[Rumour] AMD eyes Samsung 4nm for zen 5c chips

M

Marees

2[H]4U
Joined
Sep 28, 2018
Messages
2,094
The c-cores are normally used for server products, but AMD has recently started using them in low power consumer products such as phoenix 2 (2 zen4 + 4 zen4c)

It is speculated that, Sonoma Valley, a successor to AMD Mendocino (a very low power APU) could utilize zen 5c that is fabricated on Samsung 4nm

The lower-end models in the Zen 5c lineup are expected to be made using Samsung’s 4nm process, while the higher-end models will utilize TSMC’s 3nm process.

https://www.gizmochina.com/2023/11/18/amd-eyes-samsungs-4nm-process-for-zen-5c-chip-production/


If the report is true, it's extremely odd that AMD would have two versions of the same chip made by two different foundries on two processes. As far as we can recall, something like that has never been done before, so we're very curious to see what this looks like in the real world when Zen 5 arrives in 2024.

https://www.extremetech.com/computing/amd-reportedly-taps-samsung-foundry-for-4nm-zen-5c-designs


News Sources:
@Tech_Reve,
@faridofanani96
@Olrak29_

Samsung reportedly has obtained 4nm server chip orders from AMD​

https://www.digitimes.com/news/a20231116PD210/samsung-4nm-server-chip-amd.html

AMD’s Zen 5c chips could be made by Samsung’s improved 4nm technology​

https://www.sammobile.com/news/amd-zen-5c-chips-samsung-foundry-4nm/

​

Samsung lands orders for AMD's 4nm CPUs as chipmaker reportedly seeks to diversify production​

AMD's Prometheus to be made both by TSMC and Samsung.
The report from DigiTimes claims that AMD is set to produce Zen 5c-based 'Prometheus' at two contract semiconductor chipmakers: Samsung Foundry and TSMC. The former is reportedly projected to make 'basic' versions of Prometheus on one of its 4nm-class process technologies, whereas the latter is expected to manufacture 'more advanced' versions of Prometheus on a 3nm-class production node.
Click to expand...
https://www.tomshardware.com/pc-components/cpus/samsung-lands-orders-for-amds-4nm-cpus-report

Samsung secures orders for AMD's 4 nm Zen 5c cores​

Sources close to DigiTimes inform that Samsung managed to secure an important order from AMD. The title of the report refers to server chip orders, but DigiTImes then mentions Zen 5c cores on 4 nm nodes. This type of cores is similar to Intel’s efficiency cores that do not use hyperthreading, but AMD’s c-cores have only been included in the mobile-grade Phoenix2 processors up until now, so the information regarding server chips could be exaggerated.

In any case, the shift towards another foundry is definitely a new strategy for AMD. Over the last decade, Team Red has solely collaborated with TSMC on the processor core side. For the upcoming Zen 5 Prometheus architecture, however, it looks like AMD is looking to diversify the production venues. According to DigiTimes, Samsung would produce 4 nm “basic” versions of the Prometheus chips, while TSMC would handle the “more advanced” cores on the latest cutting-edge 3 nm nodes.
Click to expand...
https://www.notebookcheck.net/Samsung-secures-orders-for-AMD-s-4-nm-Zen-5c-cores.769623.0.html
 
Last edited:
Interesting. Given how expensive TSMC 3nm is said to be, it makes sense AMD may want something cheaper for their lower-end products to help margins.

I don't recall this kind of split happening CPUs in quite the same manner, but there is precedent elsewhere: Nvidia's Pascal. GP100/102/104/106 were TSMC 16nm while GP107 and GP108 were on Samsung 14nm.
 
The C cores are made possible by TSMC redesigned 6T pseudo dual-port bitcell. That’s their tech and their design and I don’t see how AMD could just take that elsewhere.

Furthermore Samsung 4nm is worse than TSMC 4nm and is cheaper because of that reason it would be a step backwards in many respects.

Seems too far fetched to me, the DigiTimes article says it’s for server components, lots of parts in an Epyc CPU that aren’t C cores that could take the upgrade. The 6nm MCD’s for instance could see a performance uplift by going to 4nm and it may help with some of the power and heat issues their GPU chiplets are having.
 
