Mediatek is IMO the most innovative when it comes to ARM core type choices. I went with a phone with a Dimensity processor after seeing the genius they did with the Dimensity 8000 series. Here they also beat Qualcomm by choosing older but more efficient arm-architectures. Genius! The phone's software makes me hesitant to do this again but that isn't MediaTeks fault to be clear
For someone who emulates game consoles on their phone, the gpu is the real issue. CPUs on the high end SOCs have been good enough since sd866 and great since sd8g2.
I'm on the laymen side when it comes to mobile and arm chips, always had the idea that Qualcomm = good and Mediatek = bad, so this video was a huge eye opener to me. The algorithm did good recommend this!
Mediatek used to be really bad like the old Helios models, but since 2020 when they launched the new dimensity 9000 and 8000 series they have been head to head with Qualcomm in SoC performance
MediaTek have themselves to blame for this aversion in the Custom ROM community by failing to release source code in an open and timely manner. To date, Custom ROMs have been the ONLY way to maintain software support for more than a few years. But with both Samsung and Google now offering essentially lifetime software support for devices, manufacturer support now arguably exceeds that of the Custom ROM community. Perhaps for this reason, Custom ROM appears to be being actively targeted (for killing off) by Google. Both MediaTek and Qualcomm both need to pull up their socks on long term chip support. I for one, will not consider a device with less than 5 years software support (that used to mean choosing a device with known/assumed excellent Custom ROM support, therefore Qualcomm) and at the moment the vast majority of devices packing chips from those vendors do not have such support (not even the high end devices).
Wide cores(big cores) with moderate clock speeds are more efficient than small(efficiency cores) at high clock speeds, they also perform better. That's why the K8 from AMD and M series from Apple are/were so great. But that makes them a lot more expensive.
So now Mediatek only needs to provide long updates. They have(had?) a reputation of not pushing drivers upstream. Devices are so fast today that they are perfectly usable for a long time.
They also had reputation of having crappy SoCs in general - their SoCs had trouble with reading touchinput up to the end of life of Helio SoC line. Crazy.
Trading die area for efficiency and perf to compete with Qualcomm is genius. And they’ll still be a better deal. Surprised no one else has tried to compete this way. I wonder if this is the trick apple uses. I’d kill for a deep dive on their architecture to know what they’re doing so differently
It absolutely is, Apple's both efficiency and performance cores are gigantic compared to rest of the industry's contemporary cores from what I understand. And they move heaven and earth in search of single core performance at any cost (apart from power efficiency of course). And yep Apple clocks their cores really low as well (pretty sure larger cores automatically forces a lower clock because timing issues).
@@damianabregba7476 I could see hobbyists pushing them to help with an open source Linux driver. If they did that, they’d explode in popularity outside of phones aswell
I always felt like the A5... Series just exist to padd out a SoC so you have a Octa Core on paper. I feel like starting with Cortex A76 the gulf between big and little cors got so large that in relative terms they don't much value.
Well, they are in-order cores, similar to 2008 Intel Atom cores, having 4 cores for "background tasks" is ridiculous, 2 cores is more than enough, so they just used a500 cores to fill the octa core marketing as you said.
About those SRAM power reductions, the actual amount saved (4.6%) is lower in the because all those savings by the 12T cells are all in dynamic power. I’m not quite sure how number of fins switching fits into the dynamic power equation for transistors (although I think it should be linearly related). Regardless, while 12T sram cells save more dynamic power, it will also consume 2x more leakage power, by virtue of having 2x the transistor count. Dynamic power is also only consumed when the transistor is toggling, so toggle rate is highly important. All this is to say that, Mediatek made a wonderful decision to use 12T cells in the L1$, but the complete power equation is so much more complex that a 60% - 70% reduction in power of a single component only equates to 4.6% in the grand scheme of things. Also, thanks for pointing out that paper, lots to learn for people like me
That's fair - 2x transistors but half the SRAM cell density so, it works out neutral. (not quite vlbit you know what I mean) but you don't design SRAM size in a vacuum, so it's 'all things considered'
Great analysis! $25 for this kind of efficiency improvement is a steal. It'd be interesting to see how Google's Tensor SoCs look like in the efficiency comparison. The raw geekbench scores certainly paint an unfortunate picture. And it's crazy to see how strong off-the-shelve ARM cores an be in ST with the Dimensity 9400.
