AMD "Zen 3" Tested with the Faster Boost Clock-Speed Ramping Speed than Snapdragon SoCs

[btarunr]

AMD's "Zen 3" architecture, particularly in its low-power mobile iterations, change their clock speeds at a very high rate of speed (switching between lower idle clock-speeds to higher boost clock bins), finds a study by Chips and Cheese, which tested 17 processors across brands and machine-architectures, including mobile SoCs. The interesting finding here is that the Ryzen 7 5800U "Zen 3" mobile processor has a much faster speed-ramp than even SoC powering handhelds, such as the Qualcomm Snapdragon 821, returning a ramp-time of just 1.6 ms, compared to 19.6 ms on the Snapdragon. We now see why AMD likes its processors to run detached from the 10 ms tick-rate of Windows internal power-management (the rate at which the OS reports its workload to the processor, so it could respond with a higher performance state). A rapid boost clock ramp rate allows the processor to better ration its power budget in response to workload.



View at TechPowerUp Main Site | Source

 

[Denver]

I wonder what stops AMD from creating its own Arm Soc for smartphones...

 

[btarunr]

I wonder what stops AMD from creating its own Arm Soc for smartphones...

They first need an E-core. Maybe "Zen 2" on 4 nm with aggressive power-management could be a nice E-core, with each CCX having comparable die-area of a "Gracemont" cluster.

 

[1d10t]

Zhaozin KaiXian KX-6640MA, wasn't that mainland manufacturer licensed the first Zen?

 

[pantherx12]

I wonder what stops AMD from creating its own Arm Soc for smartphones...

I thought they sold off a lot of their arm related business ( and therefore patents) when they stopped development of their arm server SOCs.

 

[Deleted member 185088]

I wonder what stops AMD from creating its own Arm Soc for smartphones...

Android supports x86, they can make a low power Zen, maybe the market isn't appealing for AMD.

 

[R0H1T]

No they didn't, they stopped working on K12(?) at the same time Zen was being developed. But still have all the IP from what I remember.

 

[prtskg]

I thought they sold off a lot of their arm related business ( and therefore patents) when they stopped development of their arm server SOCs.

I haven't read anything of that sort. Maybe you're remembering something else.

 

[Mussels]

This is why they have those fun temp spikes, as they rapidly change from idle to WHEEEEE

 

[Wirko]

Where are Intel mobile chips for comparison? That Goldmont Celeron isn't exactly the best example one could find...

 

[R0H1T]

Pretty high (or bad) last I checked.

 

[trsttte]

I wonder what stops AMD from creating its own Arm Soc for smartphones...

Engineering resources to dedicate to a new market where they have little to no experience. They have Arm licenses and designs and have in the past that they're ready to provide Arc soc's if the demand is there but Qualcomm has the phone market pretty cornered and AMD is resource (and even production) starved to try to enter a new market.

 

[Denver]

Engineering resources to dedicate to a new market where they have little to no experience. They have Arm licenses and designs and have in the past that they're ready to provide Arc soc's if the demand is there but Qualcomm has the phone market pretty cornered and AMD is resource (and even production) starved to try to enter a new market.

They are technically already in this market, the Eclipse GPU is based on RDNA2, it just lacks custom cores...

I don't think it's a big risk they could even develop the custom cores together with another company, maybe Samsung, Broadcom etc.. or just use the standard "Cortex" core developed by ARM + Xilinx's NPU/ASIC and their RDNA3 GPUs

 

[TechLurker]

They are technically already in this market, the Eclipse GPU is based on RDNA2, it just lacks custom cores...

I don't think it's a big risk they could even develop the custom cores together with another company, maybe Samsung, Broadcom etc.. or just use the standard "Cortex" core developed by ARM + Xilinx's NPU/ASIC and their RDNA3 GPUs

AMD already has a joint agreement with Samsung to at least share dev resources on making their GPUs mobile compatible. As well, they also negotiated a non-compete agreement to prevent AMD immediately stepping on Samsung's toes in the mobile space and vice-versa, preventing Samsung from using whatever hybrid SoC to directly compete against certain portable offerings from AMD.

