• Guest, The rules for the P & N subforum have been updated to prohibit "ad hominem" or personal attacks against other posters. See the full details in the post "Politics and News Rules & Guidelines."

AMD Bristol/Stoney Ridge Thread

Page 60 - 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
3,320
853
136
AMD doesn't have the wafer capacity to waste time with crap like a successor to AM1 or anything else in the bargain bin. Which sucks for people that actually like the bargain bin but, what can you do?
AMD has $1.6B worth of wafer capacity with GlobalFoundries over 2022-2024.

The purpose of FDSOI is to reduce average selling price. The target of CMT is for the low-power SMT aspect w/ linear power draw from CMT aspect. Rather than get 4 ALU SMT2 core at 2.3 GHz, it is better to get a (2x)2 ALU core(s) @ 3.3+ GHz, etc.

Given overall improvements a reworked x86 CMT core towards client-embedded. It is more likely AMD to choose FDSOI to cut off Multi-Vt/Leakage optimization process steps/masks that are required for FinFETs.
 
Last edited:

Insert_Nickname

Diamond Member
May 6, 2012
4,276
820
126
If you are a fan of firesale sub-$200 cheapo systems that AMD used to have in the past . . . you're out of luck! Practically nobody is making that now.
Good riddance I say. I still shudder at the thought of those, especially when equipped 5400RPM drives. Brazos was decent for running a low power XP system, or other x86 version of Windows/Linux. Don't try to run anything x64 on it. Unless you're masochistic. Jaguar/Puma was not that much better, but could at least handle x64.

Gotta say, the reality distortion field is strong in this thread.
 

krumme

Diamond Member
Oct 9, 2009
5,922
1,541
136
Good riddance I say. I still shudder at the thought of those, especially when equipped 5400RPM drives. Brazos was decent for running a low power XP system, or other x86 version of Windows/Linux. Don't try to run anything x64 on it. Unless you're masochistic. Jaguar/Puma was not that much better, but could at least handle x64.

Gotta say, the reality distortion field is strong in this thread.
Bobcat and jaguar were cost balanced processors for a 5400/7200 hd in a notebook at that time. Damn fine processors vs Atom. So much Intel dumped billions into the atom market to make it adapt Atom. With no effect. Insane decision.

Today we lack small, lean, cheap and fast processors for the lowend segment. At least the painfull time of mechanical hd is over.

What about the new small cores from Intel. What size is those?
 

NostaSeronx

Diamond Member
Sep 18, 2011
3,320
853
136
Jaguar/Puma was not that much better, but could at least handle x64.
Mind you, there is a bit of little cores we never saw; Leopard, Catamount, Margay, Zen-L, etc.

14.48 mm2 = 2c Excavator - [28nm] 114CP/90Mx/9T
12.4 mm2 = 4c Jaguar - [28nm] 114CP/90Mx/9T
8.15 mm2 = 4c Jaguar - [16nm] 90CP/64Mx/9T
11 mm2 = 2c/4t Zen - [14nm] 78CPP/64Mx/10.5T

This is where it is an Ultra-Dense Quad-core CMT Module and 12FDX as an option. 8.15 mm2 => 4 cores inefficiently not-sharing units and area. Don't forget to include 12FDX shrinking from 90CP/64Mx/9T to 84CP/56Mx/7.5T, etc.
What about the new small cores from Intel. What size is those?
Tremont = 1c ~0.841 mm2 / 4c ~3.58 mm2
JPL being 63.8(CPU+GPU+?)+45.6(Chipset) ~~ 109.4 mm2

Gracemont = 1c ~1.63 mm2 / 4c ~6.5 mm2
No N-products to compare yet.
 
Last edited:

NTMBK

Diamond Member
Nov 14, 2011
9,467
3,005
136
Bobcat and jaguar were cost balanced processors for a 5400/7200 hd in a notebook at that time. Damn fine processors vs Atom. So much Intel dumped billions into the atom market to make it adapt Atom. With no effect. Insane decision.

Today we lack small, lean, cheap and fast processors for the lowend segment. At least the painfull time of mechanical hd is over.

