Discussion Mediatek SoC thread

[FlameTail]

New Mediatek Chromebook SoC coming:

Another new MediaTek ARM chip for Chromebooks is coming

For quite some time, we didn't see much movement in the new SoC arena from MediaTek. They first launched the now-named Kompanio 500 with the introduction of the original Lenovo Chromebook Duet and not too long after, we got the Kompanio 800 and Kompanio 1000 series. Those chips hit a handful of...

chromeunboxed.com

Say hello to the MT8189, a chip that is brand new to the Chromium Repositories and – at least according to the model number – should provide a performance bump over the Kompanio 838 (internally called the MT8188) we expect to see this fall in the Chromebook tablet named ‘Ciri’.

Where is Geralt?

 

[FlameTail]

Dimensity 9400 leak:

https://twitter.com/x/status/1819372792577679834

That would put it at about 67 TOPS for NPU.

 

[FlameTail]

Mediatek leak dropped:


It is very interesting that Mediatek (Fa Ge) was involved in the design of Cortex X925 (Blackhawk).

I am not extremely surprised. I watched Mediatek's Computex 2024 keynote, where the CEO hinted that they had a deep collobaration with ARM.

 

[FlameTail]

Matches the peak ST performance of the 8 Gen 3 at only 30% of the power.

30% higher peak ST performance than D9300.

The efficiency of the D9400 seems to be exceptionally good, if this rumour is to be believed.

 

[FlameTail]

The D9300's efficiency is ahead of the A17 Pro and 8Gen3 at higher wattage levels, but if you look at lower wattage levels, the script is completely flipped!

Is this the price they are paying for not having Cortex A5xx?


The difference is there in GB5 too, though it not so drastic.

 

[DZero]

I am thinking how efficient will be the D9400 and the D8400.

Also D7300 started to go Quad Big... so maybe the new formation will end like this:

D9400: X Cores (4) + A7XX Cores (4)
D8400: A7XX Cores (8)
D7400: A7X/A7XX Cores (4) + A5X/A5XX Cores (4)
D6400: A7X (2) + A5X Cores (6)

 

[FlameTail]

D9400 = X (4) + A7xx (4)
D8400 = X (2) + A7xx (6)
D7400 = A7xx (4) + A5xx (4)
D6400 = A7xx (2) + A5xx (6)

(Wishful thinking).

 

[DZero]

D9400 = X (4) + A7xx (4)
D8400 = X (2) + A7xx (6)
D7400 = A7xx (4) + A5xx (4)
D6400 = A7xx (2) + A5xx (6)

(Wishful thinking).

Indeed, I want it too, but to be realistic...

D9400: X Cores (4) + A7XX Cores (4) - The only ones who are expecting it to keep it

D8400: A7XX Cores (8) - The best case, another case would be A7XX (6) + A5XX (2), but maybe they want to bring a fully out of order configuration ditching the in order cores. Is to me the most breaking config they delivers.

D7400: A7X/A7XX Cores (4) + A5X/A5XX Cores (4) - Realistic too since there is the D7300. Maybe the next D7400 will be Quad + Quad with Quad GPU config.

D6400: A7X (2) + A5X Cores (6) - While I want to agree with your thinking, the cost would make them stick still with that config, but maybe moving at last to A78 cores.

 

[FlameTail]

Dimensity 9400 rumour:

For reference:

 

[FlameTail]

More leaks:

Level 3 tracing with BVH?

 

[DZero]

What I am more excited is the D8400, if is full 8 out of order cores, it would be a massive change in the game.

 

[FlameTail]

Dr. Ian cutress (TechTechPotato) interviews Mediatek;

 

[FlameTail]

Dimensity 9400 leak:

alps k6991v1_64 - Geekbench

Benchmark results for an alps k6991v1_64 with an ARM ARMv8 processor.

browser.geekbench.com

alps k6991v1_64 - Geekbench

Benchmark results for an alps k6991v1_64 with an ARM ARMv8 processor.

browser.geekbench.com

ARM Immortalis G925 MC12 GPU

 

[NostaSeronx]

Mediatek should be spoiling a new Helio G40~G60 range for low-end, within six months.
- 8x APXM-6200 (Supports the bare minimum for RISC-V Android)
- DXT-8-256 (GE8320 -> DXT; 64 FP32 FLOPs/Clock -> 256 FP32 FLOPs/Clock && 4ppc -> 8ppc)
- 6nm FinFET compact (FFC) (also called N6e™)

 

[DZero]

Mediatek should be spoiling a new Helio G40~G60 range for low-end, within six months.
- 8x APXM-6200 (Supports the bare minimum for RISC-V Android)
- DXT-8-256 (GE8320 -> DXT; 64 FP32 FLOPs/Clock -> 256 FP32 FLOPs/Clock && 4ppc -> 8ppc)
- 6nm FinFET compact (FFC) (also called N6e™)

No please, no. Despite RISC-V android idea is not bad, even more, is interesting, the low tier Helio G series should rest in peace. I don't want to see the same processor being produced over and over again.

