AMD Bristol/Stoney Ridge Thread

Page 37 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

NostaSeronx

Diamond Member
Sep 18, 2011
3,683
1,218
136
GlobalFoundries didn't cancel but because there is no successor new customers might drop. Customers who buy a lot in hopes for a continued safe roadmap no longer will buy.

CPU-side 14LPP/12LP @ GF would be switched to 12FF and 7FF @ TSMC.
GPU-side 14LPP/12LP @ GF would be switched to 12FF and 7FF @ TSMC
APU-side with no option will most likely get pulled to TSMC.
//
- advanced process nodes testchips (Global Foundries 14nm and TSMC 16nm)
- advanced FinFET CMOS process technologies on TSMC/GF Foundries

So, with no 7nm and the "differentiated options" in 14LPP/12LP might be used to increase the FinFET base cost. This will however just be the sort of reason to kill it off.

http://nattokude.gozaru.jp/inspiron11_3180/review.html
http://www.pasonisan.com/dell/inspiron11-2in1/3185-a6-9220e-a9-9420e.html

Looking at the Stoney Ridge. Then, add a switch to 22FDX w/ perf boost as promised then one half of it.
A9-9420e => 2.7 GHz (peak) * 1.5 => 4.05 GHz * 1.2 w/ (1v FBB) => 4.86 GHz with the same TDP.
A9-9220e => 2.4 GHz (peak) * 1.5 => 3.6 GHz * 1.2 w/ (1v FBB) => 4.32 GHz with the same TDP.
9420e w/ Half => 2.7 GHz (peak) * 1.25 => 3.375 GHz * 1.1 w/ (1v FBB) => 3.7125 GHz with the same TDP.
9220e w/ Half => 2.4 GHz (peak) * 1.25 => 3 GHz * 1.1 w/ (1v FBB) => 3.3 GHz with the same TDP.

I'm calculating from the 1V supply and body biasing performance increase from the FBB @ 1 Vdd + 1 Vbb.

Complex IP:
https://community.cadence.com/cadence_blogs_8/b/breakfast-bytes/posts/fd-soi-2
https://community.cadence.com/cfs-f...n-Shot-2016_2D00_04_2D00_20-at-2.57.56-PM.jpg
For future FT5? with;
DDR4 support up to 3600 MHz
LPDDR4 support up to 4267 MHz
By the way: Apr 27, 2016 is the date. When Cadence Blogs showed this. PCIE4.0 w/ CCIX is supported now. LPDDR5/DDR5 haven't shown themselves yet on 22FDX, but it is only a bit of time.

If Raven2 top SKU is at $55, then successor to Stoney and 22FDX replacement to Raven2 would be $44 for top SKU. With the 12FDX solution being $40 for top SKU. As the ASP drops, but the volume sold increases.

The lowest cost scenario is like Agena to Deneb. Same architecture with higher clocks with new uncore and same to lower power.
The slightly more expensive scenario is like Bobcat to Jaguar. This however would be cheaper to the 12FDX node. As the new architecture would be validated sooner than later. Less masks now, rather than more masks later.

https://www.semanticscholar.org/paper/Low-Cost-and-Highly-Manufacturable-MOL/BEOL-in-for-Jain-Kim/30cf974cc4734ccb0240e92bd165cbdb8c3f385f
22FDX DDB-12T width is 960nm
28BLK SDB-9T width is 900nm
22FDX CNRX-8T width is 640nm
(Relative to NR2)


===
Doing some simulations:
FX (CPU-type) => 130~140 mm squared
FX would have 8-cores via CMT modules; either in 4M(2C) or 2M(4C) methodology. It will most definitely have no L3 cache. Instead, it would probably adopt Jaguar's L2 and upgrade it to 4 MB; 8x 512 KB SRAMs. Do to the aim for EPI the 96KB L1i might be shrinked for lower size. The shared FPU might turn into private FPUs which would increase LSU/FPU density by 2x~3x. A much smaller core and CMT with more modern design.

