That's fair. Just pointing out that that tweet's sentiment might be justified...
That's fair. Just pointing out that that tweet's sentiment might be justified...I won't argue. Arm IPC improvements in the last years are provable. Timelines are also public. I didn't say anything else.
Indeed: that Apple sued GW III prior to the Nuvia acquisition for leaving to found a new company is consistent w/ its general corporate mobster mentality (despite its squeaky clean public image from troves spent on skillful, virtue-signaling advertising.) It also successfully pulled the puppet strings behind many other corporate actors: when it sued Qualcomm, it was powerful enough to be able to convince OEMs to withhold royalties, ask Samsung to get aggressive in Korea, and convince the Obama administration to file a concurrent FTC complaint in its final hours. They successfully confounded the media's understanding of the distinction between a product and the specific accelerators used in its implementation (cellular network standards vs. modem + cell towers / Quake game vs. computer with voodoo card...) Other cement-shoes situations such as their gutting of GTAT, threatening to drop Dialog Semi. w/o an alternate supplier lined up, subsuming the modified Imagination GPU as their own w/o having reached licensing terms, and App store "Sherlocking" also come to mind. Qualcomm just had strong enough patents and good enough lawyers to make them blink.I guess time will tell if ARM was using underhanded tactics to prevent Qualcomm from releasing Oryon earlier as a means to prop up Apple and its own X925 core. One thing I’ll just note if this is the case, then Oryon v2 should be able to be released earlier. The reason being, the argument here is that the delay wasn’t for technical reasons but by litigation; and since Oryon v1 is now out, a quicker release for v2 shouldn’t be precluded as it’s timeline wouldn’t be shifted. This is an assumption surely, but it doesn’t seem unreasonable.
Qualcomm also could’ve acted earlier on their own as well when starting to design their own custom core. But it took Nuvia for them to start making moves.
Let’s not pretend Qualcomm is a saint Intel and Mediatek couldn’t produce 5G modems on the level of Qualcomm because of the patents Qualcomm has. Apple to go after GW3 was wrong and am glad the court found no case. Don’t forget, Qualcomm also excercises patent lawfare.Indeed: that Apple sued GW III prior to the Nuvia acquisition for leaving to found a new company is consistent w/ its general corporate mobster mentality (despite its squeaky clean public image from troves spent on skillful, virtue-signaling advertising.) It also successfully pulled the puppet strings behind many other corporate actors: when it sued Qualcomm, it was powerful enough to be able to convince OEMs to withhold royalties, ask Samsung to get aggressive in Korea, and convince the Obama administration to file a concurrent FTC complaint in its final hours. They successfully confounded the media's understanding of the distinction between a product and the specific accelerators used in its implementation (cellular network standards vs. modem + cell towers / Quake game vs. computer with voodoo card...) Other cement-shoes situations such as their gutting of GTAT, threatening to drop Dialog Semi. w/o an alternate supplier lined up, subsuming the modified Imagination GPU as their own w/o having reached licensing terms, and App store "Sherlocking" also come to mind. Qualcomm just had strong enough patents and good enough lawyers to make them blink.
As for Pegasus (Oryon V2) in Glymur, the first rumors that they were testing Hamoa came around spring of '23 with release summer '24, and that they're reported as having started testing in summer '24 for Glymur is consistent for release in 2H25. This is also consistent w/ the leaked Dell slide as well which puts the V2 release about a year and a quarter off of V1. That Dell was told V3 comes in only '27 seems to indicate an awfully long 2 year gap between V2 and V3 compared to 1 year and 1 quarter from V1 to V2 especially when Cristiano said they were full steam ahead with investment in the PC space (unless V1 had been delayed.)
Agree that phones are a game where sharp elbows are regularly thrown. However, the scale and intent of Apple's abuses is quite different. I generally think ARM/Qualcomm's eco-system enabling licensing model is much more fair than people give them credit for than say the App. store model with its very high fees for any and all transactions made on top or Intel's aggressive, decades long legal enforcement of an x86 duopoly with concepts like "contra-revenue" utilized when Intel had an engineering deficit designed to keep AMD at bay.Let’s not pretend Qualcomm is a saint Intel and Mediatek couldn’t produce 5G modems on the level of Qualcomm because of the patents Qualcomm has. Apple to go after GW3 was wrong and am glad the court found no case. Don’t forget, Qualcomm also excercises patent lawfare.
Yes Arm had to up the game, though it has to be seen how high X925 can be clocked in devices. IPC is not the whole story.X925 does have a larger than usual IPC increase...
View attachment 108692
if die size is around M4 then it might be true. It has be around 160-165mm2Dimensity 9400 is supposedly advertised as having 29 billion transistors. That is crazy.
For context, Apple M4 is advertised as having 28 billion transistors. There's no way Mediatek and Apple are counting transistors in the same way. If they are, then that would mean Dimensity 9400 is a bigger chip than Apple M4, which is absurd.
