Wow you guys seen this. Those x-rays made me feel all warm and nice inside. You can just about see the traces, I will probably be studing these for the next month.
“Feel free to use those pictures anywhere as long as you link back to the original video.
Thanks - Roman “der8auer””
Epyc interconnect (not representative of the x-ray):
The PCB between the two are definitely different. Why would they have such similar packages if they were going to redesign the PCB anyway? Why would they need two different sockets for a different pin out? None of this is making sense… Its like needing a whole new platform for a CPU refresh. Epyc and TR could have been on the same socket and had anywhere from 1-32 cores. Or TR could have been a 2 die smaller package and Epyc could have been the full 4 (they had to change the PCB between the two anyway).
I think I got it. So TR released in 2017 July-August. Winter 2018 is ending and there is stll no 32 core TR in sight. Meanwhile Epyc is undercutting every intel counter part and offering the full 32. So why does TR exist? Its sort of market scalping, they are doing the nvidia strat. I am sure AMD could release a 32 TR back in summer 2017 but they knew they had intel beat everywhere on the consumer market, so instead of selling the best they got, they will sell 16 TR until intel catches up, only to drop prices and sell a 32 TR, just like nvidia did with the 1070 - 1070 Ti. At the same time they couldn’t give up the server segment by only selling 16 cores so they had to seperate the markets by making another platform and thus Epyc was unleashed.
There is also the naming scheme for TR 1900, 1920, 1950… what about a 1970 and a 1990?
This is the best explanation I have for the existance of TR alongside Epyc.
I was I bit concered thinking that maybe a 32 TR was impossible, but now I am sure it will happen. When? Well that will depend entirely on intel.
Fairplay to AMD, I gotta give them credit for being this clever. After all they are a relatively small company and intel has been scalping and gouging since the formation of planet earth.
I wouldn’t be so sure, does a 32 core TR really need 8 channel memory and more lanes? I wouldn’t compare the 32 Epyc to possible future 32 TR. Once again they re-designed the PCBs between the two…
Anyways, I am speculating based on smell at this point. Thanks for your posts I found the articles intriguing. Maybe one day we can meet up and I’ll treat to some “Meow Mix” or “9Lives”…
Eight memory channels would be a bit of a challenge for motherboard makers to get into a ATX board I think. I guess you could get by with one DIMM per channel if they manage to put them on the board? One DIMM per channel for ultimate RAM speed or some similar marketing, lol. It would work though and is maybe the best solution?
The thing is that every die has two channels, so with only four channels either two dies gets no memory channels or every die get only one. If the latter is even possible, I would guess two dies gets two channels and the other two will have to get data over the Infinity Fabric between dies. Would probably make latency to memory for those dies interesting?
All theoretical ofc, but it would certainly be interesting if AMD managed to make a 32 core TR part work with some kind of ATX-like motherboard. Tyan has a single socket Epyc board that is EATX and only 33 cm wide, but it also doesn’t have the typical PCIe slots you find on a std ATX board (it uses riser cards etc).
This does prove all 4 dies can be active despite the BIOS not playing nice. And now we know exactly which 2 dies are truly active. That Asetek coldplate really ain’t gonna do much with how it’s spaced out. Only the Enermax cooler will do.
It proves they get power, but nothing else (yet). These were the two that were reported to be active, nothing in the video was surprise AFAIC.
We can see how thick that IHS is though and that’s interesting. That’s a lot of thermal mass to sink, but no, we already knew the Asetek was “good enough” but that was about it.
This could prove though that lapping of this CPU IHS to reduce the distance between the coldplate (assuming it’s full coverage like an Enermax one) and the die could improve thermals considering how thick the IHS is.