64 core AMD Rome CPU

[dizmo]

9 hours ago, VegetableStu said:

i wonder if AM5 (assuming) would finally go LGA o_o

One can only hope. I probably wouldn't consider AMD until then anyway.

Though, if they offer poor memory support with the start of DDR5...maybe longer.

10 hours ago, Stefan Payne said:

Why?!

You've seen the Issues that Skylake had with too much Preassure (like Scythe Ashura) and a couple of other things?

With a PGA Socket that wouldn't have happpened. They are just more robust in general...

 

LGA is just more fragile and causes Problems. Especially with higher pincounts. Or currents...

 

And in 10 years, the CPU will be fine anyway, the Motherboard is the Problem. Those are getting rarer. And when you kill a pin, its over for it...

Wasn't that more of a problem with certain coolers? And the heat spreader? That has nothing to do with PGA vs LGA...

PGA would be significantly worse with higher pin counts. Which is why AMD uses LGA for TR...

[cj09beira]

27 minutes ago, leadeater said:

TSMC fabs the 7nm chiplets and has 14nm capability so I don't see why they wouldn't source from them for the I/O dies, they still have contracts with GloFo so they also make sense since they are obligated to fab some stuff through them.

with the node being the exact same as samsungs it makes mores sense to use that, but they probably wont use any other 

[Stefan Payne]

3 hours ago, fasauceome said:

Of course there were those really nice 6 core xeons and i7s that were HEDT on the consumer socket, which may have been a subtle driving force for higher core count use in games.

Those were irrelevant as not many people had them due to very high prices.

The 6 Core LGA1366 Xeons were only relevant after they were phased out - wich was half a decade or so later. But with a Shortage of Boards, they're pretty useless anyway...

 

The Argument that would make the most sense would have been Thuban as it was reasonably priced and had 6 Cores back in 2010 as those were more affordable at the time.

 

But nobody would (or should) care about 2% or less of of the potential market.

1 hour ago, dizmo said:

Wasn't that more of a problem with certain coolers? And the heat spreader? That has nothing to do with PGA vs LGA...

PGA would be significantly worse with higher pin counts. Which is why AMD uses LGA for TR...

Look at the design of LGA and PGA.

What can happen with PGA and a lid on?
Not much because you have a hard surface beneeth it.

With LGA you have no real hard surface beneeth it and its more lying at the corners...

[leadeater]

1 hour ago, cj09beira said:

with the node being the exact same as samsungs it makes mores sense to use that, but they probably wont use any other 

It's not quite exactly the same but the tech is licensed off Samsung, it just makes sense to use TSMC for them as well since you are already fabing stuff there for the CPU. I would think TSMC is primary source for both the I/O and chiplet dies with possibly GloFlo as secondary source for I/O die.

[S w a t s o n]

9 minutes ago, leadeater said:

It's not quite exactly the same but the tech is licensed off Samsung, it just makes sense to use TSMC for them as well since you are already fabing stuff there for the CPU. I would think TSMC is primary source for both the I/O and chiplet dies with possibly GloFlo as secondary source for I/O die.

They could be using the IO die to fulfill their wafer purchase agreements. I'm sure the downfall of glofo's 7nm has triggered another renegotiation for the agreement but I doubt it's just gone

[leadeater]

3 minutes ago, S w a t s o n said:

They could be using the IO die to fulfill their wafer purchase agreements. I'm sure the downfall of glofo's 7nm has triggered another renegotiation for the agreement but I doubt it's just gone

It was renegotiated we just don't know in what way, was a news topic a while ago about it when GloFo annouced the canning of 7nm. I was wondering how AMD would actually be able to utilize GloFo after that if their whole product stack was going 7nm (CPU and GPU), was thinking some lower end CPUs would stay at 12nm or something but this I/O die should really help AMD keep their contractual commitments to using GloFo.

[porina]

https://www.anandtech.com/show/13598/amd-64-core-rome-deployment-hlrs-hawk-at-235-ghz

 

Deployment of 64-core Rome CPUs in supercomputer. They list a clock of 2.35 GHz.

[Fnige]

Gaming PC's using this...

 

[williamcll]

7 hours ago, SlimyPython said:

Gaming PC's using this...

 

Will be awful. 

[The Benjamins]

11 hours ago, porina said:

https://www.anandtech.com/show/13598/amd-64-core-rome-deployment-hlrs-hawk-at-235-ghz

 

Deployment of 64-core Rome CPUs in supercomputer. They list a clock of 2.35 GHz.

For context the top Gen 1 Epyc (7601) has a base of 2.2 all core boost of 2.7 and a Max single core of 3.2

[GoldenLag]

12 minutes ago, The Benjamins said:

For context the top Gen 1 Epyc (7601) has a base of 2.2 all core boost of 2.7 and a Max single core of 3.2

So essentially they went for power efficiency like one would expect for a server. 