What about the new small cores from Intel. What size is those?
The dual-core Zen die is really not very big.



CPU cores just aren't that big any more. The big server Zen chips use considerably more die area on the big L3 cache than they do on the actual CPU core. You can fit a "good" CPU in a cheap chip without much difficulty.

I think it will get interesting with Zen 4. AMD is making a special dense variant, Zen 4c- it has a smaller die area, in return for lower peak frequencies. Sounds kind of perfect for these sort of applications. Lower peak clock, smaller L3, and I bet you could fit a quad core cluster on a nice small chip.
 
  • Like
Reactions: Insert_Nickname

Insert_Nickname

Diamond Member
May 6, 2012
4,276
820
126
Bobcat and jaguar were cost balanced processors for a 5400/7200 hd in a notebook at that time. Damn fine processors vs Atom. So much Intel dumped billions into the atom market to make it adapt Atom. With no effect. Insane decision.
That's not really saying much. Slightly faster then slow-as-molasses is still dog slow. I will say they had a very nice IGP for their time. Again, unlike Brazos they could actually handle an x64 OS. Brazos dirty little secret was x64 instructions where (are, I guess) hacked in half and executed over two passes, which effectively halved frequency. So you'd have f.x. a 1600MHz Brazos operating at an effective 800MHz when running x64. Painful, just painful.

Today we lack small, lean, cheap and fast processors for the lowend segment. At least the painfull time of mechanical hd is over.
The faster that budget segment goes away, the better.

We already have Raven2/Dali/Pollock for low-end. I do not see any reason to go even lower in performance. If Intel ever gets around to releasing Gracemont as a standalone SoC, that Skylake-class performance should be the absolute minimum tolerated today.
 

NostaSeronx

Diamond Member
Sep 18, 2011
3,320
853
136
The faster that budget segment goes away, the better.
In today's market AMD's AM1 segment would be total-cost competitive with Qualcomm's QM215, Rockchip's RK3566, Broadcom's BCM2711.

Athlon 5350 then was $55
Unsoldered A9-9425 recently as of 2019 was $25

AM1 then was $40
Save that cost.

Recent prices
Single DIMM 8 GB DDR4-3200 = ~$30
Single DIMM 16 GB DDR4-3200 = ~$45
Two DIMM 2x8 GB DDR4-3200 = ~$33-45

$95(8GB)/$110(16GB) with recent prices and carried over prices...

Of course you are assuming the budget market is for us high GDP chums...


When it was meant for low GDP serfs...
600 million people versus 6 billion people, economies of scale, etc.

QM215:
RK3566:
BCM2711:

AMD only has to outrun the bottom of the barrel, not the top of the barrel. Why buy those over something that can at least run Windows and has legacy app access.
 
Last edited:

LightningZ71

Golden Member
Mar 10, 2017
1,005
982
136
This last part is what lot of people don't get. A LOT of the world is sensitive to a cost difference of pennies. Ultra cheap products move off the shelves there like hotcakes.

This is why I have been pushing for a half renoir port to a lower cost process at GF. A 4c ccx, 4CU Vega, single/dual (socket option) channel DDR4. It would get the job done at the bottom end like no other.
 

moinmoin

Platinum Member
Jun 1, 2017
2,897
3,875
136
If it's quantity you ask for there is a certain competitor which is in a much better place to deliver it.
 

Insert_Nickname

Diamond Member
May 6, 2012
4,276
820
126
This last part is what lot of people don't get. A LOT of the world is sensitive to a cost difference of pennies. Ultra cheap products move off the shelves there like hotcakes.
We do get it. So much in fact, that we realise most of those parts of the world will be buying mobile or cheap smart phones, because those are far cheaper, and more immediately useful for people there then a PC.

Some of what we in the 1st world consider 3rd world countries have better cellphone networks then we have actually. I was surprised when travelling in f.x. Kenya that they're actually further ahead then we are in Denmark in some areas. Now who is the developing country here? Hell, they even have more reliable rail service then we do. Of course, that's a pretty low bar to clear...
 

