EPYC 64c128t rumor strikes again!
- Thread starter bauerbrazil2014
- Start date
original source is way older then that, first im aware of was CPC tweet. that must be 6 months old now.
edit: its 9 months.....
Why stop at 64? cores and 4? chiplets.
-> Active Interposer with Fully Integrated Voltage Regulator sub-system, Northbridge(Including memory controller) and Southbridge subsystems.
The calcs put that SoC up to 72 cores(6 * 12 cores) to 144 cores(9 * 16 cores). At this point 3D stacking becomes big: 288 cores to 576 cores, so on, so forth.
-> Active Interposer with Fully Integrated Voltage Regulator sub-system, Northbridge(Including memory controller) and Southbridge subsystems.
The calcs put that SoC up to 72 cores(6 * 12 cores) to 144 cores(9 * 16 cores). At this point 3D stacking becomes big: 288 cores to 576 cores, so on, so forth.
Last edited:
LTC8K6
Lifer
- Mar 10, 2004
- 28,520
- 1,575
- 126
Organik
Member
- Jul 15, 2018
- 58
- 2
- 6
LTC8K6. Speaking of 7980XE and Intel server line ; Are Xeons faster then desktops. Ive never gotten a answer to this question but I do know one thing. Which is Xeons have more trust worthy RAM ECC and some server optimized stuff. Also that you can't overclock Xeon's. Which would make them slower core vs core. Now I might be wrong and you can OC Xeons... Im just not sure. awaiting answer to this.
LTC8K6
Lifer
- Mar 10, 2004
- 28,520
- 1,575
- 126
Xeons are generally not any faster (or slower) than their desktop counterparts clock for clock. They are certified for server duty, which is a big deal, and there are Xeon chips designed to be used in multi-cpu setups, in addition to the ECC ram capability. Xeons generally have much better ram capabilities as well, like six channel ram and the ability to use a whole lot of ram. You generally cannot overclock a Xeon, and no one in their right mind would want to overclock the CPU in a server anyway.
Here is a Xeon server board for dual Xeons and a shedload of ram slots:
https://www.newegg.com/Product/Product.aspx?Item=N82E16813145047
Here is a Xeon server board for dual Xeons and a shedload of ram slots:
https://www.newegg.com/Product/Product.aspx?Item=N82E16813145047
Wait, are there people that still don't think this will be true? Hasn't there been some server focused site (that is reputable) that has outright said that AMD straight up told them that 64core EPYC is coming, but that we'll get a 48 core version before then? Maybe AMD wouldn't let official on the record statement, but they had an article outright saying that at an AMD event they were told EPYC would be going to 64 cores, and then they corrected it to say that is a later generation, and that there will be a 48core one before then). Zen 2 seems likely to be 12 core per die, with either a revision or them saving it for Zen 3 to go 16 core per die. So will be interesting if they're doing 16core dies but leaving some off for whatever reason (perhaps its just for yields early on 7nm), or if Zen 2 is going to only be 12 core per die native and then they go to 16 with a revision or wait until Zen 3. While sure 16 cores would be incredible, I think it makes business sense for AMD to do 12 first. Still more cores, still core advantage over Intel, and they can wait for the process to mature (so they either leave redundancy so they don't have to toss chips early on, or they wait for improvements to be able to push the extra cores).
Not sure why you think that, for instance the reason why AMD's desktop chips are going up in core counts is because its using the same die as their Server chips, but the server chips are using 4 of them (so it has 4x the core count). Server chips are expanding just as fast as desktop in core count.
Not sure why you think that, for instance the reason why AMD's desktop chips are going up in core counts is because its using the same die as their Server chips, but the server chips are using 4 of them (so it has 4x the core count). Server chips are expanding just as fast as desktop in core count.
Nevermind that 7LP allow for 60% smaller size at same circuitry than 14LP (that s 2.5X smaller for those who have trouble with percentages) but still, there are people who "think" that AMD will limit the core count increasement to 1.5X....
Seems that was the plan initially because they feared that GloFo yields would be terribad. Which they apparently are even if it is in better shape than Intel's 10 nm. Fortunately for AMD they can go to TSMC. Yields will still likely be pretty meh so unless you are Amazon or Google you won't be seeing the full 16 core config dies for some time.
