- Joined
- May 13, 2008
- Messages
- 669 (0.11/day)
| System Name | HTPC whhaaaat? |
|---|---|
| Processor | 2600k @ 4500mhz |
| Motherboard | Asus Maximus IV gene-z gen3 |
| Cooling | Noctua NH-C14 |
| Memory | Gskill Ripjaw 2x4gb |
| Video Card(s) | EVGA 1080 FTW @ 2037/11016 |
| Storage | 2x512GB MX100/1x Agility 3 128gb ssds, Seagate 3TB HDD |
| Display(s) | Vizio P 65'' 4k tv |
| Case | Lian Li pc-c50b |
| Audio Device(s) | Denon 3311 |
| Power Supply | Corsair 620HX |
First of all, yes. This is a few months old. If there is a thread about it already, I apologize.
I was catching up on the news/mill, as I try to do every month or so (to stay with the current gossip), and caught wind of an old article by Geoff from TR (whom knows we love some good process-prowess arguments on this here internet) linking to this. While I kinda keep up to date on SRAM cell size for future processes (because backseat/quack futurist is how I roll), I hadn't thought to do the calculations in this paper to show a progressive pattern (and vicariously extrapolate from it). Excellent work imo, because correlation is generally food for my compulsion.
I considered posting this in GN, but thought some may be interested in it here if you haven't seen it, as it of course pertains to the future of GPUs as much as anything else. If someone feels compelled to move it elsewhere, by all means do so. I am more-so interested what others may gleam from this, or thoughts on what you feel it may mean for future GPUs.
I personally find that last graph incredibly fascinating, but I perhaps have more math geek (and recollection of former processes and their scaling) in me than others.
What I key into from that is a few things, and yes, I understand it's not perfectly indicative of final products. What's interesting is how well everything lines up.
Firstly, the density at 20nm from GF is absurd (for planar). I understood it was greater than those than came before it, but not to that extent. While that may take it's toll in clock speed/voltage allowance (28nm from GF was a good ~10-15% off from tsmc, this may increase that deficit vs 20SOC at tsmc by up to 5% or so) perhaps equalizing out to similar as tsmc 28nm in that regard, the density advantage is now essentially double that difference at the same node.
In other words, 20nm from TSMC may grant around 15-20 (or so)% in clock improvements.
CPA will have around 30% better density.
Or to put it another way, CPA will have a process that performs like tsmc 28nm but absurdly small xtors comparably.
As we move to 14nm, you can deduce that GF makes up that clock improvement (it's surely not exactly linear, but shown by the 22% increase). While not listed on the graph, 16nmff+ from tsmc will essentially make up the difference in density (around 10%), with that final trend continuing for both companies going into 10nm.
Now, if I were a betting man, I would bet the difference between 14nmELP (essentially risk production) and 14nmLPP (mature) nodes is Samsung/GF making up whatever performance difference there is with tsmc 16ff+ (not related to size, but speed/power).
Essentially, in the end, we have convergence. I find that kind of amazing. What's more amazing is that over-all, TSMC seems to be adapting to the CPA, rather than vice-versa. When you take into account the clock speed differences versus density, and their general roles in power consumption, the CPA does appear to have the philosophy advantage (depending upon limits of a design)...but it makes one wonder if tsmc hit a performance wall and are course-correcting.
All-in-all, I guess what I was/am trying to figure out was what AMD moving their lot to GF would accomplish (other than cost/allocation), and in the end came away with a different pov than which I started, and perhaps others will as well. 20nm products should have an abundance of units (or be very small) size running at lower clock (voltage?) for great power efficiency all things considered, where-as nvidia seems bent on speed. Where-as before I believed tsmc would always have that noteworthy speed advantage (however slight, which could be a significant proportional [to amd] boon to nvidia's current arch), and the CPA a small, if note-worthy, size advantage (seemingly fitting to amd's philosophy of more cores/units), I am no longer sure that will be the case. Where-as before I saw tsmc as the Apple of fabs (perhaps way too expensive for something slightly better), I now see them as equals, if not the CPA having a wiser trajectory.
I guess we'll find out starting sometime early next year, with the real showdown towards the end (if not 2016).