Great that you've remembered and referenced my question from the Podcast regarding the innovation in the industry that needs to happen. :) By the way, you fit perfectly into that PC World studio.
@@afif4738 Sure, but there's nothing a used 300$ pixel does that a brand new 200$ Xiaomi Redmi Pro can't do just as well. Talking for my use case, of course. The moment you don't care about the camera being more than good enough and don't use the phone for gaming, paying more than 200$ feels like throwing away perfectly good money.
Fr got the galaxy m31 for 225£ cuse of it 6000mah in 2020 second hand still holds great battery, good screen good camera, I'll probably have it for another 4 years and it will still hold full day charge(I only charge up to 85% for health)
This discovery of the hidden power efficient cache detail is soo good, good investigative work Dr Cutress. I was waiting for an 8gen4 chip phone (most likely motorola, I am stuck on their really nice gesture shortcuts and usb 3 display output support, apart from Samsung most other brand don't give a usb 3 port on even high end phones, and anything less than flagship level chips don't have support for multiple displays to begin with) to upgrade my still perfectly working fine 870 chip motorola phone ( I want AV1 decode). Will consider the 9400 phones from now as well.
The S25 FE could get the 9400. The X200 series from Vivo has just been released with the 9400, however it is likely that if a X200 Pro Plus comes later it will have an 8 gen 4 rather than a Dimensity, but who knows. The only thing I would like to know more about is the modem efficiency of the 9400 vs the ones in the 8 gen 3 and upcoming 8 gen 4.
@@petouser Let's hope so! Never tried Xiaomi - not a big fan of their OS customisation. But their price to benefit ratio is almost always insanely good.
Hi Ian, glad to have met you and your wife while you were in the UK. Great that you finally put out a video on your phone, was hoping you'd get around to explaining why efficiency cores aren't necessarily that efficient power wise! :D
It is mostly just a55 and a5xx being exceptionally bad, maybe it is because arm holdings keep these in order. Its hilarious how much better the overall efficiency gets when they are ditched for the a7xxs, they are larger in scale but just better
The 9400 and 8 Gen4 competition is gonna be crazy, both look like absolute monsters while being the most efficient of their lines ever. 1x3x4 vs 2x6 comparison is going to be fascinating as well.
Typing from my mediatek 9300+ device, picked that smartphone just because of this chip, amazing performance and incredible battery life. It's the redmi k70 ultra i imported, and it didnt cost me 1000 euros but 350. An amazing deal all around
Great stuff, the 9400 is looking incredible! I actually just bought a Galaxy Tab S10 Ultra that has a 9300+ in it. Predictably, the internet is full of hate that it doesn't have the 8G3 in it, despite the 9300+ being the better chip, typical. The only real criticism you can leverage at Samsung is not delaying the launch by a couple of months to wait for the 9400, but I'm sure this chip will be speedy and efficient for years.
I think the bigger issue is that it simply didn't offer enough compelling reason to upgrade over the S9, so they should've delayed it and put either the 8Gen4 or the 9400 in. Or alternatively they could've introduced a "Pro" line that features the X Elite chips, especially if they could've offered a hybrid Android/Windows option. At least then it might would be worth the price they're asking (and yes, Apple's pricing on the iPad Pro line is also insane, but would potentially make sense if it was letting you get 2 devices in one, replacing say a Macbook Air and iPad by letting you run both tabletized iOS and MacOS).
