But the IO won't stay in 12nm forever, they'll eventually move to 7nm when it becomes mature enough, cheaper enough to make the move worth it, we might see it next year or even the next.
Unless, GloFo could find a way to package the HBM with it's interposer with other silicons with much lower cost to make it viable to think about it for APU's
You won't see that happen until their CPU chiplets move to a more advanced node, AND 7nm supply loosens up. The whole point is that the I/O die doesn't scale that well, so instead of wasting wafer space, use a cheaper process. As a result, 12LP+ on GF would be a good choice over the current 12LP, if costs are reasonable.
They won't have a choice though with their upcoming APU, since it's likely a monolithic design. But the I/O diespace on such a design will also be considerably less (not to mention other cuts to cache and so forth) so it's probably not a dealbreaker.
not everyone need bleeding edge. 12nm is still incredibly powerful and stuff. I expect a lot of mid range snapdragons or mediatek or those kinds of things in there.
Their LP+ apparently has pretty substantial gains over 12LP, plus 7nm is both expensive AND supply is tight. All the big players are already gobbling up all the wafers. Designing for 12LP+ is a pretty safe bet as far as costs and wafer supply go.
For other specialty designs they've also got 12FDX.
The 12LP+ FinFET process technology is a link in the chain -- or better yet, the incorporated interposers, dies, substrates, chiplets, cores, bumps, EBIBs and 'glue that binds all that stuff' together. GlobalFoundries (and SiFive) are comfortable enough to tap-out some chips in 2020 with 'edge computing looking to optimize for TOPS-per-milliwatt performance' for 'implementation primarily used for artificial intelligence training and inference'.
It ain't about 7nm, but it is about the design and IP that can bring new levels to HPC. Ian and Andrei have articles have articles about the challenges on the frontpage.
RISC-V 'U84 IP running internally on FPGA platforms' ? That sounds like high competition with Chipziilah's (and AMD's) hetro arch to me ...
remember to that the process is relative now... actual size has been decoupled from the process naming for some time.... 7nm is not almost 1/2 of 12nm.... 12nm is more competitive against 7nm than you might think. Either way... I'm glad to see Malta get the work... They had a big layoff l when they announced they weren't pursuing 7nm..... Really affected the area.... lot of good paying jobs were lost.
cheers to that, the workers need work, if this ensures more work (hopefully also as ecologically non wasteful as possible) the more the merrier...this leads to far more competition overall, which helps to push ever more GREATNESS to the industry (which is felt world wide) ^.^
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
13 Comments
Back to Article
FreckledTrout - Tuesday, November 5, 2019 - link
I wish them all the best. They will be competing with 7nm HBM2E but hey if they can make it cheap enough maybe they will get some takers.PixyMisa - Tuesday, November 5, 2019 - link
AMD's IO dies are produced on GloFo 12nm, so there's potential there.Alexvrb - Tuesday, November 5, 2019 - link
That's an interesting thought. An APU with 12LP+ I/O die and HBM2E, connected to 7nm GPU and CPU chiplets?With that being said, I think AMD's APUs are going to be monolithic with no HBM this time around.
ET - Wednesday, November 6, 2019 - link
AMD can opt for two APU lines, one monolithic for ultraportables and one chiplet based for desktop replacements.Alexvrb - Wednesday, November 6, 2019 - link
It's not worth the (considerable) resources required to create two different designs. We *might* see something like this combined with Zen 3 or later.Xajel - Wednesday, November 6, 2019 - link
But the IO won't stay in 12nm forever, they'll eventually move to 7nm when it becomes mature enough, cheaper enough to make the move worth it, we might see it next year or even the next.Unless, GloFo could find a way to package the HBM with it's interposer with other silicons with much lower cost to make it viable to think about it for APU's
Alexvrb - Wednesday, November 6, 2019 - link
You won't see that happen until their CPU chiplets move to a more advanced node, AND 7nm supply loosens up. The whole point is that the I/O die doesn't scale that well, so instead of wasting wafer space, use a cheaper process. As a result, 12LP+ on GF would be a good choice over the current 12LP, if costs are reasonable.They won't have a choice though with their upcoming APU, since it's likely a monolithic design. But the I/O diespace on such a design will also be considerably less (not to mention other cuts to cache and so forth) so it's probably not a dealbreaker.
deil - Wednesday, November 6, 2019 - link
not everyone need bleeding edge. 12nm is still incredibly powerful and stuff. I expect a lot of mid range snapdragons or mediatek or those kinds of things in there.Alexvrb - Tuesday, November 5, 2019 - link
Their LP+ apparently has pretty substantial gains over 12LP, plus 7nm is both expensive AND supply is tight. All the big players are already gobbling up all the wafers. Designing for 12LP+ is a pretty safe bet as far as costs and wafer supply go.For other specialty designs they've also got 12FDX.
Dragonstongue - Wednesday, November 6, 2019 - link
maybe it is, maybe it is not.just cause it is 7nm does not automatically mean it IS much more pricey, all depends on the foundry (the one who makes it)
if anything, either GloFo will target a more $/unit OR try to be as cost competitive as possible, rarely has it been shown to be BOTH.
I hedge on the latter over the former...time will tell...^.^
Smell This - Wednesday, November 6, 2019 - link
Mr Trout did not think this one out.The 12LP+ FinFET process technology is a link in the chain -- or better yet, the incorporated interposers, dies, substrates, chiplets, cores, bumps, EBIBs and 'glue that binds all that stuff' together. GlobalFoundries (and SiFive) are comfortable enough to tap-out some chips in 2020 with 'edge computing looking to optimize for TOPS-per-milliwatt performance' for 'implementation primarily used for artificial intelligence training and inference'.
It ain't about 7nm, but it is about the design and IP that can bring new levels to HPC. Ian and Andrei have articles have articles about the challenges on the frontpage.
RISC-V 'U84 IP running internally on FPGA platforms' ? That sounds like high competition with Chipziilah's (and AMD's) hetro arch to me ...
HardwareDufus - Tuesday, November 5, 2019 - link
remember to that the process is relative now... actual size has been decoupled from the process naming for some time.... 7nm is not almost 1/2 of 12nm.... 12nm is more competitive against 7nm than you might think. Either way... I'm glad to see Malta get the work... They had a big layoff l when they announced they weren't pursuing 7nm..... Really affected the area.... lot of good paying jobs were lost.Dragonstongue - Wednesday, November 6, 2019 - link
cheers to that, the workers need work, if this ensures more work (hopefully also as ecologically non wasteful as possible) the more the merrier...this leads to far more competition overall, which helps to push ever more GREATNESS to the industry (which is felt world wide)^.^