AMD Bristol/Stoney Ridge Thread

[moinmoin]

Actually it is the best strategy going forward especially with GlobalFoundries. That specific market is massive in volume of units sold. It is 1 billion units in size compared to faster than Atom at 10 million units large.

Maybe you can convince GloFo to license Stoney from AMD and produce those 1 billion units.

 

[NostaSeronx]

Maybe you can convince GloFo to license Stoney from AMD and produce those 1 billion units.

Stoney is 125 mm2, which is larger than both ~75 mm2 Dual-core Bobcat and ~110 mm2 Quad-core Jaguar.

The various AMD sources indicate a new architecture for both CPU and GPU as well as a new SoC design.

Bobcat for example ~90% perf in half area. With the switch to SMT on high performance, to get that perf there must be a switch to CMT on Ultra-low-power. Which follows up on the switch to Stoney on low-end and G-series embedded. A switch to CMT, however Stoney's design is hampered by high IPC design and low clocks relative to Zen. Which wouldn't exist with a Thin-OoO CMT in vein of 2x Bobcat or 2x Jaguar, or 2x New ULP Arch.

 

[jpiniero]

Puma+Excavator (Carrizo-L/Stoney) are on a cheaper node relative to Intel's low-end while competitive towards ARM's low-end. It only needs to be faster than Cortex-A55/A510/A72.

Intel 14 nm has got to be pretty friggin cheap at this point. Seems the competition is mainly MediaTek these days, and they do have a 4xA73+4xA53 Chromebook processor (MT8183).

 

[NostaSeronx]

Intel 14 nm has got to be pretty friggin cheap at this point. Seems the competition is mainly MediaTek these days, and they do have a 4xA73+4xA53 Chromebook processor (MT8183).

Minimum cost of 16nm(90CPP/64Mx)/14nm (78CPP/64Mx) is at $3500, and you expect Intel's 14nm (70CPP/52Mx) to be cheaper?
The minimum cost for 28nm Bulk(SLP) and 22FDX is half that; 22FDX is cheaper than 28Custom(28SHP/28A/28HPA nodes).

MT8183 is a 10W TDP processor.
6700T/4.5w => 260 single-core score ===> 1.8x in FPU, and 1.8x in INT only if collaborative mode is in. 22FDX ST 260 to 468 single-core depending on weight and CMT-implementation.
9220C/6w => 375 single-core score
8183/10w => 297 single-core score.

8183/10W MT score => 950 multi-core score.
A12-9720P/15W MT score => 1313 multi-core score

A 22FDX design with 2M/4C should be faster.

 

[burninatortech4]

The thing that really killed Stoney, was the exceptionally poor SINGLE-CHANNEL DDR4 performance.

Both lack of DDR4 higher-speed RAM support, as well as inefficiencies, would often see the compute side (CPU cores) DRAM bandwidth-starved, and stalling out. Especially using VSR @ 2560x1440 on a 1080P laptop screen. It becomes notably slower.

A Zen 3050e-based laptop, I can use VSR to bump up a 1080P screen to 4K UHD res., and NO appreciable slowdown.

My A6-9220e manages 7.8 GB/s memory bandwidth with a single 8GB DDR4-2400 SODIMM. It's atrocious. On a good memory controller it should be closer to 17 GB/s. The K15 memory controller is terrible.

Youtube completely flops unless you run h264ify. General web performance is very poor. It's perfectly fine for 1080p 30 fps HEVC local playback from my NAS. Otherwise it'd be in the trash.

It wasn't powerful in early 2018 and it's completely obsolete in 2021.

 

[jpiniero]

Minimum cost of 16nm(90CPP/64Mx)/14nm (78CPP/64Mx) is at $3500, and you expect Intel's 14nm (70CPP/52Mx) to be cheaper?

Yeah easily. 14 nm is 7+ years old. The Broadwell Core M was released in Sept 2014.

 

[NostaSeronx]

Yeah easily. 14 nm is 7+ years old. The Broadwell Core M was released in Sept 2014.

Mind you, the node you are thinking about is 22FFL from Intel, 17LPV from Samsung, etc. 14nm BEOL didn't drop in prices as aggressively as the 22nm/28nm BEOL generation did.

Which is the reason >$3500 for ~14nm nodes and <$1750 for ~22nm nodes

GlobalFoundries is also doing whacky stuff with their 14nm Fab. Which means less FinFET wafers out, and more PDSOI/FDSOI wafers out.

In this case, $1.6B burns towards 14nm and low capacity or burns towards 22nm and high capacity.

AMD's use of GlobalFoundries usually reflects GloFo's node push. Right now GlobalFoundries is pushing for 22FDX and pervasive design. This reflects well with ultra-low-power/low-cost usually being on older nodes; 130nm vs 65nm, 40nm vs 32nm, 22FDX vs 14FF/7FF, etc.

