|
Interposers and similar tech can be on older nodes, right? I imagine that might fill out the majority of the WSA until 2020 even just with CPUs (if infinity fabric can use an older node).
|
# ? Aug 28, 2018 01:40 |
|
|
# ? May 23, 2024 13:29 |
|
GRINDCORE MEGGIDO posted:Welp. Be interesting to see the new agreement they create with AMD until 2020. If they needed to kill that volume they could license some DRAM IP and just churn it out like mad. There's still shortage of DDR4 that they could use to get out from it.
|
# ? Aug 28, 2018 02:01 |
|
NewFatMike posted:Interposers and similar tech can be on older nodes, right? I imagine that might fill out the majority of the WSA until 2020 even just with CPUs (if infinity fabric can use an older node). They'd have to be fairly old nodes for that to make sense, up to 40-45nm I think. I think GloFo retired those nodes?
|
# ? Aug 28, 2018 03:29 |
|
Vega's interposer was built on 28nm. GloFo keeps touting their "FDX" FD-SOI 22nm and 12nm processes, may as well leverage those. IIRC, the interposer presently doesn't even benefit from FinFETs anyways, though of course that is subject to change in the future.
SwissArmyDruid fucked around with this message at 09:14 on Aug 28, 2018 |
# ? Aug 28, 2018 09:12 |
|
Harik posted:Why would they do a new agreement? I thought what was poo poo about the agreement was AMD is committed to a volume with them no matter what processes work or don't work. Ah. I assumed part of the agreement was that they'd have to deliver pre arranged nodes. Fair enough
|
# ? Aug 28, 2018 09:52 |
|
SwissArmyDruid posted:Vega's interposer was built on 28nm. GloFo keeps touting their "FDX" FD-SOI 22nm and 12nm processes, may as well leverage those. IIRC, the interposer presently doesn't even benefit from FinFETs anyways, though of course that is subject to change in the future. Wait, why though? 28nm would be quite expensive in comparison to older nodes, passive or not, wouldn't it?
|
# ? Aug 28, 2018 16:40 |
|
Harik posted:Why would they do a new agreement? I thought what was poo poo about the agreement was AMD is committed to a volume with them no matter what processes work or don't work. Don't think it works that way but I strongly suspect that the WSA is now dead Too bad. 7nm GF had some decent press and could have been a contender.
|
# ? Aug 28, 2018 16:51 |
|
EmpyreanFlux posted:Wait, why though? 28nm would be quite expensive in comparison to older nodes, passive or not, wouldn't it? Could you clarify your question? I couldn't tell you why Vega's interposer was reported built on 28nm. And I couldn't tell you why not a larger node, except to say that there were probably some genuine benefits to be found in terms of heat generation that come from smaller node sizes, though, something that Vega struggles with even under the best of circumstances, let alone the stacked chip design.
|
# ? Aug 28, 2018 18:24 |
|
SwissArmyDruid posted:Could you clarify your question? I couldn't tell you why Vega's interposer was reported built on 28nm. And I couldn't tell you why not a larger node, except to say that there were probably some genuine benefits to be found in terms of heat generation that come from smaller node sizes, though, something that Vega struggles with even under the best of circumstances, let alone the stacked chip design. I mean, I don't understand the reasoning for selecting a smaller node for null silicon. Active Interposers might make sense, but if it's not doing anything than why us a more expensive node, say compared to hitting up UMC or TSMC for 65nm. EmpyreanFlux fucked around with this message at 21:11 on Aug 28, 2018 |
# ? Aug 28, 2018 19:13 |
|
EmpyreanFlux posted:I mean, I don't understanding the reasoning for selecting a smaller node for null silicon. Active Interposers might make sense, but if it's not doing anything than why us a more expensive node, say compared to hitting up UMC or TSMC for 65nm. Got to pay GloFlo for unused wafers anyways, might as well put em to use
|
# ? Aug 28, 2018 19:22 |
|
Malcolm XML posted:7nm GF had some decent press and could have been a contender. I'm skeptical that their 7nm process was ever going to be a contender. Considering how much Intel has struggled with a drat near unlimited budget, it wouldn't surprise me if the GF 7nm process was just not going to hit its numbers on yield and/or performance, which lead to AMD bouncing. "We made a great product that is in huge demand so were cancelling it" is essentially the story GF is selling and ehhhh seems like bullshit to me. The big question for me with Intel, Samsung, and now Glofo all struggling at (roughly) this same size process is... how is TSMC pulling it off? Or are their 7nm yields just terrible and we should expect inflated prices? Or maybe the perf/watt isn't as hyped? Interesting times.