Dont think that they ll use TSMC for CPUs, GF s Patton statement that they have not enough capacity to fill AMD need was undoubtly in reference of both CPUs and GPUs being produced at GF.
Overall GPUs are more area consuming than APU/CPUs, so it make sense to reserve GF s output to the most process performance dependant products.
Indeed, that will be the case for all the industry but on AMD s case they will benefit from their versatile uarch to make some gains in the X86 market.
Rumour has it that Zen 2 will have 10-15% better IPC than Pinnacle Ridge, dunno if that s mainly related to new instructions but since Lisa Su stated that Zen 2 will be much better in games we can assume that it s for a broad set of applications.
Gideon
Golden Member
- Nov 27, 2007
- 1,625
- 3,650
- 136
They don't have to, but it still might make sense for AMD to port the same 4 core CCX design (there may be other improvements, I'm just talking about the overall number of cores) they have now to 7nm.
First of all, it makes the move easier because there are fewer changes. This means that there aren't going to be any surprises in that area as they know what they have now works good in reality, not just on paper. It also means a smaller CCX, which is going to result in more working dies. Making your dies as small as possible makes financial sense when it's your first product on a new process. Furthermore, there are some market segments that have no need or want for a 6-core CCX or that are better suited to a 4-core CCX. While they could certainly salvage dies to fill those segments, they may have to intentionally gimp some dies that could bin higher and it's better to avoid that problem if possible. Finally, having a 6 or 8 core CCX in their pocket means that they have a compelling bump that can be released in the following year.
Keep in mind that AMD is finally just starting to become financially healthy. There's a lot of things that they could do, but they don't have the resources to execute on all of those opportunities at once. Lisa Su has been doing a great job though, and I expect that she's taken the time to figure out what the best path for the company to take is.
AMD might have just finally started to make profits again and not able to branch out widely, but these designs have also been decided on a very long time ago. Zen 2 & Zen 3 are baked in already. There are basically no major mods to be done. They are working on the projected competitor status and what was needed for the appropriate release timeframe. If anything, Intel's stumble on 10nm will allow the superiority for them. Had Intel's problems not happened, we would be in a different situation.
PeterScott
Platinum Member
- Jul 7, 2017
- 2,605
- 1,540
- 136
I am very skeptical that we will see 16 core chip as AMD next desktop Ryzen update.
That just seems like absurd overkill. Cool if it happens, but I really don't expect it.
CatMerc
Golden Member
- Jul 16, 2016
- 1,114
- 1,149
- 136
This is false.
I can't say more since the article is paywalled, but don't assume a repeat of Summit Ridge.
It make sense for APUs up to 8C/16T but for server parts, wich will be used also for high core count DTs, i would expect a 8C CCX since 2X+ density allow for doubling of the internal buses width.
Widening by 2x the data paths and keeping the frequency unchanged would still allow for an IF that is roughly 25% less power consuming than the one used in Summit Ridge while providing the same bandwith/core.
Given the characteristics of 7LP and a die that is about 200mm2 they have the following options :
16C server dedicated CPU using 8C/CCX with IF datapaths widened 2X and 32MB L3.
8C APU using 4C/CCX with IF width increased by 2x since there s a GPU to feed, this latter can be extended to 1024 SPs, wich would still be far from 2X the SP count in Raven Ridge, thus making room for somewhere between 8 and16MB L3.
Still, a 6C CCX would be a nice boost, 8C seems to be a bit much AND probably it will run intro issues with the only mem phy it has nearby. You could have 6C APUs and 12C max.
What im not looking forward is to a 16C AM4 with 4C CCXs.
What im not looking forward is to a 16C AM4 with 4C CCXs.
Why?
If they improve the "uncore" ie , distributed home agents, low latency between CCX's then it could be very nice.
TRENDING THREADS
-
Discussion Intel current and future Lakes & Rapids thread- Started by TheF34RChannel
- Replies: 23K
-
Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)- Started by DisEnchantment
- Replies: 9K
-
-
Discussion Intel Meteor, Arrow, Lunar & Panther Lakes Discussion Threads
- Started by Tigerick
- Replies: 7K
-
Facebook
Twitter