[RUMOR] AMD EPYC 7004 Genoa Zen 4 CPU Allegedly has 96 cores and 192 threads with 12-channel DDR5-5200 memory support

[poochyena]

Summary

 

AMD EPYC 7004 Genoa Zen 4 CPU Allegedly has 96 cores and 192 threads with 12-channel DDR5-5200 memory support

 

Quotes

Quote

Leaker and Twitter user @ExecuFix claims to have the inside scoop on one of AMD's eventual Genoa chips, one that wields 96 cores and 192 threads. They say it will also feature 12-channel DDR5-5200 memory support, as well as 128 PCI Express 5.0 lanes (160 lanes for 2P configurations), and have a 350W TDP configurable up to 400W.

 

My thoughts

I'm amazed by AMD continuing to advance CPUs so quickly so continuously. Intel would give us a few bumps in performance every few years, but AMD has impressed every year repeatably. I was kinda expected them to get a little complacent by now.

 

Sources

https://hothardware.com/news/amd-epyc-7004-genoa-zen-4-cpu-ddr5-lga-6096-socket

[Downkey]

2 minutes ago, poochyena said:

Summary

 

AMD EPYC 7004 Genoa Zen 4 CPU Allegedly has 96 cores and 192 threads with 12-channel DDR5-5200 memory support

 

Quotes

 

My thoughts

I'm amazed by AMD continuing to advance CPUs so quickly so continuously. Intel would give us a few bumps in performance every few years, but AMD has impressed every year repeatably. I was kinda expected them to get a little complacent by now.

 

Sources

https://hothardware.com/news/amd-epyc-7004-genoa-zen-4-cpu-ddr5-lga-6096-socket

DDR5. That’s all I have to say

[WhitetailAni]

I can't wait for 24 gamers 1 CPU.

[Guest]

That's disgusting.

 

Spoiler

I really don't want to say there's a good kind of disgusting, but I don't mean this negatively lol.

 

[Helpful Tech Witch]

can't wait for a 192core server video from ltt.

[flibberdipper]

Think I'm about to bust, holy shit.

[Doobeedoo]

Oho... this will be insane to see, I wonder will there be even high core count with this series even. Can't wait to see stacked dies.

[Kisai]

37 minutes ago, FakeKGB said:

I can't wait for 24 gamers 1 CPU.

Considering that getting 24, let alone 96 GPU's or USB ports is going to be an issue.

 

One of the key reasons high density cpu's don't end up in desktops is that they typically run at half the clock speed. You know, at a speed that a game will be noticed, but server applications will not.

 

[Electronics Wizardy]

29 minutes ago, Kisai said:

Considering that getting 24, let alone 96 GPU's or USB ports is going to be an issue.

 

One of the key reasons high density cpu's don't end up in desktops is that they typically run at half the clock speed. You know, at a speed that a game will be noticed, but server applications will not.

 

Id like to see LTT do the multi gamers one system with something like vmware horizon. Then people can use thin clients, and multiple users can share a single gpu. Also its fully supported so there should be much fewer issues than doing passthough.

[leadeater]

Just remember, if something sounds too good to be true... it's probably fake or inaccurate.

 

• 12 Chiplets: Possible but also physically and electrically problematic [Edit again] (Read the source, actually claiming 12 chiplets. I should read sources first lol) 12 core CCD makes far more sense than 8 core CCD with Zen 4 if this many cores.
• 12 DDR5 Memory Channels: Possible, unnecessary and also physically and electrically problematic
• 128/160 PCIe 5.0: Likely correct
• 350W-400W TDP: Unlikely, lower TDP parts are already proprietary water cooling only and this is 100W+ more than those

Random meaningless credibility score from random dude on the internet: 3/10

 

Quote

Supposedly, anyway. Another leaker, a "retired engineer and businessman" who goes by @chiakokhua on Twitter, replied to the mockup saying, "Bet you $6 this is incorrect." This prompted a response by @ExecuFix, who said they will "gladly take your $6."

I'll take that bet and 10x it, $6 is no fun.

[Windows7ge]

7003 hasn't launched yet that I'm aware of so I'm skeptical of this rumor. It'll be fascinating if they do squeeze 96 cores into the SP3 socket though.

 

I'll also want one.

[Dabombinable]

And my family are all still using CPU ranging from mid 2000's dual cores, all the way up to pre 2015 quad cores.

[Dutch_Master]

Insane. Some doubt if the announcement/rumour is false, there is a chance it is. But I'm also aware that it could well be true as it would make sense for AMD to push the boundaries while Intel is still down, making use of their momentum in the server market. So a 6/10 for credibility IMO.

 

Price-wise, I'd say well above the 10k for the current 7000 series, around the $15-17k (US) I reckon. Gonna be a very popular chip for supercomputing clusters and other virtualisation tasks I think.