@@Emoziga amd is the only manufacturer that may be willing. They already kind of are. Maybe not necessarily for their mobile counterparts though. Which of course are the adreno gpus on the Snapdragon socs
@@Emoziga probably not. The thing is, you may be able to find some adreno drivers floating around. But never find mali drivers. Particularly because mediatek used to release only MEDIocre processors paired with terrible gpus, and so no one paid any attention to reverse engineering their drivers. I mean, some people are trying with the panfrost/panfork but otherwise nah. Whereas, tegra and adreno have consistently provided good performance in their release time. And were popular for hardware accelerated applications, gaming, etc. So, people gave a dayumm. And added to that, partial open sourcy nature of amd, caused significantly more attention towards driver development of adreno gpus than mali. Also, Qualcomm are probably collaborators with Amd and can't open source anything. They're manufacturers, arm is not an open source architecture, they're just licensing it (the architecture) to sell processors based on that architecture. And they're usually better than other arm soc manufacturers. Except maybe apple.
I actually bought a Vivo X100 Pro earlier this year with the same SoC. The 9300 is a beast and I would've kept it as my main phone if it wasn't for the fact that I ended up with the Chinese model, which has a lot of OS issues compared to the International model.
That's my gripe with the Find X7. Still too much Chinese stuff I can't get rid of. Find X8 is getting an international model, so waiting patiently for that. Don't think Vivo has announced an international model for the X200 yet
@@TechTechPotatoI had to go back with my tail between my legs and purchase an S23 Ultra, so I think I'll be sticking with that for the foreseeable future, but yeah, I hope more Mediatek SoCs come out in the west because their designs look very impressive and is a welcome change from the regular Qualcomm/Apple duopoly we've seen at the high end of mobile SoCs.
I'd be interested in the area increase of L1 cache with 12T and whether that area when used in additional 6T L1 cache capacity leads to an overall system power reduction and what that delta is between the two options
Great stuff. I'm definitely eyeing a Dimensity 9400 powered phone next. Currently running three with different variations of the 7300 and I feel it's time for me to try the top end unit.
Awesome Video, The cache part , I never knew and I REALLY REALLY Appreciate it now, I will search more on the topic . I'm just saying chip makers shouldn't shy away from giving us the nitty gritty details, we love it Thank YOU for this video 👍
@@AkilanKamarajan i don't think intels e-cores are a good example here as they are smaller, cheaper and more efficient. I think with Metero-Lake it was about 4 e-cores had just about the same die-footprint as 1 p-core while offering about 50% more performance. They just have way worse latency.
Love the video, but can you expand a little on why 12T SDRAM is more power efficient than 6T? What I know about electronics would fill a very short and empty book indeed, but intuitively, I would have assumed "more circuits = more states to change = power consumption". Wondering what is actually happening to reduce the power consumption.
In a 6T design, read or write, youre using most of those transistors every op. In a 12 T design, you end up using fewer per read and fewer per write in a less dense arrangement, so fewer active and lower parasitics. You can also optimize the ones used in read only vs write only. In 6T, all have to do all, so there are restrictions. If you think at the circuit level, it may sound counterintuitive, but at the materials and silicon level, everything is optimized.
The thing is that there are compromises. the 12T cell requires less energy for a state-change but has higher standby power. So with high read/write loads it is efficient, idling it sips power.
@@TechTechPotato @ABaumstumpf thank you both for the explanations. For someone in my industry (infrastructure & security) it is embarrassing how little I understand at the PHY level.
I wonder how much of the little/5-series cores doing so poorly is due to lack of investment into their design; IIRC before ARM v9 and the A510, the A55 had been several years old but many chips still used the A53 in part due to their lower licensing costs (and possibly Area). This could have disincentivized ARM from putting in as much work as they do into their more profitable X and Neoverse cores, resulting in an underwhelming improvement. However, I understand that this might also be the inherent limitations of in-order execution, kept for Area reasons.
in the beginning you said the A50 cores were efficient. Then you show the graph at 5:50 and say they're not efficient. Why not call them "low power" cores?