LightningZ71

Golden Member
Mar 10, 2017
1,005
982
136
A lot of those countries have phone and rail networks that were put in decades after the rest of the world got them, using the lessons learned in the first world. Also, while the backhaul bandwidth on those networks may not be massive when you leave the cities, most of the towers are lightly used, so it's not oversaturated. Backhaul in the cities is usually modern high speed networks that were laid relatively recently. In most places, all it takes is nice payoffs to a friendly government to get your "right of way" approved, and there aren't 5000 groups of NIMBYs fighting you every step of the way.
 

NostaSeronx

Diamond Member
Sep 18, 2011
3,320
853
136
This is why I have been pushing for a half renoir port to a lower cost process at GF. A 4c ccx, 4CU Vega, single/dual (socket option) channel DDR4. It would get the job done at the bottom end like no other.
I don't think that will be smaller than Raven2. The die has to be small while having 28nm~20nm level of development costs.

New CPU architecture
New GPU architecture
New systems architecture

Which needs to funneled through a lower costing design development cycle and smaller die to return investment margins.

For example APU+Mobo for Brazos ~~"AMD had told us to expect Brazos boards to start at “somewhere under $100.”~~
For example APU+Mobo for Bhavani ~~"Motherboard and APU solution starting at $60 est"~~

It clearly needs to follow the above for APU.

CPU feature needs ~135 mm2 die-size maximum like Seattle and be able to compete against for example something like Phytium's D2000 2.3-2.6 GHz 8-core, 25W, 132 mm2 die, 128-bit DDR4, 4x8 PCIe 3.0, etc. As well as being cheaper, a lot cheaper than it.
APU feature needs ~125 mm2 die-size maximum like Stoney and and potentially scale both dual-core with higher frequency and potentially quad-core at same frequency at 50%~75% low-power.
GPU feature needs ~125 mm2 die-size maximum like Topaz and needs to do Compute AI/ML from CDNA and Gaming from RDNA.
The above needs to be able to semi-custom so if a customer licenses thru-chip interface, they can stair ladder the CPU dies and get 512 GB/s between dies. Or make a combined CPU and GPU package like Varnheim/Varnhelm/Veinheim/(Whatever was that Zeppelin+Vega64 combo) but for low power.

With later designs down the road doing 3D rather than shrinking nodes. Thus, keeping costs down more aggressively by staying away from 7nm shrink.
3dcpu.png
^-- from Thermal Analysis of a 3D Stacked High-Performance Commercial Microprocessor using Face-to-Face Wafer Bonding Technology, 2020.
 
Last edited:

LightningZ71

Golden Member
Mar 10, 2017
1,005
982
136
I was trading off the development costs against the porting costs. The CCX for Renoir is a finished design. The VEGA iGPU is a modular design by nature. There's no need to include the 16x PCIe lanes for the GPU, just retain the pair of x4 groups and shrink the I/O section accordingly. One Zen2 4 core CCX, 3-4 VEGA CUs, a few USB ports, a pair of SATA ports and a pair of four lane PCIe 3 complexes. Use the high density, LP library options for GF12+ to keep from having to spend to convert away from a FF process. It'll certainly be larger than Raven2/Dali, probably smaller than Raven Ridge due to the increased density of GF12+. That would handle the bottom of the market for a long time to come.
 

amd6502

Senior member
Apr 21, 2017
878
307
136
AMD doesn't have the wafer capacity to waste time with crap like a successor to AM1 or anything else in the bargain bin.

Not if they tapped GloFo. But I understand why Lisa would refuse to do much business with them after what their management did to screw over IBM/Power.
 

burninatortech4

Senior member
Jan 29, 2014
393
124
116
Bobcat and jaguar were cost balanced processors for a 5400/7200 hd in a notebook at that time. Damn fine processors vs Atom. So much Intel dumped billions into the atom market to make it adapt Atom. With no effect. Insane decision.

Today we lack small, lean, cheap and fast processors for the lowend segment. At least the painfull time of mechanical hd is over.

What about the new small cores from Intel. What size is those?
I don't think this is true. Dali 2C/4T is very fast w/ SSD at 6 watts. 148.55mm² on 14nm can't be very expensive (IdeaPad 1 routinely is <$200). That's why I don't understand this continued discussion about Stoney. There are current products in this segment that fulfill the same TDP target while performing so much better.