 

The nice thing is that they are probably overclockable. The 7601 was at least.

[cj09beira]

 

that my friends is what 8 milan cpus with 8x8 sticks of ram all watercooled

the density here is insane, if using rome it would be 512 cores 1024 threads, in this little space (ram up to 16 TB)

(edited due to bad math)

[The Benjamins]

29 minutes ago, cj09beira said:

 

that my friends is what 8 milan cpus with 8x8 sticks of ram all watercooled

the density here is insane, if using rome it would be 256 cores 512 threads, in this little space (ram up to 16 TB)

Milan, as in Epyc 3?

 

Where did you get this?

 

EDIT:

Found this, it also shows Nvidias Volta next

https://www.tomshardware.co.uk/amd-epyc-milan-shasta-exascale,news-59429.html

[cj09beira]

2 minutes ago, The Benjamins said:

Milan, as in Epyc 3?

 

Where did you get this?

here https://www.tomshardware.com/news/amd-epyc-milan-shasta-exascale,38067.html

[The Benjamins]

Remember this leak
https://wccftech.com/amd-epyc-rome-7nm-64-core-cpu-performance-benchmark-leak/#comment-4100195712

 

Its starting to look realistic.

If 2 Epyc 7601 can get 6000-7000 with a max all core boost of 2.7 then I can see a 2S Rome getting around 12500 at 2.35

[schwellmo92]

55 minutes ago, cj09beira said:

 

that my friends is what 8 milan cpus with 8x8 sticks of ram all watercooled

the density here is insane, if using rome it would be 256 cores 512 threads, in this little space (ram up to 16 TB)

<snip>

8x64=256?

[leadeater]

6 minutes ago, The Benjamins said:

If 2 Epyc 7601 can get 6000-7000 with a max all core boost of 2.7 then I can see Rome getting around 12500 at 2.35

Wouldn't Rome be more similar to the 2 7601? I thought CB wasn't heavy AVX2?

 

Edit:

Ohh 2x Rome, durp durp.

[cj09beira]

11 minutes ago, schwellmo92 said:

8x64=256?

you're right, dam that means 1024 threads per node O.O just woW

[The Benjamins]

13 minutes ago, leadeater said:

Wouldn't Rome be more similar to the 2 7601? I thought CB wasn't heavy AVX2?

 

Edit:

Ohh 2x Rome, durp durp.

Ya i fixed that.

 

So around 6000-7000 for 1S Rome and 11000-13000 for 2S Rome

[Mihle]

On 11/9/2018 at 12:26 AM, leadeater said:

It was renegotiated we just don't know in what way, was a news topic a while ago about it when GloFo annouced the canning of 7nm. I was wondering how AMD would actually be able to utilize GloFo after that if their whole product stack was going 7nm (CPU and GPU), was thinking some lower end CPUs would stay at 12nm or something but this I/O die should really help AMD keep their contractual commitments to using GloFo.

In the interview that was a few days ago I think they said they are indeed using GloFlo for the IO die. If you didn't already know that. I am on my phone so am not going to find a link rn. 

[Bananasplit_00]

On 11/8/2018 at 1:12 AM, 2Buck said:

Oh yes, this. People argue that bent PGA pins are easier to recover than bent LGA pins, which is true, but I've never gotten into a situation where I needed to recover a bent LGA pin. I just don't see how you could possibly bend an LGA pin. Just be careful when installing the CPU and store the motherboard in a box (hopefully it's own box). On the other hand, bending a PGA pin is easy. You could drop the CPU from an inch high and have a a disaster on you hands. I've spent hours recovering bent pins, it isn't fun. Storing and transporting a PGA CPU is a lot harder than LGA CPUs.

i have fixed both PGA and LGA pinns and PGA is much much easier but its not like its that hard to bend back an LGA socket if you take your time. personally i dont care about PGA vs LGA but i just thought id mention that.

[Taf the Ghost]

 

Slightly different slides for a Chinese event.

[cj09beira]

2 hours ago, Taf the Ghost said:

 

Slightly different slides for a Chinese event.

this is important, it means that the diagram might actually be what amd used, so we might actually see 16 cores (or 12) on am4, now, i just need to find some money hmmm

[Fasauceome]

6 minutes ago, cj09beira said:

this is important, it means that the diagram might actually be what amd used, so we might actually see 16 cores (or 12) on am4, now, i just need to find some money hmmm

They'll probably go up to 16 because it's 8 cores per ccx now

[cj09beira]

5 minutes ago, fasauceome said:

They'll probably go up to 16 because it's 8 cores per ccx now

its 8 cores per die, ccx core count will probably stay the same