|
# ? Aug 28, 2018 19:33 |
|
EmpyreanFlux posted:I mean, I don't understanding the reasoning for selecting a smaller node for null silicon. Active Interposers might make sense, but if it's not doing anything than why us a more expensive node, say compared to hitting up UMC or TSMC for 65nm. An interposer is basically just a chunk of silicon with a bunch of metal layers, and MAYBE some conditioning or timing bits added in, you can fab it on whatever is fastest/cheapest, using an older more clapped out node would make good sense, but nothing is stopping you from have some crisp as gently caress line edges on your .2mm copper traces by using 28 or 22nm if that's what you have spare capacity for.
|
# ? Aug 28, 2018 20:29 |
|
Cygni posted:The big question for me with Intel, Samsung, and now Glofo all struggling at (roughly) this same size process is... how is TSMC pulling it off? Or are their 7nm yields just terrible and we should expect inflated prices? Or maybe the perf/watt isn't as hyped? Interesting times. TSMC's 7nm process is roughly equivalent to Intel's 10nm and ultimately they are all measuring different aspects of the silicon to hit those numbers. The process size numbers are largely just branding at this point and all that will really matter are the actual characteristics of the product they build on it.
|
# ? Aug 28, 2018 20:49 |
|
Yeah, but TSMCs 7nm is roughly equal to Intels 10nm in a lot of gate size and density metrics, and Intel's 10nm may never see the light of day in a consumer product and TSMCs 7 is going in everything from cellphone SoCs to reticle limit GPUs (supposedly) to CPUs.
|
# ? Aug 28, 2018 20:55 |
|
All this talk about interposers seems wildly out of proportion to the number of products AMD is selling that actually have interposers. They are on TR and Epyc CPUs, which are halo & server products. High margin, low volume. Even if AMD does really well in that market, that's not a high number of units that need interposers. Then there's Vega, which has been completely uncompetitive in everything except crypto since the day it launched. Maybe they sold a lot of those last year but I can't imagine they're gonna sell many going forward. Navi is still a year (?) away, and it seems unlikely that it will be using HBM in the first go-round. AMD does not conceivably need enough interposers to make a dent in any remaining GloFlo agreements, and likely need so few that it's completely trivial to their overall business where they get made. I dunno whether GlobalFlounderies cancelling 7nm means AMD gets to tear up their agreements. If they're still on the hook for a bunch of wafers in 2019, they'll probably just pay off the penalty for most of them.
|
# ? Aug 28, 2018 22:25 |
|
Epyc and Threadripper don’t use interposers.
|
# ? Aug 30, 2018 13:03 |
|
They currently don't but there have been rumors the Zen2 based versions might. There have been LOTS of odd rumors (current rumor list I know of right now: possibly uses a interposer, possible 8 core CCX configuration, some redundant components moved off CPU die to reduce silicon usage + 7nm shrink allowing them to effectively nearly double core counts for the same amount of silicon, some sort of IF bus secret sauce improvements to improve power efficiency, etc) about Zen2 TR/Epyc and its really hard to give any of them serious consideration at this point since there isn't much solid information on any of it right now.
|
# ? Aug 30, 2018 19:49 |
|
I wish someone at AMD would explain just why it is that when multiple Zen CCXes are put together, they are laid out rectangular rather than square. You'd think that abutting the L3 caches up against each other would decrease inter-CCX latency even more.
|
# ? Aug 30, 2018 21:01 |
|
TheJeffers posted:Epyc and Threadripper don’t use interposers. I saw a lot of references to "organic interposer". Guess that's a different thing, but they're still made by chip fabs right? How else do you make all those tiny wires than photolithography? PC LOAD LETTER posted:its really hard to give any of them serious consideration at this point since there isn't much solid information on any of it right now. Also considering the rumours for zen+, versus the reality where it was a pure process switch and nothing inside the chip changed at all.
|
# ? Aug 30, 2018 21:07 |
|
Klyith posted:I saw a lot of references to "organic interposer". Guess that's a different thing, but they're still made by chip fabs right? How else do you make all those tiny wires than photolithography? It's litho on the fiberglass layers the CCXs are bedded to. You can get the fiberglass layers really thin these days, and litho on fiberglass masks is cheap as hell compared to having to use silicon for it. Incidentally litho is how all modern circuit boards are made once you get out of the 'CNC mill from copper coated stock for 1 off prototype' production levels.
|
# ? Aug 30, 2018 22:32 |
|
Methylethylaldehyde posted:Incidentally litho is how all modern circuit boards are made once you get out of the 'CNC mill from copper coated stock for 1 off prototype' production levels. Everyone uses CNC for the vias on small and large runs.
|
# ? Aug 30, 2018 23:53 |
|
Klyith posted:Also considering the rumours for zen+, versus the reality where it was a pure process switch and nothing inside the chip changed at all. There isn't anything obviously wrong with any of the Zen2 TR/Epyc rumors. They all just suggest drastic differences and improvements vs Zen/Zen+ some which are probably unlikely but not impossible or ridiculous off hand.
|
# ? Aug 31, 2018 00:12 |
|
Harik posted:I don't think you can use a CNC process for any relatively modern board, the feature-width is way too fine to try to mechanically scrape it off. All the one-day-turn PCB order places use litho, and if you're doing it in-house you've got the resources to have a litho process as well. I meant the 'I have access to a Makerspace with a little CNC router and Eagle, so I did one of those single layer boards with them'. Perfectly fine to make real low cost one-offs, especially when you're doing lots of little iterative test boards. For anything that's actually surface mount, especially surface mounting big 100+ pin chips, litho is the only way to go. I mean, technically you could also do litho for that one off, but the mess tends to get annoying to set up and do vs. just making the CNC mill cut the copper layer off as long as your feature sizes allow it.
|
# ? Aug 31, 2018 00:38 |
|
Methylethylaldehyde posted:I meant the 'I have access to a Makerspace with a little CNC router and Eagle, so I did one of those single layer boards with them'. Perfectly fine to make real low cost one-offs, especially when you're doing lots of little iterative test boards. For anything that's actually surface mount, especially surface mounting big 100+ pin chips, litho is the only way to go. Wait, what's the opposite of fun, again?
|
# ? Aug 31, 2018 01:11 |
|
Harik posted:I don't think you can use a CNC process for any relatively modern board, the feature-width is way too fine to try to mechanically scrape it off. All the one-day-turn PCB order places use litho, and if you're doing it in-house you've got the resources to have a litho process as well. I make parts for a commercial? hobbyist? Something in between? desktop milling machine that does .006" trace and space. On my own actual CNC machines, I can get thinner. SwissArmyDruid fucked around with this message at 13:07 on Aug 31, 2018 |
# ? Aug 31, 2018 02:29 |
|
Harik posted:You're getting borderline on those nowadays too, with the embedded controllers getting smaller and smaller. I get what you meant now, though. We turn a ton of prototypes but do them all at pcbexpress or similar because in-house just doesn't make a ton of sense. What's really fun is soldering 0201 components by hand or BGA with a heatgun. The word you're thinking of is 'frustrating'. And yeah, now with how absurdly tiny things are getting, litho is more or less the only way to mount it, unless you wanna spring for those little spring-pin chip holder jobbies. Edit: Or you have a really nice mill with tools that are both sharp AND spinning good and fast.