At my Memory Technologies class we talked about the 6T, 8T and other types of SRAM cells. I've always wondered where they are actually used. I was under the impression that the default SRAM cell that Intel, AMD, Apple, Qualcomm and others use is still the 6T. Awesome to see that Mediatek actually managed to make a good use out of these more exotic SRAM cell designs. It makes me think that there are really capable and innovative engineers at Mediatek.
I'm in the firm belief that the actual passionate and talented engineers go to the lesser companies and slowly raises them up until the cycle begins again. It's more interesting and lucrative to work for the underdog. It's happened with both AMD and Mediatek.
I've thought for a long time that chipmakers should forgo the A5x cores in favor of the even lower performance A3x cores - 4 * A76 + 2* A35 would sip power. All smartphones are fast enough now that I don't care about performance anymore. A midrange Snapdragon from 5 years ago is still fast enough. Everything else matters so much more, software, display, and camera quality being top of list, I hope MediaTek's ISP is as good as Qualcomm's.
I would consider Apple SoCs “all-big” as well. Unlike ARM’s A500-series cores, Apple’s efficiency cores are out-of-order and provide much better performance. Consequently, they are not solely the low-power standby cores, but also contribute significantly to the SoC’s multi-core performance. p.s. I have spent some time with devices powered by both the D9300 and the S8G3 (from the same manufacturer as well). Surprisingly there isn’t much of a performance gap despite the D9300 having a huge advantage on paper, and the battery life seems ever so slightly better on the S8G3 device as well. It appears that while MTK invested more into higher performance IPs, Qualcomm’s experience in silicon optimization did a lot to narrow the gap.
after seeing the Geekbench ranking of chips graphed it makes me really appreciate the performance in a phone chip now just wish I could do desktop computing on it instead of needing to interface with tiny touchscreens and touch optimized app designs.
There are attempts at it, the problem is android isn't really a great OS in general for general productivity. High end Samsung and Motorola phones have a desktop mode (dex and ready for respectively) that enables a desktop like experience if you connect an external monitor, mouse and keyboard. Can do web based tasks from the mode quite easily like email and research. Rest of the stuff however isn't so good.
This sounds like a great approach to chip design, targeting real performance rather than what seems almost like an exercise in on-paper min-maxing for most conventional designs these days. That L1$ thing seems like an absolute no-brainer regardless of the core design chosen though, given how relatively small those caches are. Doubling the transistor count for 32-48k of cache per core doesn't sound like it'll matter much on the scope of a whole SoC with billions of transistors.
Ironically all cores have a small die area nowadays. Back in the day a desktop quad core would be 250-300mm^2, now an octa-core zen4/zen5 is tiny 71mm^2 with a 100mm^2 or so I/O die. Or intel's 8p+16e is 257mm^2 with an iGPU being part of it. While compared to old ARM chips, all cores are big.
@@panjak323 That's not exactly true. They could have made the cores wider and kept the overall mm^2 per core the same. But they wanted to maximize profits by lowering both R&D and production costs.
@@rattlehead999 But making the core mindlessly wider has diminishing returns. There many more factors at play... Especially the limits of decoding, caching, and superscalar efficiency. Also you don't get higher efficiency with having simply wider cores, because for it to increase any performance, the larger chip area has to be active - drawing more power.
@@panjak323 yes, true, but they always said that there are diminishing returns, until suddenly there weren't, for ages people said that sandy bridge to skylake were the limit of how wide a core can go, then boom Apple and AMD have a wider core and suddenly intel can also have a wider core... oops. And yes, simply making the core wider doesn't make it more efficient, but making it wider while lowering clock speed does, because frequency doesn't scale linearly with power and vice versa. There is a sweet spot for each node where it achieves the best clock per watt and in the past 5-6 years AMD, intel and Nvidia have pushed way beyond it.