Sempron FX/Sempron (APU-type) => 100~120 mm squared
These products would probably peak out at 4-cores via CMT modules. Also, either 4-core module or a two 2-core modules. Largely following the above CPU enhancements. While, also reducing the GPU from Bristol Ridge or enhancing the GPU from Stoney. These would probably use 2 MB L2, with 4x 512 KB SRAMs. With Small Super-SIMD CUs w/ small caches we could still see 8 CUs. This would probably point why we haven't seen 22FDX products yet. The GPU validation isn't complete and probably by the time it is complete DDR5/LPDDR5 will be out.

GlobalFoundries releases AMD from FinFETs, then GlobalFoundries shackles AMD with FDSOI. It is the only way for GlobalFoundries to keep big names at the foundry. I highly doubt AMD wouldn't run a processor that will get cheaper than 28nm by Q4 2018. Adaptive Voltage Frequency Biasing Scaling; AVFBS would probably be an extreme benefit to 25x20. Essentially, getting to it in 2019, rather by 2020.

New Ryzen/Athlon => Premium
New FX/Sempron => Budget
 
Last edited:

amd6502

Senior member
Apr 21, 2017
971
360
136
GlobalFoundries didn't cancel but because there is no successor new customers might drop. Customers who buy a lot in hopes for a continued safe roadmap no longer will buy.

CPU-side 14LPP/12LP @ GF would be switched to 12FF and 7FF @ TSMC.
GPU-side 14LPP/12LP @ GF would be switched to 12FF and 7FF @ TSMC
APU-side with no option will most likely get pulled to TSMC.

Why would GF abandon 12LP, especially just when it's in a golden age (relatively cheap, and a top performer)? With the supply agreement AMD will try to make as much volume on GF as they can, leaving the 7nm exclusively for premium for the next two years. 7nm is for high margin and maybe dollar sales can almost compare to 12nm dollar sales, but by transistors 12nm will dwarf 7nm indefinitely. 12nm is the new 28nm, it'll be around forever.

12nm should eventually be complemented by a 28nm successor for cheaper smaller APUs. If your frequencies are even close, ~4.5ghz, then even a dual thread Stoney would almost be useful. A merged puma+xv 4 thread CCX would have a much wider appeal and better perf/watt for mobile use; we all know how inefficient XV becomes when you go over 3.5ghz.
 

NostaSeronx

Diamond Member
Sep 18, 2011
3,683
1,218
136
GlobalFoundries would abandon 14LPP/12LP for the same reason 7LP was abandoned. No customers other than AMD, which of course is winding down:
"This will be in time for various devices for high-growth markets, but AMD will naturally wind down its 14LPP/12LP orders to GlobalFoundries over the 2019 – 2020 timeframe, reducing the company’s revenue and profits."
- https://www.anandtech.com/show/13277/globalfoundries-stops-all-7nm-development/2

Why go FinFETs at GlobalFoundries if their FDSOI is so much better? Higher capacity, higher performance and power efficiency, better SoC integration, better analog & digital asymmetricals, etc.
///
AMD winding down at GlobalFoundries' FinFET nodes means the WSA between AMD/GloFo will target the FDSOI nodes. Which are being aimed for high capacity plays... 22FDX will be deployed at all major 300mm foundries at GlobalFoundries.

GlobalFoundries is in talks with SOITEC to get 300mm Fab3 placed near Malta. The closest place for that would be in Marcy which is <100 miles away from Malta.

Fab8 will be FDSOI More than Moore. Monolithic 3D, Stacking, 450mm, etc.
Fab1 will be FDSOI More Moore. 12FDX, 7FDX, 22RF, 12RF, 22I(photonics), 12I, and possible HV/UHV on special 22FDX/12FDX wafers.

///
Architecturally, the design AMD is holding onto will be a continuation of 15h. While dropping majority of the HPC enhancements that plague energy(&area) efficiency on 15h. So, whatever it will be number 19h and beyond or 15h with newer family models. It will be more efficient than Jaguar and Zen, if they were ported over.
 