 

Can't wait for the server builds from LTT, Level1Techs and STH   Particularly Linus opening some system monitoring tool: "Look at al those coooooooooooooooooohooooooooooooooooooohooooooooooooooores! It's not gonna fit on the screen!! We need a bigger screen, where's that 65" TV?!"

[TetraSky]

That's a lot of cores!

And single core performance will be able as good as a pentium 3, isn't it.

[leadeater]

28 minutes ago, Dutch_Master said:

But I'm also aware that it could well be true as it would make sense for AMD to push the boundaries while Intel is still down, making use of their momentum in the server market.

The number of memory channels is very unrealistic and the TDP even more so. There might be some custom chips for HPC that high but they will be from the likes of Cray only with water cooling.

 

12 memory channels is basically impossible to fit physically in to the most common standard server formfactors and with the current 8 of DDR4 the move to DDR5 should mean 8 channels would be sufficient for more than 96 cores.

 

You cannot make this wider, so if the socket is larger and there are 12 memory channels where are the slots supposed to go? Rotate the socket 90 degrees and make it longer, well not everyone wants much longer than what is in usage now.

[Vishera]

2 hours ago, leadeater said:

• 12 Chiplets
• 350W-400W TDP

Those two are very good reasons to doubt it,

It doesn't make sense to have 12 chaplets - it will create latency problems,potential bandwidth bottlenecks and more.

And it doesn't make sense to have a 350W~400W CPU in a market that is all about performance per watt,which is reflected by it's lower clock speeds.

 

2 hours ago, leadeater said:

Random meaningless credibility score from random dude on the internet: 3/10

I agree,a lot of things just don't make sense here.

[leadeater]

Just now, Vishera said:

It doesn't make sense to have 12 chaplets - it will create latency problems,potential bandwidth bottlenecks and more.

I'm not even sure it's possible to connect them to the IOD, there's actual traces in the sub-straight and suffice to say shit gets complicated as you increase the number of die you need to interconnect.

[Mling]

No one will be able to afford to license Windows and SQL server on this thing.

[Vishera]

8 minutes ago, leadeater said:

I'm not even sure it's possible to connect them to the IOD

Good point.

8 minutes ago, leadeater said:

there's actual traces in the sub-straight and suffice to say shit gets complicated as you increase the number of die you need to interconnect.

Yeah,I know that.

From my calculations 8 CCDs is possible,but 12 is a fantasy (for now).

[Bombastinator]

6 hours ago, SGT-AMD said:

96 cores is enough for a herd of cows......

A single core is more than enough for a herd of cows.  Cows can’t type.

[Spindel]

3 hours ago, leadeater said:

• 350W-400W TDP: Unlikely, lower TDP parts are already proprietary water cooling only and this is 100W+ more than those

Area you really share the dissipated power is a problem considering it being loads of different dies on the CPU and thus spreading out the heat over a larger area?

[leadeater]

24 minutes ago, Spindel said:

Area you really share the dissipated power is a problem considering it being loads of different dies on the CPU and thus spreading out the heat over a larger area?

Yes, the HPC only 7H12 is 280W and deployed only with water cooling. Intel's Xeon 9200 series are also BGA only and come in water cooled systems. Once you go above 250W on a socket air cooling in servers become problematic. You can but it rules out a lot of configurations so no many want CPU with that much power usage. Then you also start encountering power density problems in racks, not ever hosting provider and business has 20kw+ per rack capability

 

[Spindel]

1 minute ago, leadeater said:

Yes, the HPC only 7H12 is 280W and deployed only with water cooling. Intel's Xeon 9200 series are also BGA only and come in water cooled systems. Once you go above 250W on a socket air cooling in servers become problematic. You can but it rules out a lot of configurations so no many want CPU with that much power usage. Then you also start encountering power density problems in racks, not ever hosting provider and business has 20kw+ per rack capability

 

BTW sorry for the really bad english in the post you quoted, autocorrect made a really wierd number on that message and I didn’t proof read

[CarlBar]

Whilst i think the doubts raised are solid i do want to raise a couple of counterpoints.

 

1. Transistors per core generally goes up as IPC does. Even with a smaller node there's a fundamental limit to how many transistors they can squeeze into a given amopunt of silicon and the whole point of chiplets was to trim silicon area per die down so they're not going to want to push it back up. And that means if AMD in the long term wants to ever increase core counts again they have to solve the interconnects issue somehow, it's a non-optional issue for them. So whilst it might be a valid criticism, it's also at the same time somthing AMD is eventually going to debut an answer for in some form. This generation could be that debut, then again it may not. We'll see.

 

2. Memory channel count. Nothing says that every board has to support more than 8 channels. People have been saying current epic can hit memory bottlenecks with just 64 cores in some scenarios. Building extra memory channels into the IO die from the start may be an attempt along with the DDR5 to give those with such workloads an option.

[leadeater]

4 minutes ago, Spindel said:

BTW sorry for the really bad english in the post you quoted, autocorrect made a really wierd number on that message and I didn’t proof read

Don't worry my post just then is full of errors too hahaha