AMD Bristol/Stoney Ridge Thread

Page 51 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.
Apr 21, 2017
448
53
61
I think it took almost 200W for Piledriver to hit 5GHz (well, maybe 100W for the boosting core and the rest for the uncore and other 7 cores). Even with the advancements, I don't think this is doable at ultra low power.

Following the news today on the new R1000 (and V1000 series), it's starting to look like there will be an upcoming native dual core APU die (either now, or about 6 months down the road), based on Picasso. This should be Zen+ ~100mm2 group of products, optimized to address the ~10W market (from 6W to 15W).

At 2c/4t it should be a little more capable than BR at CPU tasks, but have only well under half the GPU perf of BR. Should have almost 3B transistors (vs 1.2B for Stoney and 3.1B for BR) and so it should still cost significantly more to produce than Stoney yet not too significantly more than BR (which is why Stoney remains in production while BR probably does not).

Lower production costs to the point of recovering development costs and product premiums would be needed to justify an FDX alternative positioned below Picasso-L. I think a 4 thread Stoney++ on 22FDX could still possibly do that, but the odds of seeing this in the future appear to be low.
 
Last edited:
Sep 18, 2011
2,297
95
126
I think it took almost 200W for Piledriver to hit 5GHz (well, maybe 100W for the boosting core and the rest for the uncore and other 7 cores). Even with the advancements, I don't think this is doable at ultra low power.
The voltage scales across the SOI nodes. It is also easier for FDSOI to get really low delay and thus really high clocks compared to PDSOI/FinFET.

32nm PDSOI <1.6V for 5 GHz -> 28nm HP-FDSOI <1.4V for 5 GHz -> 22nm FDX <1.2V for 5 GHz -> 12nm FDX <1.0V for 5 GHz. The FDSOI nodes have FBB which can be used to increase speeds at lower voltages. 28nm HP-FDSOI <1.3V for <5 GHz w/ FBB -> 22nm FDX <1.1V for <5 GHz w/ Conventional FBB -> 12nm FDX <0.85V for <5 GHz w/ DITO FBB.

Scaling wise, with increased energy efficiency pulled across from Zen's logic and memory compilers, etc. What is 220Ws in 32nm PDSOI with energy optimization would be 15W~35W in 12FDX. Absolutes is 22FDX on power is 1/5th 32PD, and 12FDX in turn can be 1/2th of that. With all the optimizations on bulk(SR/XV) and some agnostic(Zen). Eight core with no UMC(universal memory controller; MCT/DCT/PHY) and only some instances of infinity links. Easy, efficient, and cheap alternative to Zen, in CPU chiplet cases. While, pushing SKUs in likeness of Bristol/Stoney into extremely efficient budget parts.

Under the same math, since 28nm bulk on the low power is similar to PDSOI. So, a9-9420e(9220c)/a6-9220e(9120c) is 1.0V is 2.7/2.4 GHz, 0.9~1.0V is 2.0 GHz. Which would place a 12FDX product at 0.4V for 2 GHz and with FBB w/o power increases to ~2.5 GHz. At that voltage a 6W Stoney Ridge should scale to 0.6W~1.2W. This product does however need some stuff from Zen;
- Excavator's stock voltage range is 0.7V to 1.35V. This needs to adopt Zen's 0.4V to 1.0V stock voltage range.
- Excavator has a less accurate AVFS system and it might be slower at switching than Zen's. This needs to not only adopt it but modify it as well for faster freq/volt/bias switching and to add ABS for AVFBS. Both of which are FDSOI exclusive, body-bias and the shorter time needed for going between power states, ex: Px to Py.
 
Last edited:
Sep 18, 2011
2,297
95
126
Not sure if it matters but the cost to buy/bid for all the new stuff is near these values;
Ryzen 3200U is $100
Athlon 300U is $80
A6-9220C is $30
A4-9120C is $25

//Misc 12FDX stuff;
Lib cells from 28nm to 12FDX should be around 0.38x smaller?, still pending.
6T SRAMs are on-par with Intel; 14LPP/12LP SRAM is 0.0806 mm2 to 12FDX/Intel's 14nm is ~0.058mm2 which is a 0.72x shrink?
 
Last edited:
Apr 21, 2017
448
53
61
Not sure if it matters but the cost to buy/bid for all the new stuff is near these values;
Ryzen 3200U is $100
Athlon 300U is $80
A6-9220C is $30
A4-9120C is $25

//Misc 12FDX stuff;
Lib cells from 28nm to 12FDX should be around 0.38x smaller?, still pending.
6T SRAMs are on-par with Intel; 14LPP/12LP SRAM is 0.0806 mm2 to 12FDX/Intel's 14nm is ~0.058mm2 which is a 0.72x shrink?
Seems like too small a premium between Stoney top bin (A6) and medium (A4) bin; the boost difference is 12%, 300Mhz, 2.4 vs 2.7ghz.

That's a good cost advantage. So Stoney around $30. I'd guess Picasso-L around $50-$60.

Compared to Pentirum they are not doing too bad at the moment, just because Stoney can go down in price so much (being just ~100mm2 at what is now a very old node) while Pentirums right now have a shortage of 14nm capacity. The main deficit is really the number of threads:

http://www.cpu-world.com/Compare_CP...,Intel_FH8066802979803,Intel_FH8066802980002/

I think pentirum may also have low load and idle efficiency advantages, just because the atom cores can clock down so low (somewhat below 600mhz if I remember right from using something like an n2820).

Both number of threads and low p-state wattage is something that I think a next gen jaguar with on-demand coarse grain multi-threading could help with. Probably it won't happen, and the low end x86 mobile market will ride the dual thread as long as it's still acceptable.

It would make sense if Stoney would in fact get company from 6W single core Zen (1c/2t) Picasso-L die salvage sometime in 2020; then Stoney production would end within half a year to a year of that, and all dual thread x86 consumer SoC would cease to be produced early 2022 and fade into history that year.

The other approach, of which there are some hints, is to revive Puma (maybe on FDX). At 28nm it was about 4 cores in under 15mm2. So octacore in 30mm2. Salvaged quads can be funneled into 10W and below market, while hexacores and 8c go to the other markets. This was some leak by tum_apisak who has sprung proper leaks before. Search "A9-9820" and there is a reddit thread + various articles like https://www.guru3d.com/news-story/a...820-reveal-themselves-in-3dmark-database.html

So RX models go to consoles, while A9 models go to things like all-in-ones and even laptops, and maybe even two or three AM4 packaged skus. The GPU might be comparable to almost to an R7 250, maxing out DDR4 2133 or 2400 dual channel, being something like 5 or 6 CU at 1300MHz. The really bad single thread and small die area of ~200mm2 at 28nm (or less at 22nm) would make this family strictly all budget.
 
Last edited:

Similar threads



ASK THE COMMUNITY

TRENDING THREADS