Last edited:

amd6502

Senior member
Apr 21, 2017
971
360
136
"This will be in time for various devices for high-growth markets, but AMD will naturally wind down its 14LPP/12LP orders to GlobalFoundries over the 2019 – 2020 timeframe, reducing the company’s revenue and profits."
- https://www.anandtech.com/show/13277/globalfoundries-stops-all-7nm-development/2

I'd think they would want a mix of fin fet and fdsoi. Whether Ian Cutress believes it or not, 12LP is not going away. I think the authors above are calling it pretty wrong. I'll wager that GF's 12nm finfet will be the workhorse for years to come, at least until there are breakthrough and the disadvantages of FDSOI can be resolved.
 

dark zero

Platinum Member
Jun 2, 2015
2,655
138
106
GlobalFoundries would abandon 14LPP/12LP for the same reason 7LP was abandoned. No customers other than AMD, which of course is winding down:
"This will be in time for various devices for high-growth markets, but AMD will naturally wind down its 14LPP/12LP orders to GlobalFoundries over the 2019 – 2020 timeframe, reducing the company’s revenue and profits."
- https://www.anandtech.com/show/13277/globalfoundries-stops-all-7nm-development/2

Why go FinFETs at GlobalFoundries if their FDSOI is so much better? Higher capacity, higher performance and power efficiency, better SoC integration, better analog & digital asymmetricals, etc.
///
AMD winding down at GlobalFoundries' FinFET nodes means the WSA between AMD/GloFo will target the FDSOI nodes. Which are being aimed for high capacity plays... 22FDX will be deployed at all major 300mm foundries at GlobalFoundries.

GlobalFoundries is in talks with SOITEC to get 300mm Fab3 placed near Malta. The closest place for that would be in Marcy which is <100 miles away from Malta.

Fab8 will be FDSOI More than Moore. Monolithic 3D, Stacking, 450mm, etc.
Fab1 will be FDSOI More Moore. 12FDX, 7FDX, 22RF, 12RF, 22I(photonics), 12I, and possible HV/UHV on special 22FDX/12FDX wafers.

///
Architecturally, the design AMD is holding onto will be a continuation of 15h. While dropping majority of the HPC enhancements that plague energy(&area) efficiency on 15h. So, whatever it will be number 19h and beyond or 15h with newer family models. It will be more efficient than Jaguar and Zen, if they were ported over.
Same reason they won't ever release 22nm: no customers. Without AMD GloFo is nothing.

AMD already killed other uArchs in favor to Ryzen. Even K12 is dead.


AMD is at 16 nm at worst now.

If GloFo leaves 12nm, they are done for.
 

ET

Senior member
Oct 12, 1999
521
33
91
I'd think they would want a mix of fin fet and fdsoi. Whether Ian Cutress believes it or not, 12LP is not going away. I think the authors above are calling it pretty wrong. I'll wager that GF's 12nm finfet will be the workhorse for years to come, at least until there are breakthrough and the disadvantages of FDSOI can be resolved.

12nm will be mainstream for AMD throughout 2019 and probably into 2020, in the form of Picasso (which I assume it 12nm). The question is, who will use 12nm after AMD releases a 7nm APU?
 
  • Like
Reactions: amd6502

NostaSeronx

Diamond Member
Sep 18, 2011
3,683
1,218
136
Same reason they won't ever release 22nm: no customers. Without AMD GloFo is nothing.
22FDX so far is fine without AMD.
>7 product tapeouts in July/August.
>10 more product tapeouts by December.

http://www.game234.com/yxrd/20180730/3526727.html
=> http://09.imgmini.eastday.com/mobile/20180730/20180730122655_8b346b5403dbd748cd52098df09b06bb_7.jpeg

AMD's actual mainstream will most likely be 22FDX+. At best another Steamroller(GF28SHP) -> Excavator(GF28A). Since, 22FDX-UHP was supposed to drop in on top of 28HPA(Bristol Ridge/Stoney Ridge).

GlobalFoundries relationship with AMD deprioritize FinFET products to open up FDSOI products.
 
Last edited:

NostaSeronx

Diamond Member
Sep 18, 2011
3,683
1,218
136
BD and its derivatives are dead, long live Zen.
Any AMD FDX x86 core will naturally be derived from a Bulldozer core. Either in thought as a new architecture with CMT or in actual design that is based on an extension of Excavator.

Cluster-based Multithreading is just that power efficient.

Praire Falcon => 2016
Prairie Falcon Refresh 1&2 => 2017-2018
Stoney Ridge => 2016
Stoney Ridge Refresh 1&2 => 2017-2018

There is significant demand to continue Stoney Ridge but maintaining cost over performance. Which is heavily orientated towards 22FDX with reduced total process costs.
 
Last edited:
  • Like
Reactions: amd6502

NostaSeronx

Diamond Member
Sep 18, 2011
3,683
1,218
136
Which is why AMD ditched it for Zen :rolleyes:
Where is Zen's small APU? :rolleyes:
It seems Zen was thrown in a ditch dug by an Excavator. :fearscream:

Then there is 22FDX announced in 2015, but known by AMD since 2012. Two cores one SMT and another CMT made together with only one being released first.

If there is no 7LP Zen2. Then, it has to be the companion core on 22FDX which takes the spotlight at GlobalFoundries from now on. GlobalFoundries won't allow depreciated products in the WSA. AMD will need to move on to something more productive. Like 22FDX which goes up to <120,000 wafers per month for 22FDX at Chengdu(AMD-HMC) and Dresden(AMD).

The prolonged support of Stoney Ridge and Prairie Falcon is indicative that AMD's Zen is not cost-sensitive enough. That or AMD has a product on 22FDX which will ramp up in 2019 replacing Stoney Ridge/Prairie Falcon. Which the successor to the two would need to provide a lower complete cost.

Banded Kestrel is 20% more expensive than a hypothetical 22FDX device. River Hawk is 35% more expensive than a hypothetical 12FDX device. This however is ignoring end costs to consumers. So, the BK/RH devices will be more expensive to consumers. While a 22FDX/12FDX device will be less expensive from Stoney Ridge & Prairie Falcon. Minimizing the end-cost to optimize to move from CMT to SMT is also worthwhile in such a device, so CMT stays. With the 22FDX shrink, there is several optimization to target.

Total device shrink of 0.6x: 125 mm squared(SR) to 75 mm squared(HypoDev). Or, have the 22FDX device consume the total markets Stoney Ridge up to Bristol Ridge. Allowing salvaged dies to have even lower costs at the end user.
 
Last edited:
  • Like
Reactions: amd6502

dark zero

Platinum Member
Jun 2, 2015
2,655
138
106
Where is Zen's small APU? :rolleyes:
It seems Zen was thrown in a ditch dug by an Excavator. :fearscream:

Then there is 22FDX announced in 2015, but known by AMD since 2012. Two cores one SMT and another CMT made together with only one being released first.

If there is no 7LP Zen2. Then, it has to be the companion core on 22FDX which takes the spotlight at GlobalFoundries from now on. GlobalFoundries won't allow depreciated products in the WSA. AMD will need to move on to something more productive. Like 22FDX which goes up to <120,000 wafers per month for 22FDX at Chengdu(AMD-HMC) and Dresden(AMD).

The prolonged support of Stoney Ridge and Prairie Falcon is indicative that AMD's Zen is not cost-sensitive enough. That or AMD has a product on 22FDX which will ramp up in 2019 replacing Stoney Ridge/Prairie Falcon. Which the successor to the two would need to provide a lower complete cost.

Banded Kestrel is 20% more expensive than a hypothetical 22FDX device. River Hawk is 35% more expensive than a hypothetical 12FDX device. This however is ignoring end costs to consumers. So, the BK/RH devices will be more expensive to consumers. While a 22FDX/12FDX device will be less expensive from Stoney Ridge & Prairie Falcon. Minimizing the end-cost to optimize to move from CMT to SMT is also worthwhile in such a device, so CMT stays. With the 22FDX shrink, there is several optimization to target.

Total device shrink of 0.6x: 125 mm squared(SR) to 75 mm squared(HypoDev). Or, have the 22FDX device consume the total markets Stoney Ridge up to Bristol Ridge. Allowing salvaged dies to have even lower costs at the end user.
They have and is the Smach Z and if not Subor will take it.

Even Mediatek is moving to 12nm
 

amd6502

Senior member
Apr 21, 2017
971
360
136
Where is Zen's small APU? :rolleyes:
It seems Zen was thrown in a ditch dug by an Excavator. :fearscream:
Everything is very hush about this so far; it seems top secret; zero leaks to disclose the low end strategy. Yet I think we'll find out really soon, by H1 2019 or earlier. I'd like to not rule out one last hurrah for XV, but imho odds are still a good bit higher that it's a 12nm Zen+ or Zen2 APU. The ~20% savings mentioned below just seems not quite significant enough for the tradeoffs.

If there is no 7LP Zen2. Then, it has to be the companion core on 22FDX which takes the spotlight at GlobalFoundries from now on. GlobalFoundries won't allow depreciated products in the WSA.[....] The prolonged support of Stoney Ridge and Prairie Falcon is indicative that AMD's Zen is not cost-sensitive enough. That or AMD has a product on 22FDX which will ramp up in 2019 replacing Stoney Ridge/Prairie Falcon.[....] Banded Kestrel is 20% more expensive than a hypothetical 22FDX device. River Hawk is 35% more expensive than a hypothetical 12FDX device.

Above, you must mean 12nm Zen2 not 7nm, because the latter is a certainty. I wouldn't have thought before the GF news broke, but now I think 12nm Zen2 looks likely.

It's hard to believe the bit about WSA not allowing depreciated products, unless I'm misinterpreting what you mean by depreciated products.

I do think there is a ramp up soon of something to supplement and replace Stoney, with the likeliest guess the above mentioned dual core Zen APU.

Lastly, I'm wondering if the team B that is working on next gen architecture is in any way (maybe as secondary project) busy with XV derived designs. Maybe this is totally unrealistic, but I could imagine side projects being decent academic exercises when there's a good upside potential.
 

NostaSeronx

Diamond Member
Sep 18, 2011
3,683
1,218
136
Everything is very hush about this so far; it seems top secret; zero leaks to disclose the low end strategy.
GlobalFoundries is super aggressive on 22FDX. The ideal scenario is AMD tests it, samples it out, and starts production within four months. The aggressive MPW schedule of 22FDX was GloFo increasing DPML(days per mask layer) efficiency.

http://www.coolchips.org/2018/?page_id=10#keynote1 <= AMD at Cool Chips
http://www.coolchips.org/2019/ <== April 2019

AMD would need to finalize the core around the Zen2 design. Zen2 as far as I know has been design complete. So, the time frame for a companion core wouldn't be far behind with 22FDX. AMD has yet to show absolute Zen2 changes from Zen either. All of it is pretty hush, hush at the architecture level.
Revive the market of Desna, Hondo, Temash, Mullins.
Zen2 on N7 at HiPerf and --- on 22FDx at ULP.
It's hard to believe the bit about WSA not allowing depreciated products, unless I'm misinterpreting what you mean by depreciated products.
The WSA has been always orientated at future products or products to be released. So, a seventh amendment will ensure 22FDX(and successors) supply for AMD. As that is the only profit stream left to exploit at GlobalFoundries. Malta isn't getting any more FinFET, but Dresden and Chengdu are doing leading-edge FDSOI. Converting Malta into a FDSOI fab in SOITEC's 3rd phase of FDSOI ramp up(3 million wafers per year) wouldn't be expensive.
Just look how 200GE obliterates A12-9800 in almost every metric.
If based on Stoney Ridge shrink:
- 1.8W w/ a9-9400e or a9-9420e
- 6W w/ a9-9420 or a9-9425

If on a new core based on next-gen ULP IPs. Then, there is some fuzzy math.

Return to tablets, revitalize netbooks/AIOs/etc, ∞D Freedom Fabric for revival of UHD servers.
 
Last edited:
  • Like
Reactions: amd6502

amd6502

Senior member
Apr 21, 2017
971
360
136
GlobalFoundries is super aggressive on 22FDX. The ideal scenario is AMD tests it, samples it out, and starts production within four months. The aggressive MPW schedule of 22FDX was GloFo increasing DPML(days per mask layer) efficiency.

http://www.coolchips.org/2018/?page_id=10#keynote1 <= AMD at Cool Chips
http://www.coolchips.org/2019/ <== April 2019

AMD would need to finalize the core around the Zen2 design. Zen2 as far as I know has been design complete. So, the time frame for a companion core wouldn't be far behind with 22FDX. AMD has yet to show absolute Zen2 changes from Zen either. All of it is pretty hush, hush at the architecture level.
Revive the market of Desna, Hondo, Temash, Mullins.
Zen2 on N7 at HiPerf and --- on 22FDx at ULP.
The WSA has been always orientated at future products or products to be released. So, a seventh amendment will ensure 22FDX(and successors) supply for AMD. As that is the only profit stream left to exploit at GlobalFoundries. Malta isn't getting any more FinFET, but Dresden and Chengdu are doing leading-edge FDSOI. Converting Malta into a FDSOI fab in SOITEC's 3rd phase of FDSOI ramp up(3 million wafers per year) wouldn't be expensive.
If based on Stoney Ridge shrink:
- 1.8W w/ a9-9400e or a9-9420e
- 6W w/ a9-9420 or a9-9425

If on a new core based on next-gen ULP IPs. Then, there is some fuzzy math.

Return to tablets, revitalize netbooks/AIOs/etc, ∞D Freedom Fabric for revival of UHD servers.

From 6W to 2W seems a bit of a stretch, but I can sort of buy 15W to something close enough to 6W. I think a9-9425 actually clocks high enough to be useful with the limited threads it has.
 

NostaSeronx

Diamond Member
Sep 18, 2011
3,683
1,218
136
From 6W to 2W seems a bit of a stretch, but I can sort of buy 15W to something close enough to 6W. I think a9-9425 actually clocks high enough to be useful with the limited threads it has.
The issue with the whole port of Stoney Ridge is why wait? So, it is an option that isn't really an option.

22FDX started in Q1 2017 with it increasing to 500K from SOITEC in Q3 2017. Then, there is the advanced IP's PCIe4.0(16 Gb/s-25 Gb/s) and potentially LPDDR5/DDR5 via DFI5.0. Which for PCIe 4 is finished the DDR4/DDR5+LP haven't finished yet.

The same for Stoney Ridge would also apply to Mullins.
A10 Micro-6700T => 1.2 GHz and 2.2 GHz w/ 500 MHz GPU
28B -> 28FD = 30% perf boost at -100 mV
28FD -> 22FD = 50% perf boost at -100 mV
The Mullins product is more ideal in the Vdd department as well:
https://hothardware.com/reviews/amd...instream-and-lowpower-2014-apus-tested?page=3
0.925V for 2.2 GHz => 80% perf would push that up to 3.9 GHz at 0.725V
0.575V for 1.6 GHz => 2.8 GHz around 0.5V, etc.
0.250V for 1.0 GHz

Mullins 4.5W/2014 was succeeded by the Stoney Ridge 6W/2016&2018. So, Puma doesn't cut at the port down.

There is only two choices with 22FDX; Excavator and another architecture. The Excavator would be the lowest budget, but technically we would have seen it earlier than later. Excavator can have a port down but Llano soured the pool and Bobcat 28nm was canned. So, that leaves a new architecture the only choice for that market.

https://www.pcworld.com/article/2084883/amd-prototype-pc-could-be-mistaken-for-an-envelope.html
A return to this, but better than last time. Much like how EPYC did for Servers.

If Verisilicon can do it, why can't AMD?
https://i.imgur.com/I7HnywV.jpg
When it is actually this: https://electrek.co/2017/05/22/tesla-nvidia-supercomputer-self-driving-autopilot/
https://en.wikipedia.org/wiki/Drive_PX-series

Maintaining Excavator IPC/work per cycle, but having a new architecture.
Would lead two outcomes:
1. Higher Frequency @ same power.
2. Lower power @ same frequency.
While work per cycle stays the same, work per second can grow.

So applying "Power Efficiency of Zen" slide particularly Pure Power and Physical Design. The added components would be Pure Power Plus for Body Biasing(AVFBS), and FDSOI. The removed components would be FinFETs and potentially architecture.