|
# ? Aug 31, 2018 06:21 |
|
The 2950X became available today and has apparently already sold out in some places? Heh.
|
# ? Aug 31, 2018 16:17 |
|
Combat Pretzel posted:The 2950X became available today and has apparently already sold out in some places? Heh. You aren't getting the 2990?
|
# ? Aug 31, 2018 16:24 |
|
Want one, price and memory bandwidth issues bug me. Gonna spend the money on an 2080 Ti as soon they come down in price, when the Pascal stock is gone. Also, babby's first water cooling kinked the budget. :[
|
# ? Aug 31, 2018 18:08 |
|
Combat Pretzel posted:Want one, price and memory bandwidth issues bug me. Gonna spend the money on an 2080 Ti as soon they come down in price, when the Pascal stock is gone. Also, babby's first water cooling kinked the budget. :[ It is quite a lot of money agreed, but it sure is a nice processor. Didn't get an AIO? They are not that expensive, mine was 120 Euro's or something. If I had know the NZXT Cam software for variable pumprate and cooling would require an account with them I would have bought something from a different vendor.
|
# ? Aug 31, 2018 20:31 |
|
Wanted cooling that covers the whole IHS, since that apparently makes a noticeable difference (and maybe helps PB2/PBO eke out a bit more), which only left the Enermax Liqtech as an option. Given the drama surrounding that cooler, I wanted to sidestep those issues. And I'm probably going to loop in the future graphics card, since better cooling can apparently also help the card boosting higher.
|
# ? Aug 31, 2018 20:39 |
|
I'm basically in the same boat, doing my first custom loop except with two gpus (not SLI, the other is my 1080ti) for a Windows gaming VM. Not sure where you are but the 2950x is still available on Amazon for shipping some time next week. I stayed up a bit late to ensure I could order it as soon as it was available. Wish I could have just pre-ordered it. E: misread, looks like you did manage to order one and you were just commenting on it being sold out. Desuwa fucked around with this message at 05:41 on Sep 3, 2018 |
# ? Aug 31, 2018 20:45 |
|
I'm in the EU and ordered one at CaseKing. The stock situation on this continent seems weird. CaseKing claim they get stock Wednesday. Amazon EU doesn't list it at all, some Amazon UK representative claims they get stock end of September to mid-October (per some reddit guy), Alternate.nl claims to have stock already, some other shop claims the 17th, etc.
|
# ? Aug 31, 2018 20:55 |
|
https://twitter.com/Ashraf__Eassa/status/1035511047514468352?s=20 https://twitter.com/Ashraf__Eassa/status/1035687084487581696?s=20 N7 HPC has about 10-13% clockspeed advantage over N7 SoC.
|
# ? Sep 1, 2018 02:25 |
|
EmpyreanFlux posted:https://twitter.com/Ashraf__Eassa/status/1035511047514468352?s=20 16c/32t 6 ghz 3700x on air confirmed.
|
# ? Sep 1, 2018 02:41 |
|
.
sincx fucked around with this message at 05:50 on Mar 23, 2021 |
# ? Sep 1, 2018 03:03 |
|
Confirm it all you want but no way AMD isn't charging at least $500-600 for 16c/32t @ >5 Ghz. Still would.
|
# ? Sep 1, 2018 05:49 |
|
Seamonster posted:Confirm it all you want but no way AMD isn't charging at least $500-600 for 16c/32t @ >5 Ghz. Oh if they can hit that while Intel is still stuck on 14nm, I fully expect the prices to be eye watering.
|
# ? Sep 1, 2018 05:51 |
|
Can't wait to deactivate half of my chiplets to play the next waifu simulator with peak ideal optimal supreme memory response time
|
# ? Sep 1, 2018 07:22 |
|
|
# ? May 23, 2024 13:29 |
|
AdoredTV put up a video about GlobalFoundry. Didn't know their 14nm node was licensed from Samsung, and besides that, they mainly just hosed up all the way. So now I'm not surprised that they cancelled 7nm.
|
# ? Sep 1, 2018 13:04 |