AMD: Speculation, Rumors, and Discussion (Archive)

Status
Not open for further replies.
On the topic of GFX IP level, I am curious whether increased FP16 vs FP32 throughput can sneak under the same version number, or if that is different enough to be visible to the driver.
That would require new instructions, so it would be visible.
That said, if you'd simply use previously reserved bits for encoding those and all existing instructions still work all the same, you could probably pretend there weren't any changes (only when you need the new instructions then you'd actually need to distinguish). But why would you do that, you could just say it's gfx ip level 8.1 or so in that case, and noone would have any clue what changed neither...
 
That would require new instructions, so it would be visible.
That said, if you'd simply use previously reserved bits for encoding those and all existing instructions still work all the same, you could probably pretend there weren't any changes (only when you need the new instructions then you'd actually need to distinguish). But why would you do that, you could just say it's gfx ip level 8.1 or so in that case, and noone would have any clue what changed neither...

The policy behind the version numbers apparently has changed with 8.0 versus the 7.x generation, although it's not clear from the outside how much change is enough to warrant a new number.
The alleged perf/W benefit of dropping down to 16-bit math might explain some of the math for Vega's improvement, and I thought I saw an earlier mention about a possible 8.1.
 
Damn, I hope Polaris isn't delayed....!

- http://www.cnet.com/news/short-Amd-cancels-Samsung-Global-Foundries-contracts/
- Amd cancels 14nm manufacturing deal with Samsung and Global - EETimes
- http://www.tomshardware.com/news/AMD-cancels-contract-Samsung-Global-Foundries,31486.html

daily-photos-16-a-humiliated-statue-paris.jpg
 
Last edited:
Lol, you magnificient bastard! :LOL:

I consider myself only HALF falling for it tho, because I've got video autoplay blocker enabled. ;)
 
I'm pretty sure the Polaris lineup is actually graphic IP v9.0 already, at least internally naming. Doesn't mean though there's any major difference on the publicly visible interface compared to VI, so the drivers may not differentiate much.

Which means that unknown graphic IP v9.0, 4096 shader SoC could as well have been intended as just another GDDR5 powered chip for the Polaris lineup. Just as the resume says, biggest chip of the v9.0 series. Possible that this chip won't be part of the final lineup, as it would somewhat be collide performance wise with Vega 10, but at a much higher power consumption.
 
I don't really have a source but I don't expect 14nm to ramp faster than 28nm did and the time between the first 28nm part (Tahiti) and the first really large chip (GK110) was about a year later. 14nm GPUs will likely launch mid 2016 and a refresh line up with a larger die flag ship can launch in 2017. It makes too much sense given how it worked out like this for 28nm and it will probably be a long time until 10nm node comes online.
20nm initially ramped quite a bit faster than 28nm... based on how long it took to reached 20% of the overall wafer revenue. (Edit- TSMC initially projected 20nm to get to 20% of OWR in the same length of time as 28nm. The numbers they actually used included 16nm with 20nm... so they fibbed a little to make it look good.) No idea where I got the original info but I remember comparing the numbers and timeframe. This article references the same if my memory isn't a good enough source, it probably shouldn't be anymore.
14nm should ramp even faster than both, for multiple reasons.

Fabs are bypassing EUV for 10nm... Samsung and TSMC are both projecting on shipping 10nm before the end of 2016.
So late 2017 would be best case and sometime 2018 for GPUs being released would be the most probable, IMO.

2nd Edit- Found my original post about TSMC 20nm being 20% of TWR in Q4... my post was from Sept '14. This is *likely* the original article that I got those numbers from.

Too lazy to do the math but 16/14nm should still ramp faster than both 28nm and 20nm.
 
Last edited:
20nm initially ramped quite a bit faster than 28nm... based on how long it took to reached 20% of the overall wafer revenue. (Edit- TSMC initially projected 20nm to get to 20% of OWR in the same length of time as 28nm. The numbers they actually used included 16nm with 20nm... so they fibbed a little to make it look good.) No idea where I got the original info but I remember comparing the numbers and timeframe. This article references the same if my memory isn't a good enough source, it probably shouldn't be anymore.
14nm should ramp even faster than both, for multiple reasons.

Fabs are bypassing EUV for 10nm... Samsung and TSMC are both projecting on shipping 10nm before the end of 2016.
So late 2017 would be best case and sometime 2018 for GPUs being released would be the most probable, IMO.

2nd Edit- Found my original post about TSMC 20nm being 20% of TWR in Q4... my post was from Sept '14. This is *likely* the original article that I got those numbers from.

Too lazy to do the math but 16/14nm should still ramp faster than both 28nm and 20nm.
Well the ramp should be fast then which is likely why Vega is coming out very soon after Polaris. It is good to see my prediction from about a year ago playing out pretty much exactly as I said. I guess Apple paying for most of that 20nm production really pushed TSMC's profits through for those early 20nm months. 14nm seems to have been very fast ramp at least on the ARM side of things. GPUs have been a bit slower than I originally thought tho.

If 10nm is really coming as fast as Samsung and TSMC are projecting then it would line up with Navi. I don't feel confident about 10nm being ready for GPUs in 2018 tho. Even Intel is slowing down to get to 10nm. If TSMC and Samsung can deliver ARM socs by 2017 then they would be hitting at the same time as Intel's first 10nm commercial parts. The gap would truly be closing on Intel if that were the case. Heres hoping there won't be too much delays.
 
If 10nm is really coming as fast as Samsung and TSMC are projecting then it would line up with Navi. I don't feel confident about 10nm being ready for GPUs in 2018 tho. Even Intel is slowing down to get to 10nm. If TSMC and Samsung can deliver ARM socs by 2017 then they would be hitting at the same time as Intel's first 10nm commercial parts. The gap would truly be closing on Intel if that were the case. Heres hoping there won't be too much delays.

True, they could be talking about any sort of 10nm chips... early test SDRAM? Shuttles? Risk? Actual start of mass production?
silentguy would probably have better insight than me on translating their marketing speak on "10nm by the end of '16".

I don't really follow Intel's nodes too closely. Is their 10nm actual close to what real 10nm should be or did they change their nm designations like TSMC and Samsung?

TSMC/Samsung's 10nm will be an actual node shrink from 20nm, in regard to gate pitch. So what real 14nm should have been before marketing people jumped on the nm bandwagon and decided that moving to FinFets deserves a smaller node designation.

Edit- And you are right to second guess my speculation on when we will see GPUs on a new node. I have historically been quite optimistic in that regard... usually ~6months to a year too early. Though to be fair, I took that into consideration this time and went a bit conservative.
 
Last edited:
This is two weeks old news, I guess they forgot to go through all the leaks from that german forum. oh look there's GREENLAND!!!!11!!

https://forum.beyond3d.com/threads/...ors-and-discussion.56719/page-48#post-1901907

https://www.google.co.in/search?q=https://compubench.com/device.jsp?benchmark=compu15d&os=Windows&api=cl&D=AMD+67FF%3AC8&testgroup=info&ie=utf-8&oe=utf-8&gws_rd=cr&ei=uvAHV-efMc6cugS-y5qIDw

If Polaris 11/Baffin only has upto 1280 cores, it's peculiar that zauba shipping manifests show it at 60% of Fury prices. (48k to 80k INR)
I think people put too much thought to Zauba pricing - they're prototypes, they can be priced at whatever
 
Status
Not open for further replies.
Back
Top