Marees said:
If the report is true, it's extremely odd that AMD would have two versions of the same chip made by two different foundries on two processes. As far as we can recall, something like that has never been done before, so we're very curious to see what this looks like in the real world when Zen 5 arrives in 2024.
Click to expand...
If you look at cell phones, Samsung has offered its Galaxy line in both Samsung's native Exynos chip (International market) and SnapDragon chips (primarily US only). So different chips, but same components that the chips have to work with.

They may be doing it as an experiment or to get comfortable using Samsung for x86 CPUs. Samsung has also been partnering with AMD to co-develop its Exynos line of chips (ARM) for cell phones - there is a relationship there. More competition to TMSC is good.
 
sleepeeg3 said:
If you look at cell phones, Samsung has offered its Galaxy line in both Samsung's native Exynos chip (International market) and SnapDragon chips (primarily US only). So different chips, but same components that the chips have to work with.

They may be doing it as an experiment or to get comfortable using Samsung for x86 CPUs. Samsung has also been partnering with AMD to co-develop its Exynos line of chips (ARM) for cell phones - there is a relationship there. More competition to TMSC is good.
Click to expand...
Competition with TMSC is good. A competitive foundry not in a potential war zone would be even better. Let's not forget that DRPK, the "dear leader" country, is a crazy man who would knock over South Korea in an instant, if given the chance.
 
Maybe the advanced chiplet tech AMD has now developed allows such mix and match techniques to be implemented? One "C" core may be different then another "C" core from a different foundry?
 
funkydmunky said:
Maybe the advanced chiplet tech AMD has now developed allows such mix and match techniques to be implemented? One "C" core may be different then another "C" core from a different foundry?
Click to expand...
I think it is more like the Nvidia Ampere situation, where they get a deal from Samsung that they just can't refuse.

The result for nvidia was that RDNA 2 caught up with Ampere but luckily the chip shortage & DLSS kept them ahead.

For AMD this would be something like the chromebook market, I guess.

Would make sense, if AMD is planning on flooding the laptop market in 2025/2026 with cheap Samsung made chips
 
Last edited:
Marees said:
I think it is more like the Nvidia Ampere situation, where they get a deal from Samsung that they just can't refuse.

The result for nvidia was that RDNA 2 caught up with Ampere but luckily the chip shortage & DLSS kept them ahead.

For AMD this would be something like the chromebook market, I guess.

Would make sense, if AMD is planning on flooding the laptop market in 2025/2026 with cheap Samsung made chips
Click to expand...
The below is an example of the budget range that we are discussing about:

https://www.notebookcheck.net/HP-st...s-with-AMD-Mendocino-processors.692994.0.html

HP starts selling budget laptops with AMD Mendocino processors​

csm_c08434652_d4001b0caa.pngAMD's Mendocino APUs conserve battery power more efficiently. (Image Source: HP)
The new AMD Mendocino models from HP start at US$329.00 and feature 17, 15 and 14-inch form-factors, plus up to a Ryzen 5 7520U processor with Radeon 610M iGPU. RAM is limited to 8 GB LPDDR5, but the NVMe storage can be expanded up to 1 TB.
 
This is disappointing to me, is Samsung actually going to be good? Just look at the Snapdragon 8 Gen 1 (disaster in efficiency and thermals) fabbed by Samsung, then the Snapdragon 8+ Gen 1 came out that was an identical chip fabbed by TSMC and it's significantly better. Tensor/Exynos mobile chips fabbed by Samsung are inherently inferior to everything fabbed by TSMC on the same process nodes.

I hope they can catch up and legimitately compete with TSMC.
 
Dopamin3 said:
is Samsung actually going to be good? Just look at the Snapdragon 8 Gen 1 (disaster in efficiency and thermals) fabbed by Samsung, then the Snapdragon 8+ Gen 1 came out that was an identical chip fabbed by TSMC and it's significantly better. Tensor/Exynos mobile chips fabbed by Samsung are inherently inferior to everything fabbed by TSMC on the same process nodes.
Click to expand...
Samsung always inferior to TSMC.
I hope they are cheap enough for AMD to secure sufficient laptop wins. Only competition in that bracket is atom/e-cores by Intel. So it should be do-able. TSMC 4nm/3nm is not the competition for these chips but rather TSMC 6nm. They need to beat TSMC 6nm in both price & performance for this to work for AMD
 
Marees said:
TSMC 4nm/3nm is not the competition for these chips but rather TSMC 6nm. They need to beat TSMC 6nm in both price & performance for this to work for AMD
Click to expand...

