature which results in a lower maximum clock speed.

 

thats true until you reach epyc levels of core counts, where the problem start to be cooling it and powering it so you cant keep them at high clocks

[Drak3]

12 minutes ago, cj09beira said:

probably threads as i dont think an Os can distinguish between them 

Windows can distinguish threads and cores. It's 256 cores, regardless of multithreading shenanigans.

[Taf the Ghost]

Been discussing this other places.

 

"Rome", which is the Zen 2-based Epycs, was already known to be a 48c/96t part (max). That was on AMD's roadmap as of early 2016. This isn't surprising, now, as 7nm from GloFo is something like 2 node jumps from 14nm. So AMD has a lot of room to play with.

 

However, it should be noted that Epyc uses both normal Zen packages, and a set of flipped packages. There is actually 3 full chip designs in the Ryzen/Epyc series. Raven Ridge, Zen and Zen-flipped. This is important.

 

What I think we're going to see is 12c (Ryzen), 24c (Threadripper) and 48c & 64c (Epyc). They might run out a set of 32c Threadrippers, but not sure on that. I think part of the "news" with the rumor is that AMD had taped out their "big core" design that they didn't run with for the first Zen. AMD is using a modular design system now, so they can practically drag & drop design bits now. I just do hope the "Zen 2" core has AVX2 instructions running better, that'll help AMD a lot in the server space.

[ScratchCat]

2 minutes ago, cj09beira said:

thats true until you reach epyc levels of core counts, where the problem start to be cooling it and powering it so you cant keep them at high clocks

The rumor mentioned around 240W, given that the 32 Core Epyc requires 180W I'll sit here with my mountain of salt until this is announced. This kind of power drop would require a large process change or a painfully large reduction in clocks.

[Dylanc1500]

3 minutes ago, Drak3 said:

Windows can distinguish threads and cores. It's 256 cores, regardless of multithreading shenanigans.

You can see this by using powershell.

[BingoFishy]

56 minutes ago, D Levy said:

Money is not the issue for HEDT, i9 7900x for example... can be OC'd to 4Ghz easily.

4Ghz is plenty for gaming, compound by the massive core count for games utilizing vulcan support properly...

You could game in 4K, record it, and stream on the same PC... probably even do a little rendering at the same time lol.

No but really... 4Ghz or higher with more cores than 7900x and more lanes... would be just... fing amazing!

who said 4ghz

[Taf the Ghost]

2 minutes ago, ScratchCat said:

The rumor mentioned around 240W, given that the 32 Core Epyc requires 180W I'll sit here with my mountain of salt until this is announced. This kind of power drop would require a large process change or a painfully large reduction in clocks.

2 node improvement going from 14nm to 7nm at GloFo. It's actually in their published design docs. Intel's 10nm node was so bad it's been delayed by 4? years now. GloFo's was so bad it was canceled.  They'd be able to fit 64c at 3 Ghz under that TDP.

[Falconevo]

This is going to take a while, have only just got the EPYC evaluation builds in for testing. 

 

Currently testing the dual socket version, 64cores, 128threads.

[cj09beira]

@Taf the Ghost what do you mean with flipped chips, as cpus are all already flipped, or do you mean it in another way?

[Sauron]

28 minutes ago, cj09beira said:

probably threads as i dont think an Os can distinguish between them 

It can distinguish between them (plus the cpu usually reports it), however it doesn't give them any special treatment - it's up to the cpu to handle the OS' requests efficiently. I don't know if core support includes them or not though.

[Taf the Ghost]

While it's still way too early (but AMD should be taping out two designs on 7nm about now), but I really do think the 64c rumor is about a second Zen-package design that's bigger than the normal Desktop one. It's a risk for AMD, but Intel won't have EIMB online until 2020 at the earliest. This is a chance for AMD to really push the Zen design to the max against what Intel does with a single CPU. It's a "go big" moment for AMD.

[Sauron]

Just now, Taf the Ghost said:

While it's still way too early (but AMD should be taping out two designs on 7nm about now), but I really do think the 64c rumor is about a second Zen-package design that's bigger than the normal Desktop one. It's a risk for AMD, but Intel won't have EIMB online until 2020 at the earliest. This is a chance for AMD to really push the Zen design to the max against what Intel does with a single CPU. It's a "go big" moment for AMD.

Epyc is already on a massive socket, I don't think package size is a big deal in this case. Surely any external problem that might arise from the die area could be fixed with specific motherboard and cooler designs, which are not a problem for the potential customers of something like this.

[Taf the Ghost]

1 minute ago, cj09beira said:

@Taf the Ghost what do you mean with flipped chips, as cpus are all already flipped, or do you mean it in another way?

The Epyc die shots show mirrored Zen packages. It puts the inter-connects between the packages in the middle, which would shorten up distance between connections and make everything fairly even. My assumption is that the normal Ryzen chip package can accept either direction.

[Taf the Ghost]

3 minutes ago, Sauron said:

Epyc is already on a massive socket, I don't think package size is a big deal in this case. Surely any external problem that might arise from the die area could be fixed with specific motherboard and cooler designs, which are not a problem for the potential customers of something like this.

I assume AMD would need a new Socket for a 64c CPU, but the entire CPU package doesn't have to be larger. I just mean the physical CPU would be a larger design. Say 280 mm² for Epyc 2 vs 180mm² for Ryzen 2. 

[leadeater]

Just now, Liltrekkie said:

Is that 256 physical cores (with 512 threads?) or 256 threads (with 128 physical cores)?

Physical cores, however when virtualizing Windows they are shown as virtual cores but the limit is the same. The down side to VMs is that when you configure vCPUs that's 1:1 under Windows to this limit even though the hypervisor isn't necessarily giving the VM that number of real physical cores under the hood.

 

 

[Taf the Ghost]

Here's a link to someone running the numbers.

 

https://forums.anandtech.com/threads/cpchardware：2nd-gen-amd-epyc-will-have-64-cores-256-mo-l3-8x-ddr4-3200-and-128-pcie-4-lines.2525152/page-3#post-39148715

 

7SoC 6T  is the process that AMD is likely using for both Mainstream & Server CPUs. Though, as I said, my assumption is we're actually getting a "server" CPU design from AMD, rather than just "one zen package to work in all of them". 

 

If I'm also doing wild speculation, let's think about the 3 spaces. The APUs are going to be staying 4c/8t more than likely. The smaller & more power efficient, the better there. Mainstream is the question. If AMD is splitting the Mainstream & Server, that means we could see 8c/16t as Zen 2, but clocking near 5 Ghz. But that could mean Threadrippers at 32c/64t. 

 

Or they really could go all-in on rearranging the CPU market and drop a Ryzen 7 2800X in Q2 2019 with 16c/32t at 4.4 all-core and 5 Ghz XFR.

[leadeater]

47 minutes ago, ScratchCat said:

Not always. More cores on a single die results in a lower clockspeed, if you 'glue together' multiple dies you can keep the same clocks with more cores just like Threadripper (yes this partly has to do with the fact that they are using binned dies). MCM can work magic by making each die smaller therefore having a lower chance of having a misfeature which results in a lower maximum clock speed.

 

MCM still doesn't change this reality. If you have a power envelope of say 200W and a CPU core needs 20W to sustain 4.5Ghz then 10 cores is the limit, MCM or not.

 

MCM helps with heat dissipation which does effect power/thermal runaway and all that but it still comes down to more cores = lower frequency.

 

This is why I said same architecture, you're comparing Intel to AMD which isn't directly comparable in the statement I was making.

[cj09beira]

6 minutes ago, Taf the Ghost said:

The Epyc die shots show mirrored Zen packages. It puts the inter-connects between the packages in the middle, which would shorten up distance between connections and make everything fairly even. My assumption is that the normal Ryzen chip package can accept either direction.

i see it now, cool small detail right there.