I'm still curious if we'll see a Zen 3 "small die" on 12nm. There was some discussion on the AMD subreddit about this a few weeks ago (based on a VideoCardz rumor so massive grain of salt).

This thread should probably be re-named to "small die rumor and speculation". I don't think it's worth discussing Stoney Ridge in 2021-2022 (but that's just me).
 
Last edited:

NostaSeronx

Diamond Member
Sep 18, 2011
3,320
853
136
148.55mm² on 14nm can't be very expensive (IdeaPad 1 routinely is <$200).
It is more expensive than 28nm 107mm2 and 28nm 125mm2.

Cost per 0.1B gates increased with 14nm.
Cost per mm2 increased with 14nm.
Cost of design increased with 14nm.

However, with 22FDX/12FDX;
Cost per 0.1B gates is inline with decreasing.
Cost per mm2 is inline with decreasing.
Cost of design is inline with decreasing.

Also: IdeaPad 1 (11”, AMD) is DISCONTINUED. 14w Gen2 vs Gen1 has Pollock using lower speed DDR4, and lower resolution 768p. So, activating premium features is in favor of an A6-9220C refresh instead.
Full AIO/Laptop Intel competitor against 14w and Ideapad 1 series is the 500w Gen3/ALLDOCUBE N5100 laptops at around ~400 USD in comparison.
I was trading off the development costs against the porting costs. The CCX for Renoir is a finished design. The VEGA iGPU is a modular design by nature. There's no need to include the 16x PCIe lanes for the GPU, just retain the pair of x4 groups and shrink the I/O section accordingly. One Zen2 4 core CCX, 3-4 VEGA CUs, a few USB ports, a pair of SATA ports and a pair of four lane PCIe 3 complexes. Use the high density, LP library options for GF12+ to keep from having to spend to convert away from a FF process. It'll certainly be larger than Raven2/Dali, probably smaller than Raven Ridge due to the increased density of GF12+. That would handle the bottom of the market for a long time to come.
The problem is still FinFETs are expensive, backports are more expensive than shrinks, etc.

It is a better case to succeed Stoney with something just as cheap as it, if not cheaper with lower mm2.
There is more profits for a new design from fresh strictly built to maximize all the strong points of FDSOI, than doing an old-new design on FinFETs.

Basically, everything post-Stoney must cut costs down the line. From AMD to OEM to Consumers to make the cheapest product allowing for 1080p plus displays, 8 GB+ rams, NVMe SSD bga or m.2, etc.

If we use IBS numbers:
150 mm2 14nm FinFET = ~$15 per die... so $25 selling point ~~ just $10 margin.
125 mm2 28nm Bulk = ~$4 per die... $21 margin.
sub-125 mm2 22nm/12nm FDSOI = ~$3-$2 per die, etc.

The benefit of FDSOI is the potential absorbing of motherboards features: FIVR on-die, Audio codec on-die, etc. So, not only is it cheaper for AMD to make, it comes out cheaper for Motherboards as well.

GlobalFoundries is running 14nm FinFET at a loss, unless they want to make profits there is going to be a price hike. However, 22FDX is running at a profit, it is unlikely to get a price hike. With 14nm FinFET's price hike, it also comes with the decrease of volume of bulk FinFET wafers.

As Malta needs to get rid of its loss by adopting;
RH90 FDSOI(Skywater/Mil agreement)
RF45e/45SPCLO FDSOI(Shift IBM process to Malta).

"GlobalWafers Co., Ltd. (GWC) today announced a $800 million agreement with GLOBALFOUNDRIES® (GF®), the global leader in feature-rich semiconductor manufacturing, to add 300mm silicon-on-insulator (SOI) wafer manufacturing and expand existing 200mm SOI wafer production at GWC’s MEMC facility in O’Fallon, Missouri.

In particular, the 300mm wafers made at GWC’s MEMC site in Missouri will be used at GF’s most advanced manufacturing facility, Fab 8 in Malta, New York."

For example SOITEC's 300mm FDSOI pilot-line in Singapore was only $40 mil. With further Bernin II FDSOI investment not exceeding $180 mil.