So for the time being, it's more that AMD is content to let Samsung improve RDNA for mobile, and see if Samsung can improve their ARM IP to make a powerful hybrid SoC. Samsung seems to be floundering though, with doesn't bode well for AMD's ambition to expand some of its footprint into the mobile space.

At the same time, Lisa Su was on record stating that they kept their ARM project off to the side, warm but not entirely forgotten, in one of the early Zen interviews, but that their focus is regaining their competitive footing in the x86 space. And this was before the much more recent agreement with Samsung, so the Samsung/AMD joint project may have put their ARM side-project on ice.

 

[Minus Infinity]

They first need an E-core. Maybe "Zen 2" on 4 nm with aggressive power-management could be a nice E-core, with each CCX having comparable die-area of a "Gracemont" cluster.

It is being said that the Bergamo cores will become E-cores in Zen 5 as they also move to hybrid architecture. Zen 5c will destroy Gracemont+++

 

[doc7000]

I wonder what stops AMD from creating its own Arm Soc for smartphones...

One problem that Intel has today is that they do a lot of different things, they don't just do CPUs in the as they have like 10 different divisions. For example my AMD motherboard has Intel networking on it, When Lisa became CEO at AMD they stopped a lot of projects to focus on 3 core markets for CPUs and GPUs which is mobile, desktop, and servers/data centers. This has helped AMD take a limited amount of resources and focus more of it on each segment instead of taking a highly limited budget and spreading it across 10 different things to focus on.

Even better they have been able to make one CPU core design for both mainstream and server/data centers, it would be even better if they could use the same chiplet design on mobile as well. Now that AMD has much more revenue coming in they can think about spreading into other segments though the segments that they are in now are ones that have growth in the future (primarily server/data center).

 

[AnarchoPrimitiv]

They first need an E-core. Maybe "Zen 2" on 4 nm with aggressive power-management could be a nice E-core, with each CCX having comparable die-area of a "Gracemont" cluster.

Isn't Zen4c, the core that's going to be in Eypc Bergamo chips and can fit 16 hyperthreaded cores in a single Chiplet versus only 8 cores in the regular Zen3/4 Chiplet, kind of like an e-core for AMD? If I'm recalling correctly, AMD took the Zen4 core and regular 8 core chiplet and paired down some of the cache and other things to be able to fit 16 hyper threaded cores in the same footprint. That's how Bergamo will reach 128 Zen4 cores and 256 threads versus the 96 cores and 192 threads of Genoa. It makes me wonder though, because Genoa uses 12x 8 core chiplets to reach 96 cores, and Bergamo is using only 8x 16 core chiplets to reach 128 cores, does that mean there could be an Eypc SoC in the pipeline that uses a total of 12x 16 core Zen4C chiplets to reach a core count of 192 core/384 threads? And even one with v-cache creating an Eypc chip with 192 cores and over a gigabyte of cache? It'd be really cool to see such products....

There are even some rumors going around that AMD will take a regular Zen4 chiplet with 8 cores and pair it with a Zen4C chiplet with 16 cores to have a 24 core 48 thread CPU for AM5/X670 to counter the 13900k and its 24 cores and 32 threads. Also heard that AMD might add 3D v-cache to this hypothetical chip to not only make up the lost cache on the paired down Zen4C chiplet, but to surpass the total cache amount on the 16 core/32 thread 7950x. This may just be rumors that never come to fruition, but it sounds extremely logical to me and we know for 100% fact that AMD already has the Zen4C chiplets for Bergamo, so it doesn't seem too difficult to replace one of the regular 8 core Zen4 chiplets from the 7950x with a Zen4C chiplet from Bergamo and create a 24 core/48 thread consumer desktop chip for AM5 to neutralize any claims by Intel that they have more cores available on their consumer desktop platform (which they will assuredly make since with Zen4, AMD has for all intents and purposes, taken away the possibility for Intel to claim a frequency advantage since it reaches 5.8Ghz and could even go higher). I'd even be willing to bet that the Zen4C core is a more performant core than the e-cores found in alder/raptor lake if AMD is creating a 128 core/256 thread Eypc SoC entirely out of Zen4C cores/chiplets...at the very least I'm sure they'll be much more efficient.