4 minutes ago, Taf the Ghost said:

I assume AMD would need a new Socket for a 64c CPU, but the entire CPU package doesn't have to be larger. I just mean the physical CPU would be a larger design. Say 280 mm² for Epyc 2 vs 180mm² for Ryzen 2. 

might not, there is room to grow in that socket

i myself think and will go to 12, 24,48 i am skeptical of the 64 core one unless its like a server only sku or the gen next to that

[leadeater]

36 minutes ago, Falconevo said:

This is going to take a while, have only just got the EPYC evaluation builds in for testing. 

 

Currently testing the dual socket version, 64cores, 128threads.

Remember to share what you can, I'd like to hear what you have to say about your experiences with EPYC 

[Taf the Ghost]

1 minute ago, cj09beira said:

i see it now, cool small detail right there.

might not, there is room to grow in that socket

i myself think and will go to 12, 24,48 i am skeptical of the 64 core one unless its like a server only sku or the gen next to that

Yeah, the flipped design was noticed around Epyc's main announcement. One of the designers mentioned it was aligned for length of connections.

 

As to the rumor, I thought 12c was the right place to be for a Mainstream + Server CPU package, but 7nm will let them easily fit a 16c design. But the question is: Are AMD separating the Mainstream & Server?

 

The one thing we do know is that the baseline Zen core doesn't need too much in the way of fiddling. Keeping it fed is the real problem, though they'll have improvements to make. Still, upwards of a 60-70% area shrink going to the 7nm process. They could increase the core itself by about 10% each (for AVX2 and other improvements), and they'll still have a bunch of space for 1-2 more CCX in the same size package. Which is why I think 12c makes the most sense. AMD could roll out with 12c parts at around 150 mm². That's some money if the yields are solid.

[cj09beira]

2 minutes ago, Taf the Ghost said:

Yeah, the flipped design was noticed around Epyc's main announcement. One of the designers mentioned it was aligned for length of connections.

 

As to the rumor, I thought 12c was the right place to be for a Mainstream + Server CPU package, but 7nm will let them easily fit a 16c design. But the question is: Are AMD separating the Mainstream & Server?

 

The one thing we do know is that the baseline Zen core doesn't need too much in the way of fiddling. Keeping it fed is the real problem, though they'll have improvements to make. Still, upwards of a 60-70% area shrink going to the 7nm process. They could increase the core itself by about 10% each (for AVX2 and other improvements), and they'll still have a bunch of space for 1-2 more CCX in the same size package. Which is why I think 12c makes the most sense. AMD could roll out with 12c parts at around 150 mm². That's some money if the yields are solid.

i think separating the server and consumer market would reduce profits a bit but might help them optimize the IF for one and the other, as long as they can lower latency and improve the clocks, we will see some good competition, 

7nm ryzen seems to me like a great potential perf increase, i am more worried about ryzen+ though with coffelake out and 8 cores next year, some ipc improvements would be very helpful 

[leadeater]

4 minutes ago, Taf the Ghost said:

Yeah, the flipped design was noticed around Epyc's main announcement. One of the designers mentioned it was aligned for length of connections.

 

As to the rumor, I thought 12c was the right place to be for a Mainstream + Server CPU package, but 7nm will let them easily fit a 16c design. But the question is: Are AMD separating the Mainstream & Server?

 

The one thing we do know is that the baseline Zen core doesn't need too much in the way of fiddling. Keeping it fed is the real problem, though they'll have improvements to make. Still, upwards of a 60-70% area shrink going to the 7nm process. They could increase the core itself by about 10% each (for AVX2 and other improvements), and they'll still have a bunch of space for 1-2 more CCX in the same size package. Which is why I think 12c makes the most sense. AMD could roll out with 12c parts at around 150 mm². That's some money if the yields are solid.

Unless the market and narrative shifts by the time these new CPUs launch the 12 core option does make the most sense. AMD finally releases a competitive product stack and within moments large waves of criticism comes in about how they are terrible and a failure because of slightly lower per thread performance. The amount of claims that Zen is useless for gaming is ridiculous.

 

So based on the above I'd say it's within AMD's best interest to focus on improving that single thread performance over adding more cores.