Dimensity 9300 = 140 mm² N4P
Apple M4 = 165 mm² N3E
Dimensity 9400 = ?? mm² N3E
https://www.pcmag.com/news/mediatek...m-with-ai-enhanced-dimensity-9400-mobile-chipThe Dimensity 9400 adopts MediaTek’s second-gen All Big Core design, integrating one Arm Cortex-X925 core operating over 3.62GHz, combined with 3x Cortex-X4 and 4x Cortex-A720 cores. This design offers 35% faster single-core performance and 28% faster multi-core performance compared to MediaTek’s previous generation flagship chipset, the Dimensity 9300. Built on TSMC’s second-generation 3nm process, the Dimensity 9400 is up to 40% more power-efficient than its predecessor, allowing users to enjoy longer battery life.
The 12-core Arm Immortalis-G925 delivers super immersive gaming experiences with up to 40% faster raytracing performance compared to the previous generation. The Dimensity 9400 also brings PC-level features to smartphones with opacity micromaps (OMM) support for realistic effects. The chipset’s powerful GPU also offers 41% peak performance boost with up to 44% power savings compared to the Dimensity 9300, allowing users to game for longer. Additionally, the Dimensity 9400 supports HyperEngine technology for super resolution and impressive picture quality, which is co-developed by MediaTek and Arm Accurate Super Resolution (Arm ASR).
Packing MediaTek’s 8th Generation NPU, the Dimensity 940feature0 boasts a number of industry firsts for exceptional generative AI performance; it is the first mobile chipset to offer on-device LoRA training, high-quality on-device video generation, and developer support for Agentic AI. To allow users to take advantage of the latest agentic and generative AI applications, the Dimensity 9400 offers up to 80% faster large language model (LLM) prompt performance while also being up to 35% more power efficient than the Dimensity 9300.
Apple has a really good Ray-tracing implementation. I have heard it being described as a Level 3.5, similar to Nvidia/Intel, and far better than Adreno, Immortalis or RDNA atm.Interesting even with the beefed up GPU the Ray tracing is no match for M4 in solar bay.
M4 does around 14000 points in solar bay.Apple iPad Pro 13 (2024) Review
Apple iPad Pro 13 (2024) Tablet review with benchmark scores. See how it compares with other popular models.benchmarks.ul.com
The only thing D400 GPU leads in vs M4 is GFX bench.
No 3Dmark WLE or SNL tests
D9300 | D9400 |
Cortex X4 1 MB L2 | Cortex X925 2 MB L2 |
Cortex X4 512 MB L2 | Cortex X4 1 MB L2 |
Cortex A720 256 KB L2 | Cortex A720 512 KB L2 |
8 MB L3 | 12 MB L3 |
They do upclock the X4 (15.8% higher) and A720s (20% higher).Interestingly, Mediatek has only upgraded the prime core in Dimensity 9400. The other cores are same as Dimensity 9300.
1 MB L2
2 MB L2
512 MB L2
1 MB L2
256 KB L2
512 KB L2
This is a rare situation. No SoC vendor is quick to using ARM's new Cortex A725 core.
Mediatek held off from upgrading to A725 in the Dimensity 9400.
Tensor G4 uses last year's cores anyway (X4/A720/A520). Tensor G5 might use A725, but that's not releasing until August/September of 2025.
Qualcomm has opted to use their Oryon cores in the soon to be revealed 8 Gen 4, so they are not implementing any of ARM'S new cores, including A725.
Samsung's Exynos 2500 might be using A725, but the rumour is that it's delayed. It won't be featured in the Galaxy S25 series which comes in early 2025, but only in the next Galaxy Z Flip/Z Fold, which is 2025H2.
Hauwei's recent Kirin SoCs have been using their custom Taishan cores, in place of Cortex X/Cortex A7xx. They only use Cortex A5xx as their little core.
D9300 | D9400 |
1 x Cortex X4 3.25GHz 1 MB L2 | 1 x Cortex X925 3.626GHz 2 MB L2 |
3 x Cortex X4 2.85GHz 512 MB L2 | 3 x Cortex X4 3.30GHz 1 MB L2 |
4 x Cortex A720 2.00GHz 256 KB L2 | 4 x Cortex A720 2.40GHz 512 KB L2 |
8 MB L3 | 12 MB L3 |
Okay now that GPU is impressive. Good job Mediatek and ARM. It’s peak in SNL is basically on par with M4.Some Chinese reviewera appear to have tested the Dimensity 9400;
View attachment 108986View attachment 108987View attachment 108988View attachment 108989View attachment 108990View attachment 108991
The quality is not very good; I'll personally wait for Geekerwan's review video.
But this caught my eye;
View attachment 108992
2500 points in Steel Nomad Light.
For comparison;
View attachment 108993
*M2.Okay now that GPU is impressive. Good job Mediatek and ARM. It’s peak in SNL is basically on par with M4.
That's also peak numbers for M3/LNL?It peaked at 24FPS, it would’ve stayed there if it had sustained cooling.
View attachment 108994
Those are peak numbers, yes.That's also peak numbers for M3/LNL?
If so, yeah D9400 is impressive.
CPU energy efficiency does not look very good. The A18 Pro wins here by a long shot. And in general, the new X925 core is not very impressive.Some Chinese reviewera appear to have tested the Dimensity 9400;
View attachment 108986View attachment 108987View attachment 108988View attachment 108989View attachment 108990View attachment 108991
The quality is not very good; I'll personally wait for Geekerwan's review video.
But this caught my eye;
View attachment 108992
2500 points in Steel Nomad Light.
For comparison;
View attachment 108993