 

[jpiniero]

Mind you, the node you are thinking about is 22FFL from Intel, 17LPV from Samsung, etc. 14nm BEOL didn't drop in prices as aggressively as the 22nm/28nm BEOL generation did.

I'm not. Go look it up, 14 nm really is that old.

 

[Shivansps]

Why would an OEM use Stoney when Goldmont Plus is faster?

The tiny amount of the very low end Zen/Intel stuff released to DIY is going to be gobbled up by miners.

Goldmont Plus? ill bet that Tremont is already faster at lower power. Gemini Lake was already in Kaveri area using near to no power in comparison.

https://imgur.com/mz5evUO

https://imgur.com/sjybdgL

And that thing measures under 18W of power consumption at the DC-DC board to get those results.

 

[jpiniero]

Goldmont Plus? ill bet that Tremont is already faster at lower power.

OTOH 10 nm is too expensive because of the mediocre yield and all the depreciation that needs to happen. But it will get cheaper. Intel is doing a Gracemont Atom, and that will be fast enough that AMD must have realized that even Dali isn't going to cut it. Hence the GloFo 12 nm Zen 3.

 

[NostaSeronx]

I'm not. Go look it up, 14 nm really is that old.

Nope, you are in fact talking about 22FFL in regards to cost. It is the 14nm node with the least premium at Intel.



It is the main reason why most American companies on the cost side have dropped 14LPP/12LP/12LP+. Since 22FFL is that much more cost-effective for them.

22FFL at Intel America, 17LPV at Samsung Austin => Cheapest Advanced nodes in America.

Hence the GloFo 12 nm Zen 3.

Again, you are thinking about Mendocini which is 6nm FinFET Zen2/RDNA2. AMD isn't going to cripple themselves chasing after GlobalFoundries shrinking 14nm/12nm wafer starts. When they have access TSMC's ballooning 6nm wafer starts and no obligation to use GlobalFoundries at 14nm/12nm. Those are waived, cancelled, gone, obliterated, going forward there is no 14nm/12nm. It is now best friend N12e at TSMC: "worked in AMD ODC with industry latest technologies => tsmc 12nm"

 

[burninatortech4]

Goldmont Plus? ill bet that Tremont is already faster at lower power. Gemini Lake was already in Kaveri area using near to no power in comparison.

https://imgur.com/mz5evUO

https://imgur.com/sjybdgL

And that thing measures under 18W of power consumption at the DC-DC board to get those results.

Faster than 95W 7890K for ~10W. Stoney isn't even in the same galaxy.

OTOH 10 nm is too expensive because of the mediocre yield and all the depreciation that needs to happen. But it will get cheaper. Intel is doing a Gracemont Atom, and that will be fast enough that AMD must have realized that even Dali isn't going to cut it. Hence the GloFo 12 nm Zen 3.

Has there been anything concrete on Monet (12nm Zen 3)?

 

[NostaSeronx]

Has there been anything concrete on Monet (12nm Zen 3)?

There has been no CPUID for Monet, where Mendocino does: 8A0Fxx.
There has been no socket announced for Monet, where Mendocino does: FT6
The BIOS blob also indicates Mendocino exists, where Monet does not.

 

[burninatortech4]

There has been no CPUID for Monet, where Mendocino does: 8A0Fxx.
There has been no socket announced for Monet, where Mendocino does: FT6
The BIOS blob also indicates Mendocino exists, where Monet does not.

Ok. Good to know. I'm curious why AMD is skipping straight to 6nm when 12nm is super cheap. I see what you said above Nosta. I'm just surprised that 12nm is already being taken out of preference for cheap parts this quickly.

 

[NostaSeronx]

Ok. Good to know. I'm curious why AMD is skipping straight to 6nm when 12nm is super cheap.

AMD is targeting increased average selling price.

2c/3cu(~150 mm2) -> 4c/8cu (~160 mm2), the cost increase doesn't exceed the price increase.

150 mm2 at 30k wafer starts per month vs 160 mm2 at >150k wafer starts per month.

I'm just surprised that 12nm is already being taken out of preference for cheap parts this quickly.

My understanding is that Zen isn't aiming to be cheap. It is a performance part in between Mont and Cove.

 

[Hitman928]

Nope, you are in fact talking about 22FFL in regards to cost. It is the 14nm node with the least premium at Intel.

View attachment 54415

It is the main reason why most American companies on the cost side have dropped 14LPP/12LP/12LP+. Since 22FFL is that much more cost-effective for them.