It looks like TSMC 7nm/6nm prices will decrease next year.
That will give stiff competition to samsung 4nm !!!

https://twitter.com/dnystedt/status/1730038116764053713

Screenshot_20231201-013243_Opera.jpg
 
Dopamin3 said:
This is disappointing to me, is Samsung actually going to be good?
Click to expand...
It's either 'good enough' for AMD's goals, and they enjoy a better price.

Or it's 'TSMC doesn't have enough extra capacity to fill these orders', and AMD has to use the next option.
Although I'm less inclined to think this is the case. Because everybody cut back their orders, earlier this year. Including AMD.
 
chameleoneel said:
It's either 'good enough' for AMD's goals, and they enjoy a better price.

Or it's 'TSMC doesn't have enough extra capacity to fill these orders', and AMD has to use the next option.
Although I'm less inclined to think this is the case. Because everybody cut back their orders, earlier this year. Including AMD.
Click to expand...
Yet TSMC is still at capacity, anytime someone cuts their production run there‘s a dozen waiting to bid on it.
TSMC 7/6 is heavily booked and Samsung 4 is close enough that most of AMD’s older designs will transition easy enough. Some Zen3 RDNA2 APU’s would probably do well, give it some on die memory and their off to the races.

A Zen3 6/12 CCD paired with a 6650xt as an APU would slap.
 
Last edited:
Lakados said:
Yet TSMC is still at capacity, anytime someone cuts their production run there‘s a dozen waiting to bid on it.
TSMC 7/6 is heavily booked and Samsung 4 is close enough that most of AMD’s older designs will transition easy enough. Some Zen3 RDNA2 APU’s would probably do well, give it some on die memory and their off to the races.

A Zen3 6/12 CCD paired with a 6650xt as an APU would slap.
Click to expand...
I guess. But I don't see why they would go backwards.

Zen 5c is a shrunk down Zen 5. Uses less die space and power, but not considered a drastically worse core, like Intel's E cores. Zen5c should still be better than Zen 3 and probably 4, as well. And I'm pretty sure the next APU's are RDNA3..

I have a feeling AMD won't be F#$!ing around, with Zen 5 and 6. They have had their time to try things and now it's time to make some bigger moves. And with that, (according to leakers) AMD has not yet decided internally, if Zen 6 will be made compatible with current motherboards, or if it will actually require a new mobo.
 
chameleoneel said:
I guess. But I don't see why they would go backwards.

Zen 5c is a shrunk down Zen 5. Uses less die space and power, but not considered a drastically worse core, like Intel's E cores. Zen5c should still be better than Zen 3 and probably 4, as well. And I'm pretty sure the next APU's are RDNA3..

I have a feeling AMD won't be F#$!ing around, with Zen 5 and 6. They have had their time to try things and now it's time to make some bigger moves. And with that, (according to leakers) AMD has not yet decided internally, if Zen 6 will be made compatible with current motherboards, or if it will actually require a new mobo.
Click to expand...
They still clock much lower, the 6T SRAM cell doesn’t clock as high as an 8T one does and they are more prone to error and instability. They also only take a single instruction instead of 2 per clock cycle.

The shrink comes at a cost, what that is will yet be seen but that 6T SRAM cell used by AMD for their C cores is a TSMC patented design I doubt AMD has their permission to take that to a competing fab so I doubt Samsung having anything to do with the C cores.

But AMD is again getting flack for not being able to supply their OEM contracts, they need to find a way to get bulk mobile silicon to the market, AMD essentially paper launched much of their mobile lineup or outright cancelled it this generation. I think Samsung is better served handling those products. Samsung has the capacity, their 4N fab is actually pretty good and they have better or equal packaging facilities available to support it. Going from TSMC 6/7 and GoFlo 12 down to Samsung 4N+ and 8N++ for their more basic designs would be a step up especially for anything APU or SoC related.

