Speculation: Thread Ripper 3

nicalandia

Diamond Member
Jan 10, 2019
3,330
5,281
136
“If mainstream is moving up, then Threadripper will have to move up, up—and that’s what we’re working on,” Lisa Sue.

1. Socket SP3 compatible
2. Up to 64 cores(possibly 48 cores)
3. Same IO chip built on 14 nm(perhaps 12 nm?)
4. Top model to cost about $3,000
5. 4.7 Ghz Boost

Lets get this ball rolling

00.jpg
 
  • Like
Reactions: Drazick

Markfw

Moderator Emeritus, Elite Member
May 16, 2002
25,540
14,495
136
If there is a 64 core threadripper, I would pay 3 grand. 1800 for 32 cores today ? I even for one for $1000 (ish) but the thing will have the IO does and be far better than the 2990wx and probably for the same current draw.
 
  • Like
Reactions: Drazick

Thunder 57

Platinum Member
Aug 19, 2007
2,673
3,791
136
I think the initial release will top out at 48 cores so that it doesn't interfere with EPYC. Clocks should be high as I expect the highest clocking chips will go to TR. I expect prices to be quite high as AMD is intent on not being the "budget brand" anymore.
 

Atari2600

Golden Member
Nov 22, 2016
1,409
1,655
136
“If mainstream is moving up, then Threadripper will have to move up, up—and that’s what we’re working on,” Lisa Sue.

1. Socket SP3 compatible
2. Up to 64 cores(possibly 48 cores)
3. Same IO chip built on 14 nm(perhaps 12 nm?)
4. Top model to cost about $3,000
5. 4.7 Ghz Boost

Lets get this ball rolling

1. No. It will (almost surely) be TR4 compatible only.
2. Unlikely. 48 cores sounds more reasonable. While 64 cores is close to EPYC, X399 does seem to have a mobo limitation of 128GB DRAM [which the OEMs may be able to adjust in bios). Would 2GB/core be considered enough memory for many server applications?
3. Same IO as EPYC? Yes.
4. Definitely >$2,000, I could see 48 cores go at $2,500+.
5. Probably not - it'd be similar to Ryzen in that 4.7 would be a theoretical max, with 4.5 or 4.6 being achievable max.

I'd love to see the DRAM limit of X399 pushed up to 256GB, but I can see why they won't do that.
 

BigDaveX

Senior member
Jun 12, 2014
440
216
116
Core count is probably going to depend on exactly how good 7nm yields are. It might be viable further down the line, but I'd be VERY surprised if they were in a position to start putting out 64C Threadrippers right away.
 
  • Like
Reactions: OTG

inf64

Diamond Member
Mar 11, 2011
3,697
4,015
136
We should at least get 48C in initial batch. Clocks on par with previous gen of TR while having higher (ST) boost - 4.7Ghz?
TR3000 will utterly demolish anything that comes from intel in HEDT segment in the next 2 years. It will be brutal.
 

Kippa

Senior member
Dec 12, 2011
392
1
81
Intel have billions in reserve, couldn't then rush and come up with something to counter within the next 12 months if they put a huge amount of assets doing so? As for being brutal I am not sure it is going to get AMD huge headway because isn't HEDT on a small segment of all sales? I am thinking of upgrading to a new system in about 12 to 24 months time so my eye is on a possible Threadripper 3 system.
 

maddie

Diamond Member
Jul 18, 2010
4,738
4,667
136
Intel have billions in reserve, couldn't then rush and come up with something to counter within the next 12 months if they put a huge amount of assets doing so? As for being brutal I am not sure it is going to get AMD huge headway because isn't HEDT on a small segment of all sales? I am thinking of upgrading to a new system in about 12 to 24 months time so my eye is on a possible Threadripper 3 system.
Tick tock. One thing money can't buy is time.
 

BigDaveX

Senior member
Jun 12, 2014
440
216
116
Intel have billions in reserve, couldn't then rush and come up with something to counter within the next 12 months if they put a huge amount of assets doing so? As for being brutal I am not sure it is going to get AMD huge headway because isn't HEDT on a small segment of all sales? I am thinking of upgrading to a new system in about 12 to 24 months time so my eye is on a possible Threadripper 3 system.
Nope. Even an MCM Skylake-X - the only feasible thing they can come up with in the next 12 months, if not the next 24 months - would probably struggle to beat a 48C Threadripper 3, and would get absolutely murdered by a 64C Threadripper 3.
 

