TR 3 32 or 48?

Epyon

Gawd
Joined
Oct 25, 2001
Messages
928
Rumor mill says 48 Cores but I think this shit is moving way to fast. I love that btw. So do you think they are going to stick to 32 for a while or bump it up again? Also, if the TDP is 250 watts now what does 7nm do for power requirements? 230 watts?

If it is 48 cores does that mean the TDP would remain the same or only go up like 20 watts or what ever small % it is.
 

DuronBurgerMan

[H]ard|Gawd
Joined
Mar 13, 2017
Messages
1,340
Zen2 die is 8-core and 4-core CCX again according to last leaks/rumors.

If it is, it'll be a fail unless it matches Skylake IPC and clocks - which I seriously doubt.

OTOH, I've heard plenty of rumors about a 6 core CCX, an extra 4 core CCX per die, and an 8 core CCX - the latter seems highly unlikely, though.
 

juanrga

2[H]4U
Joined
Feb 22, 2017
Messages
2,804
If it is, it'll be a fail unless it matches Skylake IPC and clocks - which I seriously doubt.

OTOH, I've heard plenty of rumors about a 6 core CCX, an extra 4 core CCX per die, and an 8 core CCX - the latter seems highly unlikely, though.

6-core CCX was the original rumor linked to the original roadmap with 48 core Rome. Those plans were skiped and Rome is now 64 core. Rome is 8 dies with 8 core each (according to certain industry sources). I guess the chinese 8-core CCX rumor comes from the server roadmaps. I guess someone saw 64 core Rome and did the next computation

64 core / 4 die = 16 core die (aka 8-core CCX). But it is

64 core / 8 die = 8 core die (aka 4-core CCX).

I guess that TR 3000 will max at 32 core.
 

DuronBurgerMan

[H]ard|Gawd
Joined
Mar 13, 2017
Messages
1,340
6-core CCX was the original rumor linked to the original roadmap with 48 core Rome. Those plans were skiped and Rome is now 64 core. Rome is 8 dies with 8 core each (according to certain industry sources). I guess the chinese 8-core CCX rumor comes from the server roadmaps. I guess someone saw 64 core Rome and did the next computation

64 core / 4 die = 16 core die (aka 8-core CCX). But it is

64 core / 8 die = 8 core die (aka 4-core CCX).

I guess that TR 3000 will max at 32 core.

What is your source for this?
 

tangoseal

[H]F Junkie
Joined
Dec 18, 2010
Messages
9,316
If it is, it'll be a fail unless it matches Skylake IPC and clocks - which I seriously doubt.

OTOH, I've heard plenty of rumors about a 6 core CCX, an extra 4 core CCX per die, and an 8 core CCX - the latter seems highly unlikely, though.

Huh??

Your statement is like wtf?
 

drescherjm

[H]F Junkie
Joined
Nov 19, 2008
Messages
14,926
I understood what he said.

There have been rumors of 6 and 8 core CCXs. so 12 and 16 core dies.

I also can see the 6 core CCX (12C die) but not the ( 8C CCX / 16C die).

Although maybe the first generation 7nm (Zen2) will not change the number of cores per CCX at all. Maybe that will happen in Zen3.
 

DuronBurgerMan

[H]ard|Gawd
Joined
Mar 13, 2017
Messages
1,340
I understood what he said.

There have been rumors of 6 and 8 core CCXs. so 12 and 16 core dies.

I also can see the 6 core CCX (12C die) but not the ( 8C CCX / 16C die).

Although maybe the first generation 7nm (Zen2) will not change the number of cores per CCX at all. Maybe that will happen in Zen3.