GWC buying SunEdision's SOI part includes the license and patent agreement with SOITEC.

GlobalFoundries 6G Plan does not appear to include FinFETs:
"GLOBALFOUNDRIES (GF) is taking steps to establish a leadership position in 6G by collaborating with top researchers at leading universities to leverage the unparalleled benefits of its FD-SOI, RF-SOI and SiGe platforms"

So any semi-custom with an integrated 6G modem on die would need to fit in one of those three.

So, GlobalFoundries selling 14nm FinFET at a loss is bad for GFS-stock.
AMD selling fat-core 14nm FinFET with higher cost and reduced margin gap is bad.
5.5 mm2 for one Zen core.
~2.0 mm2 for one Jaguar core.

The two prior WSA agreements actually return to 1st WSA where AMD is required to support certain nodes. With the most recent WSA actually getting rid of AMD's FinFET commitment, hence no 12LP+ at AMD. Thus, allowing IOD's to move to TSMC 16nm or Samsung 14nm instead. However, AMD's commitment to FDX is still present.

Switching 14nm FinFET to something profitable say 22FDX or 12FDX since there is a source of FDSOI from Missouri.
Lower price for GlobalFoundries, carries over to lower price on the wafer given a smaller die. Which has cost distributed over many more dies.

CMT + 12FDX might allow one to fit four cores into the area of one Zen core to two Zen cores on 14nm
5.5 -> ~2.0 -> 1.375 or 2.7 per core. ==> Power-efficiency or IPC.
11 mm2 -> ~8.1 -> 5.5 or 11 per module. => Many cores in same area or same core to thread.

This scales across to CMT + 22FDX as well. Basically, two cores to two threads in the same area.
14.48 mm2 2c Excavator
6.2 mm2 2c Jaguar
AMD would need to use CMT on 6.2 mm2 2c and the shrink of 114CPP/90Mx/9T to 104CP/80Mx/8T(50% logic and 10% SRAM scaling) to get two cores within 5.5 mm2.
 
Last edited:

Shivansps

Diamond Member
Sep 11, 2013
3,400
1,004
136
Bobcat and jaguar were cost balanced processors for a 5400/7200 hd in a notebook at that time. Damn fine processors vs Atom. So much Intel dumped billions into the atom market to make it adapt Atom. With no effect. Insane decision.

Today we lack small, lean, cheap and fast processors for the lowend segment. At least the painfull time of mechanical hd is over.

What about the new small cores from Intel. What size is those?
Nah they way over did it with the segmentation on Bobcat and Jaguar. There were some C50 14"/15" that were e-waste from day 1, and that wasnt the worst cpu. AMD should have never released single cores or low cloked dual cores Bobcats, the same applies for dual cores and low clocked small cores never than that.

That was the problem, AM1, specially the 5150 and 5350 werent THAT bad, but the Sempron 2650 and 3850 were just terrible.

And in notebooks, well anything similar to the E1-6010 should have NEVER EXISTED... i hate the E1-6010, Gigabyte and Biostar dumped what they had left in my country and i was forced to deal with it, you cant even run Windows 10 with a SSD on it! And they are still coming.
The good thing is that the Biostar board (A68N-2100K) has a pcie slot and works very well with a pcie multiplier to support 4 gpus, so they mostly go to miners.

And they really killed IGP perf on Bobcat, Jaguar and anything newer by limiting them to single channel. But, it still took Intel until BAY TRAIL to match the IGP perf of Bobcat on a Atom.

@krumme

There are some Tremont Pentiums out there. Good luck getting them in a system you'd actually want.
at least in desktop, no, but there are some gemini lake around, Asus, Asrock and Biostar have Celeron J4005 and J4105 boards. The J4105 is actually decent for a cheap pc/HTPC as it is very low power and it has H265 8b and 10b as well as VP9 hardware support. Hell, it even has Quick Sync support.
 
Last edited:

DrMrLordX

Lifer
Apr 27, 2000
18,167
7,072
136
at least in desktop, no

Yup the sale link works:


Or one from Shuttle:


Looks like these units are mostly going to be popular overseas. I haven't found the AiOs with Jasper Lake in them yet.

edit: ugh apologies for the double post.
 