moinmoin

Diamond Member
Jun 1, 2017
4,944
7,656
136
Intel have billions in reserve, couldn't then rush and come up with something to counter within the next 12 months if they put a huge amount of assets doing so?
Like they did in the last 30 months, since the point when it became clear that Zen is competitive with Intel's offerings performance wise?
Intel was able to suddenly up the amount of cores of their Skylake derivatives after years of stagnation, but now they can't even do that to offer a 10c20t chip following the yearly cadence they had going doing that. Now it takes them one year more until 2020 to launch that, and they fill the gap with a 9900K derivative instead...
 

zir_blazer

Golden Member
Jun 6, 2013
1,163
405
136
1. No. It will (almost surely) be TR4 compatible only.
2. Unlikely. 48 cores sounds more reasonable. While 64 cores is close to EPYC, X399 does seem to have a mobo limitation of 128GB DRAM [which the OEMs may be able to adjust in bios). Would 2GB/core be considered enough memory for many server applications?
3. Same IO as EPYC? Yes.
4. Definitely >$2,000, I could see 48 cores go at $2,500+.
5. Probably not - it'd be similar to Ryzen in that 4.7 would be a theoretical max, with 4.5 or 4.6 being achievable max.

I'd love to see the DRAM limit of X399 pushed up to 256GB, but I can see why they won't do that.
As far that I know, TR4 and SP3 are physically compatible. Giving the fact that there is no public documentation, we don't know if the pinout is close enough to call the Sockets electrically compatible too, which is possible assuming that in TR4 all the pins for the extra 2 RAM channels and 64 PCIe Lanes than SP3 has are marked as reserved, not connected, or something, not repurposed for something else, which would make them incompatible.
I think there was some guy that tried an EPYC on a TR4 Motherboard (Or the other way around? I don't remember), and while it didn't managed to fully POST, the Motherboard debug LED pointed that it at least detected the Processor, and it didn't blow up.

After seeing the Workstation TR models using 4 functional dies of which 2 have no I/O, I don't know why AMD didn't planned for EPYCs to be used that way in the first place, as it should technically be possible. Same way that there are EPYCs that only work in Single Processor mode (The ones whose model number finishes in a P) and others that can do either Single or Dual Processor, the same EPYC could work with either reduced I/O in a ThreadRipper Motherboard or go full. Viceversa should also be possible, TR in a SP3 Motherboard with reduced I/O (Like populating a LGA 2066 Motherboard with a Kaby Lake-X instead of a Skylake-X). TR and EPYC should have been a single Socket to begin with.
I don't know if Rome can merge them considering that the same IO die would work either full in EPYC or with a reduced feature set for TR (Giving how much design AMD is reusing between lines, there is no way that TR gets its own exclusive IO die), as that is close to what I'm proposing.

The problem with the current lineup is pretty much absolute market segmentation. TR have models that clock higher than P series EPYCs, but TR Motherboards are aimed at high end overclockers and gamers, not that much of Workstation usage, and only one announced with a BMC for Server use (AsRock Rack X399D8A-2T). TR makes more sense for Workstation than EPYC, but you lack Motherboards. Viceversa also applies, there are few SP3 Motherboards with prosumer features for users that prefer EPYC over TR because they want massive I/O. If both Sockets were unified you could at least have more versatile Processor choice, mostly useful if you want high frequency low IO Workstation type Processor on a Motherboard with Server features.



The 128 GiB RAM limitation is because Motherboards were intended to be used with 8 16 GiB modules and manufacturers may not go back to upgrade the specifications even if they work with 8 32 GiB. It happened with my AM3 ASUS M4A785TD-V EVO, official specs says it supports 4 * 4 GiB DDR3 but worked no problems with 4 * 8 GiB. If you want more, you need Slots wired for RDIMM, which means EPYC.
 

Markfw

Moderator Emeritus, Elite Member
May 16, 2002
25,540
14,495
136
As far that I know, TR4 and SP3 are physically compatible. Giving the fact that there is no public documentation, we don't know if the pinout is close enough to call the Sockets electrically compatible too, which is possible assuming that in TR4 all the pins for the extra 2 RAM channels and 64 PCIe Lanes than SP3 has are marked as reserved, not connected, or something, not repurposed for something else, which would make them incompatible.
I think there was some guy that tried an EPYC on a TR4 Motherboard (Or the other way around? I don't remember), and while it didn't managed to fully POST, the Motherboard debug LED pointed that it at least detected the Processor, and it didn't blow up.

After seeing the Workstation TR models using 4 functional dies of which 2 have no I/O, I don't know why AMD didn't planned for EPYCs to be used that way in the first place, as it should technically be possible. Same way that there are EPYCs that only work in Single Processor mode (The ones whose model number finishes in a P) and others that can do either Single or Dual Processor, the same EPYC could work with either reduced I/O in a ThreadRipper Motherboard or go full. Viceversa should also be possible, TR in a SP3 Motherboard with reduced I/O (Like populating a LGA 2066 Motherboard with a Kaby Lake-X instead of a Skylake-X). TR and EPYC should have been a single Socket to begin with.
I don't know if Rome can merge them considering that the same IO die would work either full in EPYC or with a reduced feature set for TR (Giving how much design AMD is reusing between lines, there is no way that TR gets its own exclusive IO die), as that is close to what I'm proposing.