22FFL at Intel America, 17LPV at Samsung Austin => Cheapest Advanced nodes in America.Again, you are thinking about Mendocini which is 6nm FinFET Zen2/RDNA2. AMD isn't going to cripple themselves chasing after GlobalFoundries shrinking 14nm/12nm wafer starts. When they have access TSMC's ballooning 6nm wafer starts and no obligation to use GlobalFoundries at 14nm/12nm. Those are waived, cancelled, gone, obliterated, going forward there is no 14nm/12nm. It is now best friend N12e at TSMC: "worked in AMD ODC with industry latest technologies => tsmc 12nm"

22FFL is a new node (mostly based on 14 nm) with relaxed rules and additional 'features' geared more toward RF and mixed signal design. This is a different process than 14 nm (+...) It sits between 14 nm and 28 nm in price. It is more of a competitor to GF 22FDX in price and target customers. 22FDX isn't really targeted for APU (higher performance SOC) type of designs.

 

[NostaSeronx]

22FDX isn't really targeted for APU (higher performance SOC) type of designs.

22FDX has a wide operating range going from high performance x86 CPUs to ultra low power ARM MCUs. 22FDX does not require costly custom implementations like 28SHP to get high-performance. It goes on, but this idea of 22FDX is not for high performance because UHP was fused into ULP is not true.

 

[Hitman928]

Say that to the Google Tensor...

That's actually the opposite situation. Google licensing another company's design to use it in a bigger design in a market they don't compete in. Samsung is doing the same, they aren't going to turn around and license their entire SOC to AMD to be their direct competitor.

Even in worst-case scenarios the 22FDX APUs would still be faster than either 28nm Bhavani or 28nm Stoney cases.

Sure, but who cares? It's 2021 and we're talking about maybe putting budget designs on 12LP+ which 22FDX can't compete with for APU designs. Sure, 22FDX would be a little cheaper, but the wafer costs are only a part of the cost and when you include the density improvements of moving to 12LP+, the packaging, distribution, marketing, etc. costs, it makes no sense for AMD to put any kind of APU design at 22FDX at this point.

The purpose of switching is Fully-depleted transistors like FinFETs. . .

The rest is just a bunch of speculation and irrelevant tangents. 22FDX has become a pretty solid node but it took GF too long to get it really ready to be of any kind of interest for a company like AMD. Others will find 22FDX a sweet spot in terms of cost/performance for their designs, but it's not going to be the likes of Intel/AMD/Apple/etc.

 

[Hitman928]

22FDX has a wide operating range going from high performance x86 CPUs to ultra low power ARM MCUs. 22FDX does not require costly custom implementations like 28SHP to get high-performance. It goes on, but this idea of 22FDX is not for high performance because UHP was fused into ULP is not true.

Name me 1 high performance CPU design win for 22FDX. . .

If you've talked with GF about the process (as I have), you would know THEY don't even advertise it as such.

 

[NostaSeronx]

Name me 1 high performance CPU design win for 22FDX. . .

If you've talked with GF about the process (as I have), you would know THEY don't even advertise it as such.

Much like how GlobalFoundries' 28nm only got ARM CPUs till AMD64 showed up?

Sure, but who cares? It's 2021 and we're talking about maybe putting budget designs on 12LP+ which 22FDX can't compete with for APU designs. Sure, 22FDX would be a little cheaper, but the wafer costs are only a part of the cost and when you include the density improvements of moving to 12LP+, the packaging, distribution, marketing, etc. costs, it makes no sense for AMD to put any kind of APU design at 22FDX at this point.

There is no actual die density improvement on 12LP+, since it is targeting a much higher performance market. In which 22FDX is primarily aimed at returning from 125 mm2 to ~100 mm2 to as far as ~75 mm2.

The same issue that happened to 28Custom:


Will happen again to 14Custom/12Custom.

22FDX APU designs do to the new ULP CPU part and ULP GPU part will be faster than the 14LPP Pollock design. Preparing for the switch over that AMD lead up to with Stoney getting the low-cost HVM Essential APU spot over Bristol-L.

The major nodes going forward at GlobalFoundries:
90nm FDSOI on 300mm (from Skywater)
45nm FDSOI on 300mm (improvement over 45nm PDSOI on 300mm)
22nm FDSOI on 300mm

If a customer isn't on them tough luck, everything else is converting to low volume.

Monet is more large than Dali and at a shrinking FinFET fab. << Doesn't reflect GloFo's new direction >>
ULP SoC is more small than Stoney and is at a growing FDSOI fab. << Reflects GloFo's new direction >>

 

[Hitman928]

Much like how GlobalFoundries' 28nm only got ARM CPUs till AMD64 showed up?
View attachment 54417

AMD used GF because they had by contract. Even then, the GF 28 nm node certainly wasn't bad and was just a little behind TSMC. That's a completely different situation.

For the screenshot, where does this screenshot even come from?

