Discussion Mediatek SoC thread

[QuickyDuck]

Seems impressive

/watch?v=17hGvd9Lem4

 

[FlameTail]

Amazing show from Mediatek

 

[FlameTail]

[1] The new GPU architecture from ARM is looking great. I mean... Look at that Perf/Watt uplift over D9200😱. Goodbye Valhall.

[2] Seems like in the CPU department Mediatek put all the balls in the multi-core department. The Cortex X4 cores is clocked at 3.25 GHz and only has 1 MB L2 cache, whereas the SD8G2 has an X4 @ 3.3 GHz with full 2 MB L2.

[3] Seems like they are cheaping out on the CPU caches.

SD8G3 : 5 MB Total L2, 12 MB L3

D9300 : 3.5 MB Total L2, 8 MB L3.

Kinda understandable as the monstrous CPU configuration would indeed take up a lot of space.

What do you guys think?

 

[Henry swagger]

[1] The new GPU architecture from ARM is looking great. I mean... Look at that Perf/Watt uplift over D9200😱. Goodbye Valhall.

[2] Seems like in the CPU department Mediatek put all the balls in the multi-core department. The Cortex X4 cores is clocked at 3.25 GHz and only has 1 MB L2 cache, whereas the SD8G2 has an X4 @ 3.3 GHz with full 2 MB L2.

[3] Seems like they are cheaping out on the CPU caches.

SD8G3 : 5 MB Total L2, 12 MB L3

D9300 : 3.5 MB Total L2, 8 MB L3.

Kinda understandable as the monstrous CPU configuration would indeed take up a lot of space.

What do you guys think?

Apple will have to increase core counts again if mediatrk continues to improve the this pace 💻🤔

 

[capricornsunvoyager2]

As always, Geekerwan's reviews are on point!
I'm excited to see the outcome! Mediatek has shown a significant increase in efficiency on the CPU side, as demonstrated by the Geekbench 5 graphs. It's amazing that they are currently leading in efficiency, relying on stock ARM-cores and the N4 node, without using in-house cores or bleeding-edge nodes like Apple.

However, this advantage seems to diminish on Geekbench 6, which has adapted to the modern reality that regular SOC use doesn't rely on CPU horsepower as much as before. Instead, the whole SOC, including the GPU and NPU, seems to have a greater impact. As a result, Mediatek's advantage in the CPU front becomes much slimmer than I expected.

Regardless, it will be interesting to see how devices perform in real-life battery tests. But based on the CPU and GPU parts, it appears that Mediatek is the efficiency king!

A PD. Also, this speaks to the improvement of Mediatek on the front as implementors. In the past, QC vs. MT; even with the same stock cores you could see QC leading in performance and efficiency benchmarks.

 

[SpudLobby]

[1] The new GPU architecture from ARM is looking great. I mean... Look at that Perf/Watt uplift over D9200😱. Goodbye Valhall.

[2] Seems like in the CPU department Mediatek put all the balls in the multi-core department. The Cortex X4 cores is clocked at 3.25 GHz and only has 1 MB L2 cache, whereas the SD8G2 has an X4 @ 3.3 GHz with full 2 MB L2.

[3] Seems like they are cheaping out on the CPU caches.

SD8G3 : 5 MB Total L2, 12 MB L3

D9300 : 3.5 MB Total L2, 8 MB L3.

Kinda understandable as the monstrous CPU configuration would indeed take up a lot of space.

What do you guys think?

Honestly the perf/watt uplift over Qualcomm isn't really that great as even I'd have expected, probably partially because these cores aren't really at their ideal operating frequencies, and then they also don't have as much L3 or L2 cache (QC has 50% more L3 and twice the L2 on their prime X4, and then similar cache for a few of their A720's to MediaTek's X4's) - which makes a big difference to total SoC + DRAM efficiency.

In other words on the former part: in a laptop where clocks are a bit higher - and if they had the same amount of L3 and beefier L2's, I think if you compared pushing an 8 Gen 3 to the 9300 with higher clocks, the 9300 would blow it out as you inch towards the 10W point. But in a smartphone operating profile and with crippled L2 and L3 vs QC, the perf/watt just doesn't look as great.

 

[FlameTail]

DIMENSITY 9300 RETAIL VERSION REVIEW

 

[FlameTail]

This is ground breaking.

Standby battery life is on par if not better than the 8 gen 3.

Shows that the lack of little Cortex A5xx cores is not an issue.

This is unprecedented.

The era of the A5xx cores is over?

 

[hemedans]

This is ground breaking.

Standby battery life is on par if not better than the 8 gen 3.

Shows that the lack of little Cortex A5xx cores is not an issue.

This is unprecedented.

The era of the A5xx cores is over?

Where did you see standby?

 

[naukkis]

This is ground breaking.

Standby battery life is on par if not better than the 8 gen 3.

Shows that the lack of little Cortex A5xx cores is not an issue.

This is unprecedented.

The era of the A5xx cores is over?

Standby power use isn't affected whether small cores are small or big - they can be power gated to zero power use. What matters is low load power use - and there more efficient but slower cores won't always be most efficient - faster cores can race to idle faster and today we probably are already in situation where even phone's smallest cores should be ARM's intermediate cores instead of small cores - with faster small cores SOC can power gate itself to near zero power greater percentage of time and SOC power use go down. Mediatek might be on right track.

 

[soresu]

today we probably are already in situation where even phone's smallest cores should be ARM's intermediate cores instead of small cores

In all likelihood A5xx will be relegated to lower end streaming devices and cheapo phones in the coming years.

I'd be interested to know how successful Neoverse E licensing has been so far.

 

[SarahKerrigan]

In all likelihood A5xx will be relegated to lower end streaming devices and cheapo phones in the coming years.

I'd be interested to know how successful Neoverse E licensing has been so far.

My impression is that E1 was supposed to slot into a lot of networking silicon where A53 and A72 have been historically dominant - think Qoriq Layerscape, Marvell Armada, stuff of that ilk. The only places I've actually seen A65/E1 in the wild are in general-purpose subsystems in accelerators, though (namely EdgeQ in their 5G baseband ASIC, and an automotive AI company whose name I'm forgetting.)

 

[soresu]

and an automotive AI company whose name I'm forgetting

Yeah they only released an AE version of A65 so that makes sense.

 

[SarahKerrigan]

Yeah they only released an AE version of A65 so that makes sense.

I'm not actually sure that's true. They list the A65 and A65AE separately on their website and in documentation. For instance:

Cortex-A65AE | Multithreaded CPU for Safety-Critical Automotive Tasks

Arm Cortex-A65AE is the first multithreaded DynamIQ CPU for automotive and safety-critical tasks like ADAS. It offers high throughput, resilience, and DCLS for fault tolerance.

www.arm.com

Cortex-A65 | Multithreaded CPU with High Throughput and 8 Cores

Arm Cortex-A65 is the first multithreaded Cortex-A DynamIQ CPU, offering high throughput with dual-thread processing and up to eight cores, for sensor fusion and vision-based systems.

www.arm.com

They specify in a number of places that the straight A65 is for non-safety-critical uses, whereas the A65AE is safety-critical.

EDIT: Also I lied about the company using A65 for automotive AI. The company is SiMa for their MLSoC product, which has a cluster of A65 (not A65AE) for embedded vision applications - but it seems like automotive is only a secondary target market for them (after UAV and industrial use.)

 

[soresu]

I'm not actually sure that's true. They list the A65 and A65AE separately on their website and in documentation. For instance:

Cortex-A65AE | Multithreaded CPU for Safety-Critical Automotive Tasks

Arm Cortex-A65AE is the first multithreaded DynamIQ CPU for automotive and safety-critical tasks like ADAS. It offers high throughput, resilience, and DCLS for fault tolerance.

www.arm.com

Cortex-A65 | Multithreaded CPU with High Throughput and 8 Cores

Arm Cortex-A65 is the first multithreaded Cortex-A DynamIQ CPU, offering high throughput with dual-thread processing and up to eight cores, for sensor fusion and vision-based systems.

www.arm.com

They specify in a number of places that the straight A65 is for non-safety-critical uses, whereas the A65AE is safety-critical.

EDIT: Also I lied about the company using A65 for automotive AI. The company is SiMa for their MLSoC product, which has a cluster of A65 (not A65AE) for embedded vision applications - but it seems like automotive is only a secondary target market for them (after UAV and industrial use.)

So are they using the AE version or not?

 

[SarahKerrigan]

So are they using the AE version or not?

As I said in the quoted post, non-AE. Straight A65.

 

[FlameTail]

Dimensity 8300 is here!

 

[FlameTail]

Why are sub-premium chips still using only A7xx and A5xx cores?

Both the Snapdragon 7 gen 3 and Dimensity 8300 use last year's Cortex A715 and A510 cores. No Cortex X.

And it's still a 1+3+4 configuration. (With one A715 clocked higher as the prime core).
.

 

[FlameTail]

2021 : Snapdragon 778G = (1+3+4)(A78/A55)

2023: Snapdragon 7 Gen 3 = (1+3+4)(A715/A510)



2021 : Snapdragon 888 = (1+3+4)(X1/A78/A55)

2023 : Snapdragon 8 gen 3 = (1+5+2)(X4/A720/A520)



Flagship CPUs have grown wider while Premium CPUs have stagnated in width

 

[hemedans]

Throttle to 46% in just 2 minutes

Dimensity 9300 in vivo X100 Pro throttles after 2 minutes in stress test

As expected from a chip with four performance cores. As the new vivo X100 Pro gets into the hands of keen testers, we can finally get an idea of the...

m.gsmarena.com

 

[FlameTail]

Not surprising when it's pulling more than 10W at peak, which is far beyond the cooling capacity of the phone.

 

[Doug S]

Phones rarely run flat out for more than a few seconds in real world usage, so this is probably not that big of an issue.

 

[hemedans]

Phones rarely run flat out for more than a few seconds in real world usage, so this is probably not that big of an issue.

Majority of phones can sustain 60%+ of performance for more than 1 hour, for comparison Vivo X90 with Dimensity 9200 could sustain 74% in 60 minutes

But vivo X100 with D9300 manage just 46% in 15 minutes, probably it's going to be worse in 60 min.

 

[FlameTail]

Dimensity 9400 Leak:

Rumor: Dimensity 9400 CPU core config might be as follows.

Cortex-X5 MP3
Cortex-A730** MP5

** Tentative name; codenamed "Chaberton"

https://x.com/QaM_Section31/status/1745306121454420463?s=20

 

[FlameTail]

This thread dead or what?

I find the 3× X5 + 5× A725 setup of the D9400 interesting.

What does it tell us about Cortex X5 (BlackHawk)? Perhaps Blackhawk is so powerful that they shifted from a 4+4 to 3+5?

It's bizarre all the weird cote configurations in flagship mobile CPUs these days.

Back in the day 1+3+4 was the norm.

Now we are seeing 1+4+3 (8G2), 1+5+2 (8G5), 4+4 (D9300), 1+4+4 (Tensor G3), 1+5+4 (Exynos 2400).

Soon we are getting 3+5 !