The problem with the current lineup is pretty much absolute market segmentation. TR have models that clock higher than P series EPYCs, but TR Motherboards are aimed at high end overclockers and gamers, not that much of Workstation usage, and only one announced with a BMC for Server use (AsRock Rack X399D8A-2T). TR makes more sense for Workstation than EPYC, but you lack Motherboards. Viceversa also applies, there are few SP3 Motherboards with prosumer features for users that prefer EPYC over TR because they want massive I/O. If both Sockets were unified you could at least have more versatile Processor choice, mostly useful if you want high frequency low IO Workstation type Processor on a Motherboard with Server features.



The 128 GiB RAM limitation is because Motherboards were intended to be used with 8 16 GiB modules and manufacturers may not go back to upgrade the specifications even if they work with 8 32 GiB. It happened with my AM3 ASUS M4A785TD-V EVO, official specs says it supports 4 * 4 GiB DDR3 but worked no problems with 4 * 8 GiB. If you want more, you need Slots wired for RDIMM, which means EPYC.
SP3 and TR4 are NOT identical. There was an article a while back. There is one pin used or not used to signal which socet it is, but also, the other 4 channels are not hooked up. They could not get either a EPYC or threadripper (I forget which way they tried it) to post They tried to cover that one pin, but the memory must have stopped it from posting.
 
  • Like
Reactions: Drazick

zir_blazer

Golden Member
Jun 6, 2013
1,163
405
136
SP3 and TR4 are NOT identical. There was an article a while back. There is one pin used or not used to signal which socet it is, but also, the other 4 channels are not hooked up. They could not get either a EPYC or threadripper (I forget which way they tried it) to post They tried to cover that one pin, but the memory must have stopped it from posting.
Having 4 channels not hooked up I already mentioned. Workstation TR works exactly like that.

I wouldn't consider ONE pin a showstopper. Is not like we haven't seen mods like LGA 771 to 775. Or more recently, 6C Coffee Lake LGA 1151 to Skylake LGA 1151 (This one is harder because you also require to deal with the Intel ME instead of just physical pinout). Don't see any reason why EPYC and TR cross compatibility should be impossible.
 

Markfw

Moderator Emeritus, Elite Member
May 16, 2002
25,540
14,495
136
Having 4 channels not hooked up I already mentioned. Workstation TR works exactly like that.

I wouldn't consider ONE pin a showstopper. Is not like we haven't seen mods like LGA 771 to 775. Or more recently, 6C Coffee Lake LGA 1151 to Skylake LGA 1151 (This one is harder because you also require to deal with the Intel ME instead of just physical pinout). Don't see any reason why EPYC and TR cross compatibility should be impossible.
Well, if I could find the article, it was pros, and they could not get it to work,.
 
  • Like
Reactions: Drazick

moinmoin

Diamond Member
Jun 1, 2017
4,944
7,656
136
I think it was der8auer in this video (which was also reported in the press):
 

Kenmitch

Diamond Member
Oct 10, 1999
8,505
2,249
136
Intel have billions in reserve, couldn't then rush and come up with something to counter within the next 12 months if they put a huge amount of assets doing so? As for being brutal I am not sure it is going to get AMD huge headway because isn't HEDT on a small segment of all sales? I am thinking of upgrading to a new system in about 12 to 24 months time so my eye is on a possible Threadripper 3 system.

Haven't they already wasted billions on 10nm and we've seen how that's been working out.

HEDT along with desktop is just the trickle down effect of the core war. The lucrative server market is what AMD is targeting.
 

thigobr

Senior member
Sep 4, 2016
231
165
116
I would say they do work... I am using 2x32GB Samsung DIMMs (M378A4G43MB1-CTD) right now on AM4. The memory controller for Threadripper is the same so the only thing that could limit you is firmware.
 

amd6502

Senior member
Apr 21, 2017
971
360
136
I think a 4 elite binned chiplet flagship 24c/48t threatripper is most likely.

They may do a 32c extreme multithread compute-loads WX version for home or enterprise server, but it would not be the flagship and use four lower binned 8c chiplets. These would be dies that did not quite make the power efficiency cutoff for Epyc; basically a highly discounted alternative to their official server line.
 

Atari2600

Golden Member
Nov 22, 2016
1,409
1,655
136
So whats the story? I thought there was meant to be more info today at some point....?