Thoughts? What do you take from this...and what do you think it means for future gpus given their architectural differences? Does this change your thinking on AMD perhaps doing a gpu run on 20nm? Do you think each company is making/made the right choice (if AMD transitions to 14LPP for everything quickly after 20nm, and nvidia stays tsmc) or whom benefits more from their choice all factors considered? Is it simply a wash, and if so do you find that strange considering the variables?
I was catching up on the news/mill, as I try to do every month or so (to stay with the current gossip), and caught wind of an old article by Geoff from TR (whom knows we love some good process-prowess arguments on this here internet) linking to this. While I kinda keep up to date on SRAM cell size for future processes (because backseat/quack futurist is how I roll), I hadn't thought to do the calculations in this paper to show a progressive pattern (and vicariously extrapolate from it). Excellent work imo, because correlation is generally food for my compulsion.
I considered posting this in GN, but thought some may be interested in it here if you haven't seen it, as it of course pertains to the future of GPUs as much as anything else. If someone feels compelled to move it elsewhere, by all means do so. I am more-so interested what others may gleam from this, or thoughts on what you feel it may mean for future GPUs.
I personally find that last graph incredibly fascinating, but I perhaps have more math geek (and recollection of former processes and their scaling) in me than others.
What I key into from that is a few things, and yes, I understand it's not perfectly indicative of final products. What's interesting is how well everything lines up.
Firstly, the density at 20nm from GF is absurd (for planar). I understood it was greater than those than came before it, but not to that extent. While that may take it's toll in clock speed/voltage allowance (28nm from GF was a good ~10-15% off from tsmc, this may increase that deficit vs 20SOC at tsmc by up to 5% or so) perhaps equalizing out to similar as tsmc 28nm in that regard, the density advantage is now essentially double that difference at the same node.
In other words, 20nm from TSMC may grant around 15-20 (or so)% in clock improvements.
CPA will have around 30% better density.
Or to put it another way, CPA will have a process that performs like tsmc 28nm but absurdly small xtors comparably.
As we move to 14nm, you can deduce that GF makes up that clock improvement (it's surely not exactly linear, but shown by the 22% increase). While not listed on the graph, 16nmff+ from tsmc will essentially make up the difference in density (around 10%), with that final trend continuing for both companies going into 10nm.
Now, if I were a betting man, I would bet the difference between 14nmELP (essentially risk production) and 14nmLPP (mature) nodes is Samsung/GF making up whatever performance difference there is with tsmc 16ff+ (not related to size, but speed/power).
Essentially, in the end, we have convergence. I find that kind of amazing. What's more amazing is that over-all, TSMC seems to be adapting to the CPA, rather than vice-versa. When you take into account the clock speed differences versus density, and their general roles in power consumption, the CPA does appear to have the philosophy advantage (depending upon limits of a design)...but it makes one wonder if tsmc hit a performance wall and are course-correcting.
All-in-all, I guess what I was/am trying to figure out was what AMD moving their lot to GF would accomplish (other than cost/allocation), and in the end came away with a different pov than which I started, and perhaps others will as well. 20nm products should have an abundance of units (or be very small) size running at lower clock (voltage?) for great power efficiency all things considered, where-as nvidia seems bent on speed. Where-as before I believed tsmc would always have that noteworthy speed advantage (however slight, which could be a significant proportional [to amd] boon to nvidia's current arch), and the CPA a small, if note-worthy, size advantage (seemingly fitting to amd's philosophy of more cores/units), I am no longer sure that will be the case. Where-as before I saw tsmc as the Apple of fabs (perhaps way too expensive for something slightly better), I now see them as equals, if not the CPA having a wiser trajectory.
I guess we'll find out starting sometime early next year, with the real showdown towards the end (if not 2016).
Thoughts? What do you take from this...and what do you think it means for future gpus given their architectural differences? Does this change your thinking on AMD perhaps doing a gpu run on 20nm? Do you think each company is making/made the right choice (if AMD transitions to 14LPP for everything quickly after 20nm, and nvidia stays tsmc) or whom benefits more from their choice all factors considered? Is it simply a wash, and if so do you find that strange considering the variables?
Last edited: