AMD Bristol/Stoney Ridge Thread

[ao_ika_red]

All AMD has to do is drop clocks on Picasso and it's a done deal. They're already incredibly cheap to make, and they've gotta do something with all that 12nm capacity they're still on the hook to use, so why not? At this point it's just a question of how badly they want to service the low-end market. That market is at near-zero priority with Dr. Su at the helm.

How about new small die to accomodate ULP market? Like 2C/4T die and 6 CU and call it real Athlon. Then, cut it down to 1C/2T +2-3CU and voila, we have Duron.

 

[DrMrLordX]

How about new small die to accomodate ULP market? Like 2C/4T die and 6 CU and call it real Athlon. Then, cut it down to 1C/2T +2-3CU and voila, we have Duron.

Possible, and interesting, but it still costs money. AMD hasn't really taken that market seriously since they cancelled their 20nm products ages ago. There's also the possibility that they might get schooled by various mobile SoCs in that power envelope, even running emulation. Look at what a reference 855 can do in GB4. Now cut that number in half (worst case) and imagine how well a 2c/4t Picasso variant in a 5W power envelope would fare against it. Then consider that 8cx is coming this year at higher performance levels than the Snapdragon 855. That gives you some idea of what AMD might be facing, and why AMD is not exactly excited about returning to that market segment in force.

 

[BTRY B 529th FA BN]

Hi there. I'm still holding off on the CPU for my build but I was thinking of going NVMe M.2 with this upgrade. I currently have 2 SATA III SSD drives and I'd like to see what this new (to me) tech is like. I've seen the Samsung 970 Pro 1TB scores pretty good with Anandtech's benchmark.

The rig will serve as a Gaming only rig for now, and then eventually move over to a daily rig that'll occasionally be doing large 1080p video editing. Am I correct in thinking the Samsung 970 1TB Pro is still the best option? And I think I'd eventually like to use another M.2 drive for storage: eliminating the use of SATA SSDs

Thanks for the suggestions.

 

[Markfw]

Hi there. I'm still holding off on the CPU for my build but I was thinking of going NVMe M.2 with this upgrade. I currently have 2 SATA III SSD drives and I'd like to see what this new (to me) tech is like. I've seen the Samsung 970 Pro 1TB scores pretty good with Anandtech's benchmark.

The rig will serve as a Gaming only rig for now, and then eventually move over to a daily rig that'll occasionally be doing large 1080p video editing. Am I correct in thinking the Samsung 970 1TB Pro is still the best option? And I think I'd eventually like to use another M.2 drive for storage: eliminating the use of SATA SSDs

Thanks for the suggestions.

Even though I don't have PCIE4, I got the new Corsair 600, and the numbers even on PCIE3 are screamers. A one TB M.2 Corsair 600 is $250

 

[BTRY B 529th FA BN]

Even though I don't have PCIE4, I got the new Corsair 600, and the numbers even on PCIE3 are screamers. A one TB M.2 Corsair 600 is $250

Really thought I was in the Ryzen 3000 builders thread, lol

 

[NTMBK]

Possible, and interesting, but it still costs money. AMD hasn't really taken that market seriously since they cancelled their 20nm products ages ago. There's also the possibility that they might get schooled by various mobile SoCs in that power envelope, even running emulation. Look at what a reference 855 can do in GB4. Now cut that number in half (worst case) and imagine how well a 2c/4t Picasso variant in a 5W power envelope would fare against it. Then consider that 8cx is coming this year at higher performance levels than the Snapdragon 855. That gives you some idea of what AMD might be facing, and why AMD is not exactly excited about returning to that market segment in force.

A Snapdragon 855 is NOT a budget chip. It's not comparable to the Stoney Bridge craptop price range.

 

[NTMBK]

Really thought I was in the Ryzen 3000 builders thread, lol

I love the idea of someone pairing a top end PCIe4 SSD with Stoney Bridge!

 

[NostaSeronx]

Allwinner/Amlogic/Rockchip -> Stoney sucessor(22FDX/12FDX/etc) -> HiSilicon/Mediatek/Qualcomm -> Apple/Samsung -> Raven2 successor (7nm/3nm/etc)

If we are making accurate enough comparisons.

The price of a 22FDX device should be in between the AAR group and the HMQ group. While beating the AAR group in performance and nearing the HMQ group in performance. Allowing the Stoney design to use the same laptop/chromebook as the HMQ group or even cheaper laptops from the AAR group.

However, the 7nm small APU should aim to beat any fictitious ARM laptop/chromebook from Apple or Samsung. With an absurdly high ASP, but however is acceptable to the customers of the AS group.

Anything after Zen must increase selling price to vastly increase margins.
Anything after Excavator must decrease selling price while also increasing margins.

 

[DrMrLordX]

A Snapdragon 855 is NOT a budget chip. It's not comparable to the Stoney Bridge craptop price range.

The theory was whether AMD could continue selling Stoney-like APUs on newer nodes (FDX nodes, no less) with more threads in the future. Not whether the 855 is somehow comparable to a 28nm APU from 2016. Just a simple Snapdragon 855 - something which can be had in sub-$400 phones, by the way - would be competitive with a hypothetical 4T Stoney successor. Even running x86 emulation. And that's just the 855 which is already old by this point. Qualcomm has the 855 Plus and the 8cx.

There is no way AMD could compete with "small" XV in the future.

 

[NostaSeronx]

There is no way AMD could compete with "small" XV in the future.

A smaller than XV architecture is more competitive than a flat port of XV.

Excavator has retained the wasteful design practice of its predecessors. Big units, big caches, big etc doesn't always mean higher performance and it definitely doesn't mean lower power. So, AMD would want to make Excavator's successor/the next CMT architecture to be more like Jaguar and Zen. Maybe, even further partitioned for increased modularity and scalability.

-> Lower latency
-> Higher bandwith
-> If not higher IPC, more peak/effective IPS. (GOPs, FLOPs, Memops, etc are all included in this)
-> Definite increase of performance and reduction area and power, which potentially allow for more cores. (This also includes the multimedia cores and GPU cores)

XV on 28nm : NCMT on 22FDX
~1.4V at less than 3.7/greater than 2.8 GHz : ~1.4V at less than 5.2/greater than 4 GHz <== 15W if it peaks here.
~1.0V at less than 2.7 GHz/greater than 2.2 GHz : ~1.0V at less than 4.3 GHz/greater than 3.5 GHz. <== 6W if it peaks here. (some of the 1.4v results are actually 1.25v/1.35v entries which has reduced the height of 1.4V.)
~0.8V at less than 1.9 GHz/greater than 1.5 GHz : ~0.8V at less than 3.4 GHz/greater than 2.7 GHz <== 3W if it peaks here.
etc.

 

[amd6502]

The theory was whether AMD could continue selling Stoney-like APUs on newer nodes (FDX nodes, no less) with more threads in the future. Not whether the 855 is somehow comparable to a 28nm APU from 2016. Just a simple Snapdragon 855 - something which can be had in sub-$400 phones, by the way - would be competitive with a hypothetical 4T Stoney successor. Even running x86 emulation. And that's just the 855 which is already old by this point. Qualcomm has the 855 Plus and the 8cx.

There is no way AMD could compete with "small" XV in the future.

Sure it could. Even simply Stoney ported to 22 fdx with two jaguar++ cores would.

Or go with a lite version of Zen3. Or some radical nosta redesign on 22 FDX, which seems perfect for low transistor projects ~1.5B transistors.

 

[DrMrLordX]

Sure it could.

Not likely.

A9-9425:

https://browser.geekbench.com/v4/cpu/search?dir=desc&q=A9-9425&sort=score

Snapdragon 855 Plus:

https://wccftech.com/snapdragon-855-plus-benchmar/

A12x:

https://browser.geekbench.com/v4/cpu/search?dir=desc&q=A12x&sort=score

Kirin 980 (one sample):

https://browser.geekbench.com/android_devices/839

Adding two cores and 100-200 MHz will not turn Stoney Ridge into a competitor. Look for Intel's Lakefield later this year (and hopefully a worthy successor to Gemini Lake) as well.

 

[NostaSeronx]

Not likely.

It is more likely to compare to similar SoCs.

A6-9220C:
https://browser.geekbench.com/v4/cpu/13586767

Allwinner H6:
https://browser.geekbench.com/v4/cpu/13858372

Amlogic S912:
https://browser.geekbench.com/v4/cpu/13885835

Rockchip RK3399:
https://browser.geekbench.com/v4/cpu/13653586

The 22FDX design only needs to be faster than;
- Amlogic S905X3
- Rockchip RK3530
- Allwinner A100 or A200

and give the extra cost experience of Windows and x86-64. Thus the sell price target is lower with 22FDX/12FDX rather than higher with 14LPP/8LPP.

The closest thing to the Set-top boxes are the ThinkCentre M625 Tiny desktops. But, is it really worth ~10-20x the cost for Windows Pro, sata or m.2, so-dimm, USBs, etc?

 

[Shivansps]

Well if Qualcomm has something to bring to the table we would see it by now, the lack of designs win tells the whole history. The 850-855 failed in notebooks, and very badly.

The 8CX is other thing but still, i saw the numbers compared to an I5-8250U show by Qualcomm itselft and is not THAT impressive. With the IGP performing like 15% faster than an Intel UHD630 that was 100% pure crap 3 years ago, thats around Vega 3 perf. CPU numbers are also all over the place, with the i5 winning by 400 points in 3dmark cpu score.

BUT one thing is clear AMD cant get anywere near that with Stoney Ridge... Even with 8C it cant hope to match an I5-8250U 4C/8T in MT that is 1.6GHz base 3.6Ghz turbo. Thats around the perf of the 8CX. A 2C/4T Picasso would not do the job either, we are talking of performance that is a slower than a Ryzen 15W 3500U and probably faster than 15W Ryzen 3 3300U. AT least in CPU because those GPU numbers are awfull.

The 8CX is a premium chip at 7W, it remains to be see if AMD can lower those 15W of the 3500U... to something like 10W while retaining higher perf, there is no need to match the 7W of the 8CX.

 

[DrMrLordX]

Well if Qualcomm has something to bring to the table we would see it by now, the lack of designs win tells the whole history. The 850-855 failed in notebooks, and very badly.

Snapdragon 855 never made it into any notebooks. 8cx is coming this year. 850 is just a rebranded 845.

The 8CX is other thing but still, i saw the numbers compared to an I5-8250U show by Qualcomm itselft and is not THAT impressive.

More impressive than a hypothetical 4T Stoney Ridge (which is the topic of discussion).

The 8CX is a premium chip at 7W, it remains to be see if AMD can lower those 15W of the 3500U... to something like 10W while retaining higher perf, there is no need to match the 7W of the 8CX.

Now you're on the right track. No need to keep Stoney alive when they can use Picasso instead.

 

[NostaSeronx]

BUT one thing is clear AMD cant get anywere near that with Stoney Ridge...

No need to keep Stoney alive when they can use Picasso instead.

I'm not vouching for a Stoney Ridge 9230e or anything.

At minimum;
IOD with a 16B/cycle Infinity Fabric to the ACD.
16B at 2.4 GHz(DDR4-2400) => 38.4 GB/s
16B at 3.6 GHz(DDR4-3600) => 57.6 GB/s
With memory;
64-bit 3.6 GHz(ideal) => 28.8 GB/s
to 128-bit 3.6 GHz(mostly un-ideal) => 57.6 GB/s
If late enough 64-bit 4.8 GHz+ => 38.4 GB/s+. (DDR5) // should be feasible since it does support LPDDR4X at max speed(4267 MHz).
Preferably the IOD will be done with ultra-high-speed + ultra-low-leakage I/O, Logic, etc. Allow the I/O in 22FDX to be down one tenth the consumption across; PCIe, DDRx, Security, Display cores, etc.

On to the ACD side; PDSOI(HP) -> FDSOI(UHP to ULP) <- Bulk(LP)
16B/cycle Infinty Fabric enters which gets buffered by a die cache which basically is a L3 cache.
CPU core is overhauled further than Bulldozer/Excavator into extreme scaling. Low Vdd gets higher frequencies and High Vdd gets into extremes. Near-threshold(0.3vdd) to low-threshold(1.0v) goes to sub-10 watt. While super-threshold range(>1.0v to 1.55v) gets super high performances and goes towards >15W.
Multimedia cores go from UVD/VCE cores to VCN cores. The UWAVFS from FDSOI should improve multimedia enough to support four VCN cores or get away with two for the same function as four on 14LPP.
GPU core if possible should jump from GCN to RDNA. It will either lead to 3CUs to 2CU(Single dual-CU) or 4CU(Double dual-CU).

etc, etc, etc.
The longer it takes the more absurd the end results can be as 22FDX matures. It also makes it cheaper, so it definitely is better to wait.

At best it beats Raven2 with ease, at worst it is only 1.5x the performance of Stoney. While, having a successor product in 12FDX which pushes it further than Bristol. This is where 128-bit DDR4/DDR5 is probably better as a long-term.

 

[NTMBK]

If AMD want a <7W competitor to Snapdragon chips, they shouldn't waste power budget on inter-die communication. A monolithic design is much more efficient at that power level.

They would be better suited creating a dedicated chip for low power, with LPDDR5 memory interface and Zen 2 CPU with its layout optimised for maximum efficiency with reduced peak clocks. But I don't know how much value there really is in an x86 chip in a tablet or phone.

 

[DrMrLordX]

If AMD want a <7W competitor to Snapdragon chips, they shouldn't waste power budget on inter-die communication. A monolithic design is much more efficient at that power level.

Pretty sure they don't on any Zen-like CPU with just one CCX.

But I don't know how much value there really is in an x86 chip in a tablet or phone.

Stoney isn't a tablet/phone chip. Never was, never will be. The issue is that a Stoney successor has to face Atom and whatever Qualcomm is bringing to the PC market. All those AiOs that had Stoney Ridge chips? Their future selves may be running ARM, especially if AMD continues not to serve that market and Intel struggles with 10nm. That market is overdue for a replacement for Gemini Lake. There is one coming, but we don't know in what volumes or in what markets.

 

[amd6502]

1. Simply a gain of 100 to 200MHz would hardly justify a port to a new node. A grade in bin quality is greater than those frequencies.

22FDX would do a whole lot more, including dramatically lowering the power usage.

2. Comparing a product with 4x the cores and in a different price class, on 7nm versus 28nm is pretty pointless. (Not sure what you really expect when you're doing this.)

Throwing sheer numbers of cores to raise multithread is a great way to inflate benchmarks. These MC scores are unlilely to reflect real life performance and user needs. In many cases these are more marketing gimmicks used to charge more for some models of androde telephones. Quadcore makes sense on these telephones, octacore rarely does (exceptions for some users who actually use these telephones hooked up to a real display as a portable computer---and such users usually will be willing to pay nice premiums).

Now AMD doesn't need to achieve leadership status in this low end segment. It's a very large market. All that's needed is to have a viable product. As long as it isn't outclassed in performance and price, many consumers still would prefer or pick an x86 product over ARM. (Not everybody trusts the acorn ecosystem. The speed of their security updates and the risk of exploits leaves much to be desired. It is also more limited in the OSes and software you can run.)

 

[NostaSeronx]

If AMD want a <7W competitor to Snapdragon chips, they shouldn't waste power budget on inter-die communication.

For 22FDX the inter-die interlinks costs less power because of the lower SerDes power. It also reduces the wire length between all the components.

The benefit of two dies on 22FDX. Is that both dies would be smaller than a single monolithic die. It also allows them to simultaneous develop both dies on different roadmaps. A refresh on the ACD, can have a new IOD. With support with HDMI 3.0 or Displayport 2.1, add USB 5.0 functionality, etc. It would cost less than a shrink to 2.5nm onwards for those capabilities.

They can also do Excavator in an aggressive manner. (Steamroller(13T) -> Excavator(9T) manner, smaller logic libraries on same node)
Gen1 => 8T-104CPP w/ 14FD gates
Gen1+ => 7T-104CPP w/ 10FD gates +20% perf
Gen1++ => 6T-104CPP w/ 7FD gates +30% perf
The extra benefit of planar is that asymmetrical FEOL-to-BEOL.

 

[DrMrLordX]

2. Comparing a product with 4x the cores and in a different price class, on 7nm versus 28nm is pretty pointless. (Not sure what you really expect when you're doing this.)