@@rattlehead999 UA-cam deleted my previous 2 comments, just because I tried to share link to imgur with measured perf/power curves of 8G2 cores. Either way look up geekervan. The frequency scaling is only true, when the leakage current is negligible.. X3 for example uses 0.6W at 600MHz and is 30% slower vs A715 at 1.2 Ghz with the same power draw. The large cores are only more power efficient within some optimal efficiency point and much less efficient everywhere else. Even then the efficiency gains are minimal, because performance scaling is barely linear (you can only do so much to speed up inherently sequential tasks), whilst the power scales with area, not just with frequency. GPUs make much more sense, because of the smaller cores, and much better parallelism.
From a die area perspective it looks like you could fit about 6 to 8 A510 cores in the space of those 4 A720 cores. For a multithreaded load doesn't that mean you're going to get more performance from a given die area out of the A510 cores? Does it make sense to look at the cores on an individual basis? Maybe the better way to do an efficiency comparison should be on performance and energy use over a certain die area rather than on a core basis. Die size matters because it's not always about cost saving, it can be about fitting more cores into the product. And generally speaking from performance perspective, many small die size E cores wins out against a fewer count of big cores.
"For a multithreaded load doesn't that mean you're going to get more performance from a given die area out of the A510 cores?" No. 2 A510 still offer worse performance than 1 A720. "Does it make sense to look at the cores on an individual basis? " Yes - cause the little cores are NOT efficient cores but rather the opposite. a single A510 needs the same or more power than an A720. "Die size matters because it's not always about cost saving, it can be about fitting more cores into the product." That is absolutely only about cost. "more cores" is an irrelevant metric. They dont give higher performance not better efficiency. it is only about having a slightly smaller die - so lower cost (and the licensing-cost for the A5 is also lower). "And generally speaking from performance perspective, many small die size E cores wins out against a fewer count of big cores." That is the case for Intels e-core, but absolutely not for the ARM architecture.
@@ABaumstumpf Hmm yes it appears in Intel terminology a A510 would be more comparable to a LPE core infact worse thatn that as it lacks OOOE and is only 3 decode wide - I am quite surprised to find it's actual numbers are about 30% the performance of a A720. It is far worse than I assumed.
Have you even compared it to the battery life of newer Snapdragon chips? In theory it all sounds great, but I have never seen an Android phone really outlasting the newer Samsung phones with new Snapdragons in terms of battery life. They also do a ton of Software optimization and chinese phones don't last as long with the same chips. Older Samsung phones are really shit because of using Exynos chips. And in terms of gpu performance Qualcomm is completely unmatched. I don't want to rant here, but it is always people comparing older phones to these newer chips and being impressed, while it was all because Exynos was just so bad in the past. I would not be shocked if this Mediatek chips in fact is not more efficient than the newer Snapdragon chips, because believe me they are completely insane while delivering the highest performance possible and even competing with Apple chips and sometimes beating them. It is still impressive what Mediatek has achieved and is likely already far better than even the newer Exynos chips.
It's about time the smartphone SoC market moved off of those tiny A5xx cores. For fun I ran a prime number generator I coded on each of the cores of my SD 8G2 phone (through Termux). The X3 finished generating in 110 ms, the A710/A715 in 180 ms, and the A510 in 550...
Hi, the efficiency curves only show the efficiency of each core at the peak frequency of their operating range. It would be interesting to see the efficiency of the a5xx vs A7xx vs Apple E cores at their lowest operating frequencies. I would imagine this would be relevant for a lot of "screen-off" use cases like say receiving notifications or audio playback.
Yup, and I'd love it if android vendors were more flexible in letting the use set the frequency on the DVFS without having to do a ton of crap behind the scenes.
I know Geekerwan presents data on the efficiency curve for the cores in their videos. But they never clarify if they go all the way down to the lowest frequencies.
