I can't wait for these new cores to be paired with Nvidia Blackwell GPU Cores for Windows on ARM. I can imagine with the DynamIQ Shared Unit allowing for 10 Cortex-X925 Cores + 4 Cortex-A725 Cores.
It would be useful if you includes the area of the core. More throughput per unit area is another way to judge the chips. A smaller area may allow an extra processor core
If u compare Gary's M4 vid u can see lot of similarities with the Arm v9.4 architecture. Of course Apple does it's own in house improvements besides that, but Arm architecture is clearly not that far behind. Arm is on the rise as I was predicting decades ago and it will eventually capture PC and console market while x86 will HW emulated if needed or u can buy legacy x86 HW if needed.
Naming were so smooth like X1, X2 etc, & suddenly they jumped to X925 😒 Looks like the concept of power efficiency cores are now gone for flagship premium SoC
It seems to be so that it correlates with the naming scheme of a520 and a725 and then the big core being x9xx kind of like intel i9/ryzen 9 naming scheme but I wish it was just called x5
What are / do the graphs at 4:53 mean ? Also at 6:34 ? What's on the X and Y axis ? I assume it's something like X - cache size and branch prediction window, while on Y axis is how many cycles the core is doing nothing, waiting for a memory fetch ?
@GaryExplains Well-done and to the point explaination. That's exactly what I wanted to know. What was suppose to be X5 got renamed to x925 for political reasons.
Yes -- They are assembled by an OEM (such as MediaTek, Google, Qualcomm, Apple, Amazon, etc.), as ARM itself doesn't make any chips, and fabbed together in one or more dies.
It's a modular design, ARM makes a base framework of interconnect then ARM makes three types of cores that can fit into it. So in simple teams they've made a Lego city with streets and left open the blocks between the roads. They have made buildings that fit in-between, some are skyscrapers, some single family, you can mix and match how you want your city to be. Fill all the spots with skyscrapers for performance and high power draw if you want.
You are expecting multi score to be 7700 ? Why ? Even dimensity 9300 scores multi score closer to apple a17 pro. Like between 7200-7500. So, expecting a new core with higher clock speeds to be in 7700 would be weird. I expect it should be between 8500-9500 at 3.8ghz and if there us 4.2ghz tyen it should be closer to rumoured multi score of 8 gen 4 like between 9500-11000
As I said very clearly in the video it depends on the core configuration. The processors from MediaTek, Qualcomm, and Samsung all use the cores in different ways, and that will be reflected in the multi-core score. I thought that would be obvious.
your estimations are way off, the 9400 presumably will use 1 "x925" and 3x"X4" and 4x "A725" and the multithreaded perfomance would be around ~9200 points, but these are all estimations😁
@niveZz- Dimensity 9300 also used similar configuration of all high performance cores no a520 and then 8 gen 4 is also gonna be 8x high performance custom nuvia cores but at different clocks for some of the cores for efficiency
Qualcomm heavily promoted Nuvia cores, expecting them to outperform ARM designs. However, I now believe they will lag behind both Apple's A18 and the Dimensity 9400 in single-threaded performance and power consumption.The only edge they have is NPU performance and Machine learning.
Thanks for sharing! They do seem interesting! Hope you could get these chipsets in a mini PC that runs Linux and FreeBSD but I guess, I'll have to wait for X series, lol!
Looks like we may be getting to the hegemony of x86 for windows. Intel has needed this kick in the ass for 25 years. AMD took a long time to get to finally beating on Intel, but now with ARM, we are looking at really opening up competition for laptops!
-You are not going to see LLM AI code itself running on smartphones anytime soon and probably ever.- By their nature LLMs need access to large volumes of data so those tasks -will- should be handed off to "the cloud" with server machines with access to multiple Petabytes, or more, of data, multiple CPUSs & GPUs to perform those tasks to be sent back to the smartphone. Smartphones will indeed run other AI workloads that do not require as much data and require the faster response time and/or the need to run locally without a connection to the internet/"the cloud".
@@GaryExplains Then that LLM is not very large (or retrieves its data from the internet), nowhere near the scale of something like ChatGPT. You might get away with having only 200GB to store the model, but most phones, or Raspberry Pis, do not have that kind of capacity.
@@GaryExplains Also, attempting to train an LLM on a Raspberry Pi would probably take months or longer, if that is even possible, and require even more storage.
No one said anything about training. As for the rest of what you wrote, I have several videos about that and other LLM stuff here on this channel. LLMs are more than just ChatGPT.
@@wheelyflac yes. ARM is a CPU architecture. The CPU is what is built on top of it. An Intel Pentium in 2004 has the same architecture as an i9 14900k, being x86
@wheely4405 - You can either purchase premade cores designed by ARM to implement in your CPU or you can purchase a license that lets you design your own original cores that use the ARM ISA. Qualcomm is doing the latter with Oryon.
Obviously this has nothing to do with RISC anymore... No human being can keep track of 1500 instructions and the interplay thereof. Only advanced AI can fully utilize it!
They always first consult their customers. Example even after cortex 925 being more powerful, MediaTek only uses on 925 and resorting to 3 cortex X3 core on dimensity 9400. Dimensity utilizes 8 cores designs, of they would have been provided with larger cortex 925 like apple big core, they wouldn't be able to accommodate other cores then, bcz mobile CPUs are supposed to made on very small die size. Arm can increase performance of big cores only when MediaTek and Exynos tells them.
Call me when Cortex cpu give a good run for their money against zen 5 and arrowlake. Until then, meh. I'll keep my phone from 2014 for another 5 years thanks very much, all of you upgrading phones every year are the reason the earth is a cesspool.
Shout out to my buddy who is on the design team for the Arm X-series cores. Absolute legend.
👍🎉
Gary is on the design team?
@@NightWiz11 OP's friend
Where does he work? USA or France?
I think he works in Austin. X925 and X -cores are designed there
I can't wait for these new cores to be paired with Nvidia Blackwell GPU Cores for Windows on ARM.
I can imagine with the DynamIQ Shared Unit allowing for 10 Cortex-X925 Cores + 4 Cortex-A725 Cores.
Dude love that you do this and break it down so well
It would be useful if you includes the area of the core. More throughput per unit area is another way to judge the chips. A smaller area may allow an extra processor core
Arm didn't release that information, not that I have it and opted not to include it.
If u compare Gary's M4 vid u can see lot of similarities with the Arm v9.4 architecture. Of course Apple does it's own in house improvements besides that, but Arm architecture is clearly not that far behind. Arm is on the rise as I was predicting decades ago and it will eventually capture PC and console market while x86 will HW emulated if needed or u can buy legacy x86 HW if needed.
Naming were so smooth like X1, X2 etc, & suddenly they jumped to X925 😒
Looks like the concept of power efficiency cores are now gone for flagship premium SoC
It seems to be so that it correlates with the naming scheme of a520 and a725 and then the big core being x9xx kind of like intel i9/ryzen 9 naming scheme but I wish it was just called x5
Cuz it's Gen 9
@farthurf.5221 and so? Did you found the Gen 8 reference in Cortex-X1/X2/X3/X4..? If it wasn't necessery before, it's not necessary now too..
@@sbrader97 earlier it was Cortex A, Cortex R and Cortex M. Now ARM is XARM ?
Thank you for your video! 2900 Geekbench 6 score is exciting!
What are / do the graphs at 4:53 mean ? Also at 6:34 ? What's on the X and Y axis ?
I assume it's something like X - cache size and branch prediction window, while on Y axis is how many cycles the core is doing nothing, waiting for a memory fetch ?
Says in the titles. Missed predictions per kilo instructions. Less missed predictions
*Great explanation, hopefully my phone lasts so I can skip the next gen and upgrade to the X925 architecture*
Will this be used in Snapdragon Gen4 ?
No.
What is the reason for the 925 in the x core? Is this somekind of mid range armv9?
I don't understand. It is the Cortex-X5 but they have rebranded it now to fit a similar naming scheme as the other cores. 5xx, 7xx, and 9xx.
@GaryExplains
Well-done and to the point explaination. That's exactly what I wanted to know. What was suppose to be X5 got renamed to x925 for political reasons.
I'm glad my reply was useful but I thought I covered the naming in the video. Was it not clear?
@@GaryExplains
Didn't watch the video yet.
Is that architecture also important for the X Elite Oryon SoC predecessor?
Nope
Still not new small core.
Why?
I'm confused. Who decides what cores make up the system? Don't these need to be fabbed together in one die or are they assembled?
Each chip maker decides like MediaTek or Google. However Arm has a default configuration that it uses for testing.
Yes -- They are assembled by an OEM (such as MediaTek, Google, Qualcomm, Apple, Amazon, etc.), as ARM itself doesn't make any chips, and fabbed together in one or more dies.
It's a modular design, ARM makes a base framework of interconnect then ARM makes three types of cores that can fit into it.
So in simple teams they've made a Lego city with streets and left open the blocks between the roads. They have made buildings that fit in-between, some are skyscrapers, some single family, you can mix and match how you want your city to be. Fill all the spots with skyscrapers for performance and high power draw if you want.
looking very promising, lets hope phone makers can make the best of it.
Willl this arm cpu come to desktop market if yess, will we have a option to upgrade cpus like we can currently upgrade
Someone will probably make an single board computer with these cpu cores
@bonkerbanker I doubt we will see these CPUs in an SBC. Are there any SBCs with the Cortex-X4?
You are expecting multi score to be 7700 ? Why ? Even dimensity 9300 scores multi score closer to apple a17 pro. Like between 7200-7500. So, expecting a new core with higher clock speeds to be in 7700 would be weird. I expect it should be between 8500-9500 at 3.8ghz and if there us 4.2ghz tyen it should be closer to rumoured multi score of 8 gen 4 like between 9500-11000
As I said very clearly in the video it depends on the core configuration. The processors from MediaTek, Qualcomm, and Samsung all use the cores in different ways, and that will be reflected in the multi-core score. I thought that would be obvious.
ARM wiil be next NVDA? Hope ARM 10X
maybe apple first
I'm surprised there wasn't any information on the memory controller. I suppose that means that it wasn't improved much.
so.. the if the X, A etc and the 9, 7, 5 etc. in the names mean the same thing, why have both...?
MEDIATEK will utilize them best ❤
As cool as this is, software makers are probably going to waste this increased performance on telemetry or tracking.
your estimations are way off, the 9400 presumably will use 1 "x925" and 3x"X4" and 4x "A725" and the multithreaded perfomance would be around ~9200 points, but these are all estimations😁
sounds almost like a laptop chip
@niveZz- Dimensity 9300 also used similar configuration of all high performance cores no a520 and then 8 gen 4 is also gonna be 8x high performance custom nuvia cores but at different clocks for some of the cores for efficiency
Qualcomm heavily promoted Nuvia cores, expecting them to outperform ARM designs. However, I now believe they will lag behind both Apple's A18 and the Dimensity 9400 in single-threaded performance and power consumption.The only edge they have is NPU performance and Machine learning.
@@niveZz-the same was said about 9300 yet it prevailed. However I want them to go 4- x925 and 4-a725, wanna see everyone go all out in 2025 😈😈.
@@Nemesis_678 Overkill 🫰
Thanks for sharing! They do seem interesting! Hope you could get these chipsets in a mini PC that runs Linux and FreeBSD but I guess, I'll have to wait for X series, lol!
why cant they keep naming scheme unified? like apple.
Hey Gary will we ever surpass Apple in terms of single core performance?
Apple will launch its next-generation before...
3nm. Guess this all depends on the outcome of the pending war in Taiwan.
as usual, thanks Gary!
What we need is more RISC-V
Why?
@@GaryExplainswhy not?
Why not isn't a reason nor a motivation. You need a why.
@@GaryExplains Because Risc V. That's all the reason needed.
LOL that tells me that you don't know what RISC-V is!
I suppose we can expect new sbcs coming out with these chips, right? Still, great video as always and thanks for sharing your thoughts and knowledge.
I doubt there will be SBCs with the CPU cores. Are there SBCs with the Cortex-X1 even?
Nice to see ARM citing Geekbench. I hope the idea that iOS and Apple's design are privileged dies off.