^what he said, Tango. I was saying the rumors are saying all sorts of shit. One rumor is that there will be a 6 core CCX (i.e. 12 cores per die, 2x 6 core CCX). Another rumor is that an extra 4 core CCX will be added, i.e. 4 cores x 3 CCX = 12 cores/die. And the last one is that there will be an 8 core CCX, i.e. 8 cores x 2 CCXs = 16 cores/die. The last one seems unlikely. The other two are possibilities, I think. If AMD sticks with a 4 core CCX x 2 complexes, though, they will have to meet or exceed Skylake clocks and IPC to stay competitive.
 

tangoseal

[H]F Junkie
Joined
Dec 18, 2010
Messages
9,316
^what he said, Tango. I was saying the rumors are saying all sorts of shit. One rumor is that there will be a 6 core CCX (i.e. 12 cores per die, 2x 6 core CCX). Another rumor is that an extra 4 core CCX will be added, i.e. 4 cores x 3 CCX = 12 cores/die. And the last one is that there will be an 8 core CCX, i.e. 8 cores x 2 CCXs = 16 cores/die. The last one seems unlikely. The other two are possibilities, I think. If AMD sticks with a 4 core CCX x 2 complexes, though, they will have to meet or exceed Skylake clocks and IPC to stay competitive.

No I was referring to the statement about it being a fail?

Absolutely not a fail. Sure you cant fo a single core operation as fast but the aggregate compute power is through the roof compared to nearly skylakes little 18 core max.

So what you deem as a fail is not in all even remotely applied to the big picture.

IPC matters in like a game for instance where a 10 or 15% IPC advantage makes a difference due to limited core usage but in a productivity environment A LOT more of a wee bit slower cores is astronomically more powerful than way less slightly faster cores.

So where would the fail be? In some imaginary perfect gaming benchmark?
 

DuronBurgerMan

[H]ard|Gawd
Joined
Mar 13, 2017
Messages
1,340
No I was referring to the statement about it being a fail?

Absolutely not a fail. Sure you cant fo a single core operation as fast but the aggregate compute power is through the roof compared to nearly skylakes little 18 core max.

So what you deem as a fail is not in all even remotely applied to the big picture.

IPC matters in like a game for instance where a 10 or 15% IPC advantage makes a difference due to limited core usage but in a productivity environment A LOT more of a wee bit slower cores is astronomically more powerful than way less slightly faster cores.

So where would the fail be? In some imaginary perfect gaming benchmark?

Right now, Zen competes well because it has a core and/or thread advantage at every conceivable price point and market position. So being down on IPC and clockspeed isn't a big deal, since it's not down by that much - the cores make up the difference except in a few specific workloads, and even in those Zen is good enough.

However, if Intel reaches core/thread parity in the mainstream (and the 28 core Intel HEDT CPU ever gets out) - which it should with the 9900k - then that becomes a problem for AMD unless one of two things happens:

1. AMD matches Skylake IPC and clocks. Then no problem.
2. AMD brings even more cores to the table by either an extra CCX, or a bigger CCX.

AMD can do either and retain the status quo. If it does neither, then AMD has a problem. For what it's worth, I think they WILL do one of those two things. I'm banking on it, in fact, given that I built an AMD rig :).
 

Topweasel

[H]ard|Gawd
Joined
Aug 12, 2005
Messages
1,509
Zen2 is a really wierd one for me. Starship was on the roadmap of having 48 cores. Which started the guess work of wondering how the CCX configuration would be. AMD needs to keep a steep lead in core count. Zen 3 is unlikely to break away from Zen 2 that much and hasn't been recently refered to as Zen 2+. But by Zen 3 they really need EPYC to be at 64 cores.

My running theory is that Zen 2 will be 4x4 CCX so double mesh there. That the 48 cores were based on AMD being cautious of the move to 7nm and planned on basically only selling Zen2 dies with a max of 12 cores to up the Yield. This is before Zen as a whole proved to be as malable as they hoped. If the yields on Zen2 are great I think we might see 16c dies. But AMD might hold back if Intel has nothing to compete with. TR2 as it stands is AMD working out a solution for growth for TR. I don't think they ever really liked the idea of going 4 active dies. But it allows them to push up margins and makes Intel's life really annoying now that they have to compete with a 32c $1500-$1700 enthusiast/workstation CPU. They took a situation where they didn't have much they could do and made it an area of great strength. For Tr3 if its 16c dies they could do a performance uplift. Go back to 2 dies and 32 cores and see a major uplift in general performance and about half the workstation applications. Maybe even drop the price to 1-1.2k. Then with TR4 do the same thing as TR2 and go 64c. If Zen 2 is 12c then they will stick with the TR2 model and offer it in 12, 16, 24, 32, 40c, and 48c versions.
 