Is it possible to run SPEC on any Android phone (rooted if needed), or is it something done by the manufacturer or integrator? A very quick search and look at their website seems to integrate the latter, but I did see a forum post or two that implied it may be possible to compile and run on Android if one has access to a/the terminal. You dropped this video two days after I was looking for a benchmark suite that would report the amount of power consumed to do a set amount of work because I enjoy fiddling and want to see if my modifications to my phone's CPU behavior is actually improving efficiency, or if it's all in my head.
An "undertested" part of smartphones SoCs is the modem, I am curious to know whether the Qualcomm or MTK or Exynos modem is more efficient when on Wi-Fi and when on mobile data and on idle
The annoying thing is that it also depends on the antenna and tuning. One of my previous phones had a modem that had severe US issues, for example, which made travel really frustrating (and I go to the US a lot).
19:21 I agree with you, well I am following you so I am tech minded to begin with. Yet, even if I know about the Qualcomm IP mafia style, I still thought of Qualcomm as creating better products than MediaTek. If MediaTek wants to really make a dent in the high-end marker they ought to give real innovations juice to the marketing and show they are a real player not just a provider of good enough chips for subpar phones.
HI Ian, I really, really appreciated this video. I'm an EE undergrad (3rd year) really interested in compute and semiconductors and I love your content generally, but this video has helped pushed me even more to do my best to try and attend ISSCC 2025 next year!!! Like that is so cool!!!!! The devil is in the details and you never know what sort of nugget of information may come up and have you like whoa hold on!! I'm in Toronto so a trip to San Fran may not be the most economical, and a lot of the highly technical stuff will fly over my head, but I think it will be awesome and I will find a way to get there!!!! And when I do, I hope to run into you :) I actually will also need a replacement for my Galaxy S10 next year (i think I have about that left in it) and I am now convinced, at least for now, that a phone with a 9400 may be the way to go!! (Especially as a bit of a power user)
Intel E-cores showed pretty much the same thing, these kind of "effficient" cores are not really that efficient, but they're small. Could AMD have purposedly just cut a bit from Zen to make "c" versions, not making new tiny cores, in order to avoid this situation?
On one side, I really don't need more power in my smartphone. On the other side, if we want to be able to use our smartphone as a computer with usb-c dock, I will definetly not say no to more performance. I still like that concept. edit: I also don't say no to better battery life :D
Hi Techtechpotato, Could you make a similar showcase/deep dive on what parts of the Qualcomm Snapdragon 8 Elite that has made it able to equal the Mediatek Dimensity 9400 in (GeekBench 6) CPU power efficiency? :) Is it mostly that a large shared L2 cache is better than separate L2 $ pr core and a shared L3 $? If you could do the same for the GPU's that would be awesome! But I'm most curious about the CPU's and the strategies of Qualcomm vs Mediatek.
IMPRESSIVE! Mediatek's achievments here are quite the thing - hopefully Samsung and google will follow suit with similar designs soon. Glad to see that Taiwan has some very capable CPU-core designers, working for a home-grown business like Mediatek.
Mine's got a dimesnsity 8300. Quad a715's + quad a510's. One of the 715's run at 3350 while the other 3 run at 3200. Performance wise, couldn't have asked for anything better but I find the efficiency a bit lacking...
Cortex-A700 series cores running at around 1.00 GHz are better than Cortex-A500 series cores at around 1.25 GHz in both performance and power consumption. Implementation by OEMs may differ but generally it's like that.
You have to consider one thing in that graph is that compared to Apple SOC which only has 6 cores almost all android phones have 8 cores so when you compare those individual cores they might seem a little slower. So as to compensate for that I suggest you should watch the multicore scores of the SOC. And if possible multicore scores of big and little cores separately is that's possible.