M4 SoC broke the 3000 single core hurdle
Expected, apple always had it with single core
M4 is hitting 4000
@@Xevos701 Absolutely not
3800@@lekejoshua4402
@@lekejoshua4402it does but with the most ideal condition(best cooling one could think of)
Edit:on gb 6
Looks like we may be getting to the hegemony of x86 for windows. Intel has needed this kick in the ass for 25 years. AMD took a long time to get to finally beating on Intel, but now with ARM, we are looking at really opening up competition for laptops!
It's gonna cost you an ARM and a leg
Would love to see nvidia gpu cores on board
Great, just as I've bought 8Gen2 device
8 gen 4 is gonna be using the custom nuvia cores from x elite laptop chip these arm cortex cores will be used in dimensity 9400 and exynos 2500
fuck performance, i want power effiency improvments, even if performances takes a huge hit !
Then get a quad cortex Cortex A520 processor.
@@GaryExplains a520 is nothing but disgrace compared to apples little cores
Then get an iphone! 😜
WOW! Now that’s a performance claim.
-You are not going to see LLM AI code itself running on smartphones anytime soon and probably ever.- By their nature LLMs need access to large volumes of data so those tasks -will- should be handed off to "the cloud" with server machines with access to multiple Petabytes, or more, of data, multiple CPUSs & GPUs to perform those tasks to be sent back to the smartphone.
Smartphones will indeed run other AI workloads that do not require as much data and require the faster response time and/or the need to run locally without a connection to the internet/"the cloud".
What? I can run an LLM on a Raspberry Pi.
@@GaryExplains Then that LLM is not very large (or retrieves its data from the internet), nowhere near the scale of something like ChatGPT. You might get away with having only 200GB to store the model, but most phones, or Raspberry Pis, do not have that kind of capacity.
@@GaryExplains Also, attempting to train an LLM on a Raspberry Pi would probably take months or longer, if that is even possible, and require even more storage.
.
No one said anything about training. As for the rest of what you wrote, I have several videos about that and other LLM stuff here on this channel. LLMs are more than just ChatGPT.
Time to move to 0.8 nanometers.
What's arm gonna do now that Qualcomm makes its own cores
Nothing. Qualcomm still licenses arm
@gr-os4gd is oryon based on arm design too? I thought it started as a server chip , I thought that made it different bruh
it's still using the arm instruction set
@@wheelyflac yes. ARM is a CPU architecture. The CPU is what is built on top of it.
An Intel Pentium in 2004 has the same architecture as an i9 14900k, being x86
@wheely4405 - You can either purchase premade cores designed by ARM to implement in your CPU or you can purchase a license that lets you design your own original cores that use the ARM ISA. Qualcomm is doing the latter with Oryon.
Should've went with X5 and get rid of the A520 like Qualcomm is going to when its been trash for years 😂
Obviously this has nothing to do with RISC anymore... No human being can keep track of 1500 instructions and the interplay thereof. Only advanced AI can fully utilize it!
Sorry, I am a little confused, what as advanced AI got to do with this?
That mean 10% 15% better performance
36 percent is the best they can do? Was expecting 500% and 60% power reduction.
They always first consult their customers. Example even after cortex 925 being more powerful, MediaTek only uses on 925 and resorting to 3 cortex X3 core on dimensity 9400. Dimensity utilizes 8 cores designs, of they would have been provided with larger cortex 925 like apple big core, they wouldn't be able to accommodate other cores then, bcz mobile CPUs are supposed to made on very small die size. Arm can increase performance of big cores only when MediaTek and Exynos tells them.
❤
Still slower than apple with less cores. Android needs to catch up with code optimisations.
You think Geekbench for Android is less optimized than the version for iOS?
Call me when Cortex cpu give a good run for their money against zen 5 and arrowlake. Until then, meh. I'll keep my phone from 2014 for another 5 years thanks very much, all of you upgrading phones every year are the reason the earth is a cesspool.
Huawei …???
What about Huawei?
Hi 🙂
Intel is so fucked
Every year its up to around 30 something %...changing smartphones every year is a nonesense
In the future these things will be called: Cortex XARMZY4 980Z Goldilocks "i am your father" XXUWI 778