Last edited:

NostaSeronx

Diamond Member
Sep 18, 2011
3,320
853
136
Global Foundires is, allegedly, already fully-booked. And the less time AMD has to waste on their non-performant nodes, the better.
The $1.6B WSA is included in the fully-booked. AMD is literally the Number 1 customer, basically using every logic node GlobalFoundries pumped out.

#1 AMD/#2 Qualcomm in 2019 = 28%/8%
#1 AMD/#2 Qualcomm in 2020 = 21%/11%
and in 2021 = sub-20%(predicted)/15%(predictied)

As AMD EOL products on 28nm and 14nm/12nm nodes they lose total revenue share.

---
In the custom CPU/GPU market rather than licensed core, there is PowerPI from Libre-SoC to lead by example:
Tachyum Prodigy-esque CPU+GPU route.

JH7100 = 28nm HPC+ from TSMC
JH7110 = Quad-core RISC-V/BXE PowerVR
Raspberry Pi 5 = "Nonetheless, Upton states that the Pi 5 would likely have a faster SoC than the Pi 4 Model B, along with better USB I/O, greater than 8 GB of RAM and improved Ethernet/Wi-Fi connectivity."

Gunning for 22FDX/12FDX makes more sense on the shift from leading-edge pushing(everything is high-cost/high-risk) to trailing-edge pulling(everything is low-cost/low-risk).

Throw this here...
lowtdp.jpeg
 
Last edited:

VirtualLarry

No Lifer
Aug 25, 2001
52,238
7,054
126
Athlon Silver 3050e is actually a very compelling chip for <$200 netbooks. It's everything that Stoney wished it was and couldn't be.
Hells yeah! I've got a 3050e in my daily driver Lenovo, and it's all that I could want in a "basic" laptop. It's even capable of 4K screen res, using VSR, on a 1080P native IPS LCD screen. Shame that the laptop only has 4GB soldered RAM, and the Lenovo pre-installed Win10 (now Win11) image, has a commit charge above 4GB even just booting it up. :(
 
  • Like
Reactions: Insert_Nickname

Insert_Nickname

Diamond Member
May 6, 2012
4,276
820
126
Nah they way over did it with the segmentation on Bobcat and Jaguar. There were some C50 14"/15" that were e-waste from day 1, and that wasnt the worst cpu. AMD should have never released single cores or low cloked dual cores Bobcats, the same applies for dual cores and low clocked small cores never than that.
</deity>. I had suppressed any memories of single core Brazos, but you're right. They existed. If you ever suffered though those, you have my sympathy. I can't even begin to imagine.

Hells yeah! I've got a 3050e in my daily driver Lenovo, and it's all that I could want in a "basic" laptop. It's even capable of 4K screen res, using VSR, on a 1080P native IPS LCD screen. Shame that the laptop only has 4GB soldered RAM, and the Lenovo pre-installed Win10 (now Win11) image, has a commit charge above 4GB even just booting it up. :(
If not for other requirements, the Athlon 200GE in my HTPC would make me a decent daily driver. It's not fancy, but gets the job done. No fuss, no muss.
 

jpiniero

Lifer
Oct 1, 2010
10,307
2,498
136
If it's quantity you ask for there is a certain competitor which is in a much better place to deliver it.
Even though I do think the GloFo Zen 3 is real, I totally agree with this. I'm not sure why AMD thinks this is a market worth spending resources on even if GloFo wafers are cheap. Dali does need to be upgraded if they are going to continue with it though.
 

moinmoin

Platinum Member
Jun 1, 2017
2,897
3,875
136
I'm not sure why AMD thinks this is a market worth spending resources on even if GloFo wafers are cheap.
I'm pretty sure AMD currently (in the current shortage and supply chain disruption) doesn't think so. But the easy answer is: WSA. $1.6 billion need to be spent on wafers from GloFo either way, better spend it on something that can fetch a decent amount even if its a budget chip. Cue Monet.
 
  • Like
Reactions: Insert_Nickname

ASK THE COMMUNITY