Add Excavator being optimized for 15W TDP(Putting it all together: Excavator Performance) and Zen is for 30W-45W(inferred). The return to a sub-10 watt core is ready.
 
Last edited:
  • Like
Reactions: amd6502

NTMBK

Lifer
Nov 14, 2011
10,208
4,940
136
Raven Ridge is already in thin and light laptops. The 7nm successor will be even more power efficient. There isn't a great deal of room for an even lower power SoC, x86 tablets aren't exactly setting the world on fire. And the Bulldozer line isn't the best starting point if you want efficiency.
 

NostaSeronx

Diamond Member
Sep 18, 2011
3,683
1,218
136
Raven Ridge is already in thin and light laptops. The 7nm successor will be even more power efficient. There isn't a great deal of room for an even lower power SoC, x86 tablets aren't exactly setting the world on fire. And the Bulldozer line isn't the best starting point if you want efficiency.
The Bulldozer line is the best starting point for efficiency:
1. Cluster-based Multithreading has higher efficiency in transistor count and energy per workcycle.
2. AMD only has one existing core in the sub-10W zone and that is Excavator.

GPD, Chuwi, Jumper
Most of theses are with Intel's 4W-10W products. Apt to get replaced by a product on a cheaper process that can lower the cost of the overall machine. While being fabricated in China beginning at earliest July 2019.
*holding back throwup*Kickstarter, Indiegogo, Ogadget... and other places

Now would a end-user/consumer get a better perceived/objective quality from a high-cost 7nm solution over a low-cost 22nm. That is the question to be answered and its heavily implied that it will be 22FDX over FinFETs.

Looking at Intel then AMD:
Amber Lake => $393 at top SKU for 3.5W to 7W.
Gemini Lake => $161 at top SKU for 6W/10W
Athlon 200GE => $55 at that SKU for 35W.
A10-9620P/9630P => around $45 for 15W/35W
A9-9420 => around $20 for 15W-25W.

125 mm squared to 250 mm squared of 28nm is lower cost for AMD. Moving to 22FDX with its crossover to being cheaper than 28nm by Q1 2019. Would make the 22FDX lower cost while potentially getting higher margin within that low cost.

22FDX 125 mm squared design can be sold at $15 and still get a profit.
22FDX 150-180 mm squared design can be sold at $30 and still get a profit.
200GE is most likely not profitable, but AMD gets the revenue from the $169/$99 Ryzen APUs, and the EPYC >$400 per die.

Increase profits margins from FDSOI or continue low-margin on FinFETs. Which wouldn't make sense compared to Amber Lake and AMD wanting to increase ASPs on their flagship core.

There is also the roadmap of FDSOI is low-cost orientated:
22FDX -> 22FDX+ -> 12FDX -> 7FDX -> onwards are aimed at maintaining to increasing manufacturing production and decreasing cost.

So, 22FDX in 2019, with 12FDX in 2021, with 7FDX in 2023 is optimized for total(Substrate, Device, Masks, Production) cost erosion. So the FDX products that will continue off from Stoney Rigdge/Bristol Ridge, etc. Will cost less to produce for AMD and having increased production efficiency than the FinFET nodes for GloFo.
 
Last edited:

amd6502

Senior member
Apr 21, 2017
971
360
136
Athlon 200GE => $55 at that SKU for 35W.
A10-9620P/9630P => around $45 for 15W/35W
A9-9420 => around $20 for 15W-25W.

[....]

22FDX 125 mm squared design can be sold at $15 and still get a profit.
22FDX 150-180 mm squared design can be sold at $30 and still get a profit.

So after the Stoney 12nm native 2c/4t successor is released (Picasso-L) we can expect some die salvage for a 1c/2t sub 5W tdp soc. It probably will keep up with Stoney at about half the wattage. If RR is break even around $55 then Picasso-L ~$35.

The other alternative is a small core dozer branch like Nosta suggests (similar to what Atom is to core). Added advantage would be less licensing costs as it's also got native memory controller. The disadvantage would be missing volume production of 2c/4t to replace Bristol ridge type budget mainstream market, for both mobile as well as desktop.
 