WOW that was a really awesome video. Hope more companies adopt mediatek instead of sticking with Qualcomm. And if mediatek also joins the PC world, oh god it's gonna be fun
It's still big.little despite the fact that the A720 is a "big/medium/little" core design depending on use case. Dimensity 9300 has it at 2.0GHz efficiency mode making it in essence a little core, particularly given it's matched with 4 X4 cores which are considerably more capable. It's great that SOCs are being configured without traditional 50/500 little cores, but given the difference between X cores and 70/700 cores, you can't consider them anything but little in these configurations without 50/500 cores.
The L1 cache power saving is really cool. For the "all big cores" approach, it just seems to copy Apple's strategy all along: use more die size and the most expensive silicon on the latest process to get the best efficiency
Hadn’t really considered, but I update every 4 years to my wifes old phone which she updates every 4 years. So i’m constantly rocking a 4 year old phone until it’s 8 years old. I replace the battery but aside from that it’s always plenty usable the entire span
I wish there was a way to capture, at what frequency do these cores run the most efficient or at how much load. Doing the most work they can while consuming as little power as possible is far more useful than raw performance at the expense of energy and heat.
Smaller because they don't integrate the 5G, 4G and wifi modems into the SoC and instead rely on external connectivity chips. If integrated, then it could take up as much as 15% to 20% of the total area of the SoC. Sometimes, its size can even be bigger than the GPU or CPU.
Good call I wasn't considering that. So using your estimation, adding 20% die area for a modem would put it right in line with the 9400. In other words the point stands that it's interesting Apple's cache sizes are so much bigger than competitors while not having larger die area, especially given that SRAM scaling has slowed to a near standstill.
This has been a known solution for a long time for those who have been following Linux scheduler development. Doing things fast means less power usage.
It would be cool if the next Google Tensor takes this approach. I wonder if Intel is doing it with the the continous improvement of the e cores (I think AMD has done that though they did not go bigLITTLE)
Wow, that was a plot twist! I'm with you, 12T SRAM is where it's at. Still hankering for a 5G OLED respin of 2010s HTC Desire with a nicer camera. If you can't make it work with 2 cores, is it even worth doing?
The S24 FE launched with the Exynos 2400e SoC (X4 core is lower clocked and it might be a worse binned version of the normal version) worldwide. Would it be possible to get your hands on it? 👀 Appreciate the energy efficiency comparisons, as only power efficiency is measured by everyone else!
Beyond the CPU improvements, it's really hard to care about anything else. It really doesn't do any good to have a great GPU that is phenomenal in synthetic benchmarks but then has subpar driver support that makes it perform as if it is several generations older than it is.
@@TechTechPotato if you constrain yoursef to power yes, but why do you use bigger die area? to use for more performance even if overal power usage is the same the shorter but higher burst of power generate more heat
![[39] This Silicon Wafer has 100,000 Chips On It: Interview](http://i.ytimg.com/vi/pAEHBgrLYZQ/mqdefault.jpg)
![[38] We're not even at PCIe 6.0 Yet!](/img/n.gif)
I deleted facebook and my phone battery life is improved significantly)
I need to delete YT so I can get better battery life 😭
@@kotztotz3530 I deleted YT so I can get better life
**Fixed that for you**. 🤣
Facebook deleted my account and I never returned and my entire life improved significantly
I bet your life improved significantly as well!
How many screen hours were you spending on the app each day?
Mediatek is IMO the most innovative when it comes to ARM core type choices. I went with a phone with a Dimensity processor after seeing the genius they did with the Dimensity 8000 series. Here they also beat Qualcomm by choosing older but more efficient arm-architectures. Genius!
The phone's software makes me hesitant to do this again but that isn't MediaTeks fault to be clear
sure, but they should also allow you to flash their ROMs.
@@rattlehead999 Maybe, but that is not a problem for like 99.9 % of buyers.
MediaTek Helio's have beeb pretty sub par overall.
The Dimensity reboot however has been great with every major SoC design so far.
@@Purjo92 true, but it's a small thing they can do and won't cost them anything as far as I am aware.