22FDX "launched" 6 years ago. Early on they tried to push it as an alternative to 14nm/16nm FinFETs for high performance CPUs. The problem was when you actually dug into the process, not only was it not really competitive for high performance designs, but the node maturity was pretty poor for HVM and it took GF some time to get to the point a high performance, high volume customer would want to use it. On top of that, the back body biasing that is possible on FD-SOI that you see that GF uses a bunch in marketing to make you think the node is as good as Finfets (when back body biasing is applied) introduces design complexity, risk, and increases the layout size. It took the IP suppliers a few tries to get it right as well. By the time everything was really ready, it was too late.

Once again, if it was such a desirable node, surely there must be 1 high performance design on 22FDX, right?

There is no actual die density improvement on 12LP+, since it is targeting a much higher performance market. In which 22FDX is primarily aimed at returning from 125 mm2 to ~100 mm2 to as far as ~75 mm2.

The same issue that happened to 28Custom:
View attachment 54416

Will happen again to 14Custom/12Custom.

22FDX APU designs do to the new ULP CPU part and ULP GPU part will be faster than the 14LPP Pollock design. Preparing for the switch over that AMD lead up to with Stoney getting the low-cost HVM Essential APU spot over Bristol-L.

The major nodes going forward at GlobalFoundries:
90nm FDSOI on 300mm (from Skywater)
45nm FDSOI on 300mm (improvement over 45nm PDSOI on 300mm)
22nm FDSOI on 300mm

If a customer isn't on them tough luck, everything else is converting to low volume.

Monet is more large than Dali and at a shrinking FinFET fab. << Doesn't reflect GloFo's new direction >>
ULP SoC is more small than Stoney and is at a growing FDSOI fab. << Reflects GloFo's new direction >>

GF claims 15% area reduction for 12LP+ versus 12LP which is already denser than 14LP+ which is already denser than 22FDX for high performance designs (Finfets have a significantly higher drive current). I haven't used 12LP+ at all so I can't comment directly, but that's GF claim against 12LP anyway.

All older nodes get 'converted to low volume' or discontinued altogether, not sure why this is news. If 12LP+ is bought in volume, GF will continue to make it. If not, they won't. It's that simple. The "28Custom" as you call it was a node tweaked specifically for AMD's 28 nm designs. When they stopped making those chips, there were no customers for that node so it got discontinued, business as usual.

 

[burninatortech4]

Much like how GlobalFoundries' 28nm only got ARM CPUs till AMD64 showed up?
View attachment 54417There is no actual die density improvement on 12LP+, since it is targeting a much higher performance market. In which 22FDX is primarily aimed at returning from 125 mm2 to ~100 mm2 to as far as ~75 mm2.

The same issue that happened to 28Custom:
View attachment 54416

Will happen again to 14Custom/12Custom.

22FDX APU designs do to the new ULP CPU part and ULP GPU part will be faster than the 14LPP Pollock design. Preparing for the switch over that AMD lead up to with Stoney getting the low-cost HVM Essential APU spot over Bristol-L.

The major nodes going forward at GlobalFoundries:
90nm FDSOI on 300mm (from Skywater)
45nm FDSOI on 300mm (improvement over 45nm PDSOI on 300mm)
22nm FDSOI on 300mm

If a customer isn't on them tough luck, everything else is converting to low volume.

Monet is more large than Dali and at a shrinking FinFET fab. << Doesn't reflect GloFo's new direction >>
ULP SoC is more small than Stoney and is at a growing FDSOI fab. << Reflects GloFo's new direction >>

I've been waiting to see these fabled 22FDX AMD APU's you've been talking about for a few years. I'll believe it when I see it.

 

[Hitman928]

BTW, here are actual GF materials that I've still seen used today. Notice what they say 22FDX is optimized for / targets and what they don't mention.

https://www.gsaglobal.org/wp-content/uploads/2012/05/22FDX_GSA_AMS_Webinar_Schaeffer_rev2-JSA-rev-7.pdf

https://www.globalfoundries.com/sites/default/files/22fdx-product-brief.pdf

 

[burninatortech4]

BTW, here are actual GF materials that I've still seen used today. Notice what they say 22FDX is optimized for / targets and what they don't mention.

View attachment 54421

https://www.gsaglobal.org/wp-content/uploads/2012/05/22FDX_GSA_AMS_Webinar_Schaeffer_rev2-JSA-rev-7.pdf

View attachment 54422


View attachment 54423

https://www.globalfoundries.com/sites/default/files/22fdx-product-brief.pdf

Looks great for auto applications SoC, IOT, and RF. I doubt we see a consumer APU using this node.

 

[Hitman928]

Looks great for auto applications SoC, IOT, and RF. I doubt we see a consumer APU using this node.

Yes, it's a great node for some applications, very good perf/$. You'd be crazy to try to make a competitive consumer APU/CPU on it today though.