juanrga

2[H]4U
Joined
Feb 22, 2017
Messages
2,804
Zen2 is a really wierd one for me. Starship was on the roadmap of having 48 cores. Which started the guess work of wondering how the CCX configuration would be. AMD needs to keep a steep lead in core count. Zen 3 is unlikely to break away from Zen 2 that much and hasn't been recently refered to as Zen 2+. But by Zen 3 they really need EPYC to be at 64 cores.

My running theory is that Zen 2 will be 4x4 CCX so double mesh there. That the 48 cores were based on AMD being cautious of the move to 7nm and planned on basically only selling Zen2 dies with a max of 12 cores to up the Yield. This is before Zen as a whole proved to be as malable as they hoped. If the yields on Zen2 are great I think we might see 16c dies. But AMD might hold back if Intel has nothing to compete with. TR2 as it stands is AMD working out a solution for growth for TR. I don't think they ever really liked the idea of going 4 active dies. But it allows them to push up margins and makes Intel's life really annoying now that they have to compete with a 32c $1500-$1700 enthusiast/workstation CPU. They took a situation where they didn't have much they could do and made it an area of great strength. For Tr3 if its 16c dies they could do a performance uplift. Go back to 2 dies and 32 cores and see a major uplift in general performance and about half the workstation applications. Maybe even drop the price to 1-1.2k. Then with TR4 do the same thing as TR2 and go 64c. If Zen 2 is 12c then they will stick with the TR2 model and offer it in 12, 16, 24, 32, 40c, and 48c versions.

Zen2 EPYC is 64 core, and it is made of 8 dies (with 8 cores each die) according to latest leaks.

My guess is that Zen2 TR is 4 dies aka 32 cores.
 

mjz_5

2[H]4U
Joined
May 24, 2001
Messages
3,637
If it is, it'll be a fail unless it matches Skylake IPC and clocks - which I seriously doubt.

OTOH, I've heard plenty of rumors about a 6 core CCX, an extra 4 core CCX per die, and an 8 core CCX - the latter seems highly unlikely, though.

Why are people to quick to determine AMD is a failure. If amd can get 90% of the speeds of intel at 80% of the cost, it’s a winner
 

DuronBurgerMan

[H]ard|Gawd
Joined
Mar 13, 2017
Messages
1,340
Why are people to quick to determine AMD is a failure. If amd can get 90% of the speeds of intel at 80% of the cost, it’s a winner

Oh fer crissakes. Did I say AMD is a failure? Look at my sig, man. I built an AMD rig. Would I have done that if I think they suck?

I'm saying, SPECIFICALLY, that for AMD to competitive with Intel's next generation CPUs, the need to match IPC/clocks OR offer more cores (or both). None of this is exactly rocket science, and the second method is exactly what AMD did in 2017 and most of this year, and they succeeded wildly with that. More cores at any given tier to make up for single thread deficiency. Since Intel is catching up in cores at various product tiers, AMD either needs to bolt on more, or catch up in the per-thread performance in order to keep moving along as they have.

Sometimes I feel like I'm banging my head against a wall.
 
Joined
Aug 17, 2016
Messages
704
Oh fer crissakes. Did I say AMD is a failure? Look at my sig, man. I built an AMD rig. Would I have done that if I think they suck?

I'm saying, SPECIFICALLY, that for AMD to competitive with Intel's next generation CPUs, the need to match IPC/clocks OR offer more cores (or both). None of this is exactly rocket science, and the second method is exactly what AMD did in 2017 and most of this year, and they succeeded wildly with that. More cores at any given tier to make up for single thread deficiency. Since Intel is catching up in cores at various product tiers, AMD either needs to bolt on more, or catch up in the per-thread performance in order to keep moving along as they have.

