Array
(
    [content] => 
    [params] => Array
        (
            [0] => /forum/index.php?threads/amd-getting-comfortable-with-samsung-4nm.15327/
        )

    [addOns] => Array
        (
            [DL6/MLTP] => 13
            [Hampel/TimeZoneDebug] => 1000070
            [SV/ChangePostDate] => 2010200
            [SemiWiki/Newsletter] => 1000010
            [SemiWiki/WPMenu] => 1000010
            [SemiWiki/XPressExtend] => 1000010
            [ThemeHouse/XLink] => 1000970
            [ThemeHouse/XPress] => 1010570
            [XF] => 2021370
            [XFI] => 1050270
        )

    [wordpress] => /var/www/html
)

AMD getting comfortable with Samsung 4nm?


Samsung's latest 4nm Exynos 2200 was co-designed with AMD. Is AMD using the same playbook as Qualcomm?

Samsung using AMD's IP is not the same as AMD using Samsung's process technology. AMD is exclusive to TSMC at 7nm, 6nm, 5nm, 4nm, and 3nm. Beyond that I do not know. Maybe AMD will switch to Samsung for GAA? The AMD and TSMC partnership seems to be working quite well. TSMC adding Intel into the mix may be problematic for AMD so a switch could happen but I seriously doubt it. The Samsung PDKs are troublesome for early adopters which is why NVDA and QCOM are back at TSMC for N3.
 
Samsung's latest 4nm Exynos 2200 was co-designed with AMD. Is AMD using the same playbook as Qualcomm?
My assumptions (from this and other news reports) are that Samsung licensed graphics IP from AMD, not that AMD had a role in implementing that IP on the SoC. I see this more akin to licensing IP from ARM and having a customer (Samsung LSI) instantiate it on an SoC, than the Byzantine arrangement you saw between AMD and Hygon, for example.

The architecture of the RDNA2 IP, in terms of hierarchy and how many EUs / Shaders / etc are instantiated within that IP may have had to be customized by AMD, since what is in the E2200 is overall smaller than anything they have in their Ryzen APUs or the Console products, but this doesn't mean AMD needed to have access to any Samsung Foundry PDKs, as Samsung LSI may have been the group responsible for hardening the IP in the SoC design (the way they would with an ARM CPU / GPU).
 

Samsung's latest 4nm Exynos 2200 was co-designed with AMD. Is AMD using the same playbook as Qualcomm?
One of the interesting development is that the increased Intel order volume and projects at TSMC. AMD must maintain their excellent relationship with TSMC and even make it stronger to prevent Intel and Nvidia to chip away AMD's capacity at TSMC.
 
Some word on the internet in the last 24 hrs is that the 2200's inclusion of RDNA2 did not help the performance. If it pans out, then AMD would not be comfortable with Samsung 4nm.
 
Transistor density on Samsung's 4nm is roughly equivalent to TSMC's 6nm process node.

Very true, and Samsung 3nm doesn't compare well with TSMC N3 either:

 
Good historical view:

AMD has literally never been in a stronger position to face Intel’s challenge. The company has now been profitable every year since 2018. While this might not be a record yet, AMD was once infamous for its unreliable financial performance. Under Lisa Su’s management that’s no longer the case.

The fastest way for Intel to hand the pole position back to AMD would be to assume Alder Lake guarantees that outcome. Intel has never quite managed to dispose of its chief competitor, regardless of how strong its advantage in the moment might be. Gelsinger is likely aware of this. His message was probably more intended to inspire than to be evaluated by the press.

 
Good historical view:

AMD has literally never been in a stronger position to face Intel’s challenge. The company has now been profitable every year since 2018. While this might not be a record yet, AMD was once infamous for its unreliable financial performance. Under Lisa Su’s management that’s no longer the case.

The fastest way for Intel to hand the pole position back to AMD would be to assume Alder Lake guarantees that outcome. Intel has never quite managed to dispose of its chief competitor, regardless of how strong its advantage in the moment might be. Gelsinger is likely aware of this. His message was probably more intended to inspire than to be evaluated by the press.

From the quoted Extremetech.com article:

"Objects in the Rearview Mirror are Closer Than They May Appear". 🙂
 
The point that you can't have serious graphics on embedded is greatly challenged by M1.

It's 10W graphics which can challenge last gen 75W desktop cards, and do it without own DRAM. Part of this is 2-3 node advantage, but part must be the higher efficiency.

Memory access, cache hierarchies, and shader engine scheduling I believe are some of most inefficient, and "bruteforced" parts of GPU designs now.

M1 is know for having a gigantic cache for its size, but that alone also means that the logic which managed the cache is very advanced. If you just put few hundred megs L3 on top of anything, the increased latency, and over-eviction will in fact decrease the performance.
 
Last edited:
  • Like
Reactions: VCT
Back
Top