NostaSeronx

Diamond Member
Sep 18, 2011
3,683
1,218
136
Xbox One X -> Jaguar core is ~2 mm squared
Playstation 4 Pro -> Jaguar core is ~1.79 mm squared
A 7nm Jaguar-esque core would be <1 mm squared, this from 28nm Jaguar at ~3 mm squared.
Generally, however Zen is replacing Jaguar. With the Zen lineage starting at GlobalFoundries, but returning to the home ground of the Cat lineage -> TSMC.

Bulldozer through Excavator has always been at GlobalFoundries. So, GlobalFoundries would want an inflexible core that can't go anywhere else.

22FDX:
- Planar and 75% the same as 28nm Bulk. Faster development from paper.
- Interconnect/(Infinity Fabric) can use LVT/RVT w/ FBB to reduce power of such an interconnect.
- Physical Interfaces (PCIe/DDR/HDMI/DP/Multi-protocol) have significantly reduced leakage and area consumption. I/O VTs can use FBB/RBB as well.
Transform Dozer to Atomlike?
Sorry but there is Puma for that purpouse.
It isn't so much Atomlike, more that Steamroller and Excavator maintained or expanded on features at HPC from BD/PD.
- FMA in 15h are performance aimed not density aimed. There are more efficient, denser FMACs.
- Integer(+LSU) component in 15h is from 90s/00s. There are more efficient, denser ALU/AGU(+LSU) configurations.
- Front-end is huge in 15h. Thus, there is more efficienct, denser BP/IF/ID configs.
- There are several variations of CMT-implements. There be better cluster-based multithreading architecture implementations out there.

PDVS & FPCA:
- ALU -3 PMOS headers – V DDH , V DDM , V DDL
- Everything else - V DDH
then,
- Reconfigurable architecture
-- OpenRISC
-- Split into front end (FE) & processing element (PE)
- Reconfigurable interconnect
- Variable width SIMD
- Switch between SIMD and MIMD
AMD RD => Steve Kosonocky // SuperFLEX FPU and sub-threshold optimized voltage scaling.
PDVS => Panoptic dynamic voltage scaling
http://rlpvlsi.ece.virginia.edu/panoptic-dynamic-voltage-scaling-pdvs-0
Combined with: https://www.semanticscholar.org/pap...raig/045ac8fb81e4c38e2980e9583e9617e4f7086075
FPCA => Field programmable core array
www.cs.virginia.edu/~skadron/Papers/arrabi_FCCM_2014_poster.pdf

I also recommend looking at VectorBlox's Lightweight Vector Extensions to RISC-V.
--
On a more on topic note. A third refresh?
https://browser.geekbench.com/v4/cpu/compare/9787835?baseline=8745503
https://browser.geekbench.com/v4/cpu/compare/9798527?baseline=8787800

I wonder if the lower performance might be 4.5W cTDP. If the switch from 8 GB to 4 GB didn't cause any problems.

Beema A8-6410 is 15W to Carrizo-L A6-7310 is 12W-15W. So.. the equivalence is..
A9-9420e with 6W to A6-9220C with 4.5W-6W?
A6-9220e with 6W to A4-9120C also with 4.5W-6W?
If only there was a E2-9020Z, then the naming change would make sense sort of. As it would be a reference to Desna/Hondo(Z-series) and Ontario(C-series).

If this is the third refresh, then you must acquit. :coldsweat:
 
Last edited:

NTMBK

Lifer
Nov 14, 2011
10,208
4,940
136
The Bulldozer line is the best starting point for efficiency:
1. Cluster-based Multithreading has higher efficiency in transistor count and energy per workcycle.

What are you basing this on? The only implementation of CMT that we have seen sucked in the real world. Yes, I'm sure there are some old academic papers arguing that CMT would be wonderful, but in actual implementations it was an inefficient mess. AMD ditched it for a reason!
 

Abwx

Lifer
Apr 2, 2011
10,854
3,298
136
If we are to extrapolate some numbers based on Techspot review the Athlon 200GE, aka R3 2200U, allow for 1.6-1.7GHz@5W and 1.2-1.3GHz@3W, dunno why they would need to shrink XV given the cost and that a better solution, perfs and cost wise, is readily available...