Sometimes I feel like I'm banging my head against a wall.

While I get your point if AMD is not going to increase core count for TR3 it will likely gain some speed and IPC. They do not need to match Intel because of one major factor, price. Pricing on Intel HEDT and Xeon range is crazy when compared to Threadripper.
 

DuronBurgerMan

[H]ard|Gawd
Joined
Mar 13, 2017
Messages
1,340
While I get your point if AMD is not going to increase core count for TR3 it will likely gain some speed and IPC. They do not need to match Intel because of one major factor, price. Pricing on Intel HEDT and Xeon range is crazy when compared to Threadripper.

Intel is off their fucking rocker with some of the HEDT pricing. I don't expect this to hold forever, though. At some point they'll have to return to sanity.
 
Joined
Aug 17, 2016
Messages
704
Intel is off their fucking rocker with some of the HEDT pricing. I don't expect this to hold forever, though. At some point they'll have to return to sanity.

Intel isn't going to give up that profit margin and until the market share shifts dramatically it won't have to. They still have a speed advantage and many will continue to buy Intel on name alone (business).

Where Intel will compete more is in the higher end mainstream parts like the upcoming 9900k. AMD has eaten enough into their space that they have had to go to 8 core, solder and so forth. I am very keen on seeing Zen 2 clock and IPC vs the 9000 series from Intel.
 

cyberguyz

Gawd
Joined
Aug 28, 2014
Messages
710
Intel is off their fucking rocker with some of the HEDT pricing. I don't expect this to hold forever, though. At some point they'll have to return to sanity.
As long as Intel has the majority share in the desktop and server marketplaces, don't expect them to be too competitive with their prices. Now once their bottom line starts being threatened, you can bet they will think about actually competing again.
 

DuronBurgerMan

[H]ard|Gawd
Joined
Mar 13, 2017
Messages
1,340
Intel isn't going to give up that profit margin and until the market share shifts dramatically it won't have to. They still have a speed advantage and many will continue to buy Intel on name alone (business).

Where Intel will compete more is in the higher end mainstream parts like the upcoming 9900k. AMD has eaten enough into their space that they have had to go to 8 core, solder and so forth. I am very keen on seeing Zen 2 clock and IPC vs the 9000 series from Intel.

If AMD just goes with a 6 core CCX, or one extra 4 core CCX per die, for 12 total cores on AM4 (24/48 on TR X and WX), AMD will restore the status quo ante bellum, so to speak, even if IPC and clocks only change a little bit. If not - and they stay with 8 cores on AM4, they've a lot of work to do to make up the combination of IPC and clockspeed. But it's certainly not impossible for them to do that. I'm very curious as to how it will go.
 

DuronBurgerMan

[H]ard|Gawd
Joined
Mar 13, 2017
Messages
1,340
As long as Intel has the majority share in the desktop and server marketplaces, don't expect them to be too competitive with their prices. Now once their bottom line starts being threatened, you can bet they will think about actually competing again.

That almost certainly has to be happening already. We've already heard talk about Intel holding AMD under 20% of the server market, and shit like that. They are starting to feel it. Certainly, I doubt we'd be seeing an 8 core mainstream part so soon if it weren't for AMD and market pressure. And TR's value proposition is considerably better than Intel HEDT for most use cases, provided there's any budget ceiling at all.
 

cyberguyz

Gawd
Joined
Aug 28, 2014
Messages
710
That almost certainly has to be happening already. We've already heard talk about Intel holding AMD under 20% of the server market, and shit like that. They are starting to feel it. Certainly, I doubt we'd be seeing an 8 core mainstream part so soon if it weren't for AMD and market pressure. And TR's value proposition is considerably better than Intel HEDT for most use cases, provided there's any budget ceiling at all.

You are absolutely right there. It takes someone willing to stick their neck out on the chopping block to give competitors a good swift kick in the ass and get innovation moving. When there is no real competition, the products stagnate. And Intel has been getting pretty moldy with its HEDT products.
 
Top