You brought up a hypothetical 4T Stoney. I'm showing that there are multiple low-power products today that make a potential 4T Stoney already look awful. There are ARM SoCs that have +~50% ST and +~200% MT performance over 2T Stoney. Do you really think the expense of adding clocks to Stoney, increasing its IPC, and porting it to a different node makes any sense when something like a Snapdragon 855 can already be had in a sub-$400 product? Remember, that's last year's SoC. It's already long in the tooth and much lower in cost than it was at introduction.

It makes no sense. At all.

Meanwhile, 4c Picasso:

https://browser.geekbench.com/v4/cpu/search?dir=desc&page=1&q=3500U&sort=score

A hypothetical 2c 3500u variant would have enough ST power to still be useful in applications where AMD once used Stoney. It would cost less than improving Stoney.

 

[NostaSeronx]

It would cost less than improving Stoney.

It wouldn't...

Picasso to even get into Qualcomm's field needs to be Renoir.

$1.4 to $1.6 per 100 million gates. With increased development costs over the slap Cortex A76 / Cortex A55 and call em Kyro development. Increased wafer costs, etc.

28nm Kaveri($1.6 100M) -> 28nm Carrizo($1.4 100M) -> 28nm Stoney($1.3 per 100M) -> 22FDX($1.05 per 100M) == lower development cost, lower utilization cost, etc.

Cortex A76 => 3 ALUs/2 AGUs/2 16B FMACs
Zen => 4 ALUs/2 AGUs/2 16B FADD+2 16B FMUL <-- This
Zen2 => 4 ALUs/3 AGUs/2 32B FADD+2 32B FMUL <-- and this have more static leakage than the A76.
Excavator => 2 ALUs/2 AGUs/2 16B FMACs

The basic idea is more cores don't lead to higher single threaded performance. Hence, the eQuad mentality... two 3 GHz Cortex A9 cores will out perform four ~1.6 GHz Cortex A9 cores. While being on-par or greater than two ~1.8 GHz Cortex A15 cores. Bigger OoO and pack width doesn't mean higher performance as well.

A 22FDX device can easily be sub-$100 usd and compete with these big league devices of sub-$400 usd. With AMD getting a higher return than going cost to cost with Qualcomm.

 

[amd6502]

Meanwhile, 4c Picasso:
https://browser.geekbench.com/v4/cpu/search?dir=desc&page=1&q=3500U&sort=score

A hypothetical 2c 3500u variant would have enough ST power to still be useful in applications where AMD once used Stoney. It would cost less than improving Stoney.

Well the ST perf is good. For ULP 6W, maybe they will have a Raven2 1c/2t die salvaged variant. Zen is pretty impressive, and I think it does have promise for ULP. Especially if they were to do something like 4-way multithread in Zen3 or Zen4. In that case, one could have single core ULP models that still would manage MT demands for a large % of consumers.

I don't now if they can do tiny sized dies on 7nm though. What are the minimum die size requiremnts? Can they do 1.5B -2B transistor dies, and if so what is the area? If not, then there's one big advantage for 22FDX.

 

[DrMrLordX]

I don't now if they can do tiny sized dies on 7nm though.

Why not? They can do 74mm2 chiplets. TSMC ran 256Mb SRAM test cells in 2017 which were ~43mm2. Picasso is 210mm2 on 12/14nm so I would imagine that on 7nm, the entire die would shrink to ~94mm2. They could shave off half the CPU cores and even shrink the iGPU and still have a die larger than the test cell.

 

[NTMBK]

Pretty sure they don't on any Zen-like CPU with just one CCX.



Stoney isn't a tablet/phone chip. Never was, never will be. The issue is that a Stoney successor has to face Atom and whatever Qualcomm is bringing to the PC market. All those AiOs that had Stoney Ridge chips? Their future selves may be running ARM, especially if AMD continues not to serve that market and Intel struggles with 10nm. That market is overdue for a replacement for Gemini Lake. There is one coming, but we don't know in what volumes or in what markets.

All in ones have buckets of cooling capacity. They can handle a 17W CPU without even blinking. They used Stoney Ridge because it was dirt cheap, not because it was <10W.