Additionally: They could use Samsung for the Epic Embedded lineup, that also got paper launched much to my dismay, I’m still rocking my Zen 1 units because some 5 years in they haven’t actually released an upgrade to them. That said those 3251’s have been champs.
 
Last edited:
funkydmunky said:
No it is not like Intel's E cores in that Zen "C" cores have the exact same IPC as regular Zen cores, it is only power limited.
Click to expand...
And frequency limited, 6T SRAM cells aren’t a new concept but they are popular in mobile devices because their downsides don’t start showing until you get to high voltage or high clock speeds. But they are ~30% smaller than an 8T cell and they do consume significantly less energy when switching between states.
 
funkydmunky said:
No it is not like Intel's E cores in that Zen "C" cores have the exact same IPC as regular Zen cores, it is only power limited.
Click to expand...
If it read as me saying they are like Intel's E-Cores, that is not what I meant.

but not considered a drastically worse core, when Intel's E cores are a different architecture and much worse peformance.
 
Lakados said:
And frequency limited, 6T SRAM cells aren’t a new concept but they are popular in mobile devices because their downsides don’t start showing until you get to high voltage or high clock speeds. But they are ~30% smaller than an 8T cell and they do consume significantly less energy when switching between states.
Click to expand...
Power limited = frequency limited.
 
funkydmunky said:
Power limited = frequency limited.
Click to expand...
Not necessarily it depends on the size of the chip, but I will concede that for Desktops and above that is 100% the case.
But 6T cells are not uncommon in mobile devices especially ARM devices. Assuming one bothered too you could hit the frequency limitations on a cellphone CPU before you hit the power limits, you’ll melt it in the process but you’d theoretically hit the frequency wall first.
 
chameleoneel said:
If it read as me saying they are like Intel's E-Cores, that is not what I meant.

but not considered a drastically worse core, when Intel's E cores are a different architecture and much worse peformance.
Click to expand...
We’ll see when they hit the market, the clock speed decreases may hit the AMD architecture harder than expected. 35% smaller sure but could be clocking significantly lower so even if the IPC is the same the clock speeds could be half for all we know.

I’m really hoping it forces Intel to update their E cores (Gracemont), they haven’t changed overly much since 2008 when it was launched as Bonnell.

AMD’s solution is certainly the easier method and the simpler to schedule. So this is a smart play, and if it gets them to a better idle and low power state than their current offerings then I’m all for it, Intel is walking all over the AMD offerings for office equipment because AMD’s isn’t passing the low power tests for certification I would like to see more AMD on my RFPs.
 
Bifurcation of manufacturing based on profitability(of end-products), likely to become a trend :

The high wafer costs will increasingly force their(foundries') customers to, Weigh more and more which new chipp projects actually need the latest chip production. In the consumer sector in particular, the ultimate products mostly have a finite price spiral, i.e. their selling price in the end-user market cannot be increased arbitrarily. It should start here in some time that the device manufacturers make economic weighing decisions that, over time, allow more and more consumer chips to switch to the second newest or even third newest production.

In perspective, only these chips should remain in the best production at the end, where the margins are traditionally high – i.e. chips for professional equipment, thus for server / HPC and AI applications.
With consumer chips, however, you will probably see in some time that the manufacturers are starting to omit the best manufacturing process or to use it late.

With AMD are therefore at risk in any case Zen 6 and RDNA5. Zen 6 in particular could enable this separation according to customer segments, since AMD is probably of different sizes of CCDs (8C, 16C, 32C) wants to hang up. Also with nVidias Blackwell successor ruby one could act in such a way that only the HPC solutions are manufactured under 2nm production and a 3nm derivative is selected for the consumer solutions. Of course, none of this has to happen, in addition to economic considerations, the large chip manufacturers always have to see what their respective competition is doing and whether it is not performing in terms of performance, you ergo yourself to follow suit and thus be forced to do more modern, expensive chip production. This decision is likely to become more and more critical, especially among the successors of the 2nm generation – and thus at some point in the direction of "economy.

https://m.3dcenter.org/news/weiter-...ng-und-fertigungsanlagen-bis-hin-zum-2nm-node
 
Holy shit, this is really kind of cool. Not all silicon benefits from node reduction and this could mean that different parts of dies could be segmented off without the associated latency.

More chiplets, please.
 
You must log in or register to reply here.
Back
Top