Also, In order cores should dissapear for good.

 

[NostaSeronx]

No please, no. Despite RISC-V android idea is not bad, even more, is interesting, the low tier Helio G series should rest in peace. I don't want to see the same processor being produced over and over again.

Also, In order cores should dissapear for good.

I kept it to the smartphone-side, but it might also replace MTK MT9026/MT9025/etcMT902x/etcMT901x for Fire TVs.

MediaTek | MT9602 | High Performance 4K TV SoC

Designed to create highly competitive products, MediaTek MT9602 is a high-performance 4K TV SoC with HDMI 2.1a support, global HDR, AV1 & AVS2 decoders.

www.mediatek.com

https://www.mediatek.com/products/digital-tv/mt9612
^-- whatever this thing is... all of the A53/A55 stuff will be replaced by Imagination/Imagination 2025 SOCs.

In-order cores benefit more from clustered queues: 1.5x for InO vs 1.1x for OoO. (Tremont -> Gracemont -> Skymont stuff) So, I assume In-order will be here to stay.

 

[FlameTail]

Dimensity 9400 arrives in Geekbench.

vivo V2419A - Geekbench

Benchmark results for a vivo V2419A with an ARM ARMv8 processor.

browser.geekbench.com

The single core clock seems to be limited to about 2 GHz, which explains the measly 1500 score. Assuming linear scaling, it should score 2700 at it's peak of 3.6 GHz.

 

[jdubs03]

Dimensity 9400 arrives in Geekbench.

vivo V2419A - Geekbench

Benchmark results for a vivo V2419A with an ARM ARMv8 processor.

browser.geekbench.com

The single core clock seems to be limited to about 2 GHz, which explains the measly 1500 score. Assuming linear scaling, it should score 2700 at it's peak of 3.6 GHz.

Seems to almost line up with ARM's presentation about X925 performance. Maybe a bit under. Extrapolate to the 3.8GHz level they were referencing, it'd be about 100 points below. Assuming linear scaling.
From: https://forums.anandtech.com/thread...verse-µarchs-discussion.2611782/post-41242907

 

[FlameTail]

Dimensity 9400 tipped to be announced on October 9;

 

[FlameTail]

Dimensity 9400 hit Geekbench again, now with the full blooded scores. 2800+ single, nearly 9000 multi.

OPPO PKB110 - Geekbench

Benchmark results for an OPPO PKB110 with an ARM ARMv8 processor.

browser.geekbench.com

 

[jdubs03]

Dimensity 9400 hit Geekbench again, now with the full blooded scores. 2800+ single, nearly 9000 multi.

OPPO PKB110 - Geekbench

Benchmark results for an OPPO PKB110 with an ARM ARMv8 processor.

browser.geekbench.com

That’s more like it. Should be ARMv9 though, maybe GB doesn’t recognize it yet.

Also, I find it interesting that the X925 would have higher IPC than the Oryon core in that leaked OnePlus result. That’s pretty impressive, almost a 6% advantage.

 

[Tup3x]

That’s more like it. Should be ARMv9 though, maybe GB doesn’t recognize it yet.

Also, I find it interesting that the X925 would have higher IPC than the Oryon core in that leaked OnePlus result. That’s pretty impressive, almost a 6% advantage.

Yeah, well, in the end power usage is what matters. If Oryon can clock higher while using about the same, then they did a better job job but we'll have to wait and see. Anyway, interesting times ahead.

 

[FlameTail]

That’s more like it. Should be ARMv9 though, maybe GB doesn’t recognize it yet.

Also, I find it interesting that the X925 would have higher IPC than the Oryon core in that leaked OnePlus result. That’s pretty impressive, almost a 6% advantage.

Yeah, I always expected X925 would have a wee bit higher IPC than Oryon.

 

[FlameTail]

It seems this generation, Dimensity has less performance than Snapdragon. So Mediatek will have to score design wins for the Dimensity 9400 on the basis that it's cheaper and more efficient (?).

 

[ikjadoon]

Dimensity 9400 hit Geekbench again, now with the full blooded scores. 2800+ single, nearly 9000 multi.

OPPO PKB110 - Geekbench

Benchmark results for an OPPO PKB110 with an ARM ARMv8 processor.

browser.geekbench.com

That OPPO phone has been benchmarked a few times, it seems:

Geekbench Search - Geekbench

The highest 1T score is 2889, which is the same 1T score as the basically fastest 288V Lunar Lake run I'd seen.

OPPO PKB110 vs Micro-Star International Co., Ltd. Prestige 13 AI Evo A2VMG - Geekbench

Sometimes X925 is 25% slower and sometimes X925 is 25% faster. Ends up at an equal geomean with the 288V.