For someone who emulates game consoles on their phone, the gpu is the real issue. CPUs on the high end SOCs have been good enough since sd866 and great since sd8g2.
I'm on the laymen side when it comes to mobile and arm chips, always had the idea that Qualcomm = good and Mediatek = bad, so this video was a huge eye opener to me.
The algorithm did good recommend this!
Mediatek used to be really bad like the old Helios models, but since 2020 when they launched the new dimensity 9000 and 8000 series they have been head to head with Qualcomm in SoC performance
Mediatek used to be terrible but their newer chips are good
MediaTek these days has a really good performance. Custom ROM support, on the other hand.... literally barely anyone bothers with MediaTek devices.
then again people who custom rom their phones are less than 1% of smartphone users, they see no need to appeal to a dying niche @@jamesbrendan5170
MediaTek have themselves to blame for this aversion in the Custom ROM community by failing to release source code in an open and timely manner.
To date, Custom ROMs have been the ONLY way to maintain software support for more than a few years.
But with both Samsung and Google now offering essentially lifetime software support for devices, manufacturer support now arguably exceeds that of the Custom ROM community.
Perhaps for this reason, Custom ROM appears to be being actively targeted (for killing off) by Google.
Both MediaTek and Qualcomm both need to pull up their socks on long term chip support.
I for one, will not consider a device with less than 5 years software support (that used to mean choosing a device with known/assumed excellent Custom ROM support, therefore Qualcomm) and at the moment the vast majority of devices packing chips from those vendors do not have such support (not even the high end devices).
Wide cores(big cores) with moderate clock speeds are more efficient than small(efficiency cores) at high clock speeds, they also perform better. That's why the K8 from AMD and M series from Apple are/were so great. But that makes them a lot more expensive.
Out of oder vs in order as well
@@TechTechPotato good point
But also in this case, a720 and a520 are both at 2 GHz
@@TechTechPotato
I'm genuinely gobsmacked at the thought of a modern CPU core on a high end product not having out of order execution.
@@TheSyntheticSnake yeah me too, I forgot that they actually still had in-order execution on their cores.
So now Mediatek only needs to provide long updates. They have(had?) a reputation of not pushing drivers upstream. Devices are so fast today that they are perfectly usable for a long time.
They also had reputation of having crappy SoCs in general - their SoCs had trouble with reading touchinput up to the end of life of Helio SoC line. Crazy.
Trading die area for efficiency and perf to compete with Qualcomm is genius. And they’ll still be a better deal. Surprised no one else has tried to compete this way. I wonder if this is the trick apple uses. I’d kill for a deep dive on their architecture to know what they’re doing so differently
It absolutely is, Apple's both efficiency and performance cores are gigantic compared to rest of the industry's contemporary cores from what I understand. And they move heaven and earth in search of single core performance at any cost (apart from power efficiency of course).
And yep Apple clocks their cores really low as well (pretty sure larger cores automatically forces a lower clock because timing issues).
Sad about driver support. Not much of a Mesa driver existing = limited x86 emulation/translation
@@aravindpallippara1577 it also helps that they are usually a node ahead of everyone else
The A18 Pro is 4 GHz, i.e. clocked higher than this D9400
@@damianabregba7476 I could see hobbyists pushing them to help with an open source Linux driver. If they did that, they’d explode in popularity outside of phones aswell
I always felt like the A5... Series just exist to padd out a SoC so you have a Octa Core on paper. I feel like starting with Cortex A76 the gulf between big and little cors got so large that in relative terms they don't much value.
Well, they are in-order cores, similar to 2008 Intel Atom cores, having 4 cores for "background tasks" is ridiculous, 2 cores is more than enough, so they just used a500 cores to fill the octa core marketing as you said.
About those SRAM power reductions, the actual amount saved (4.6%) is lower in the because all those savings by the 12T cells are all in dynamic power. I’m not quite sure how number of fins switching fits into the dynamic power equation for transistors (although I think it should be linearly related). Regardless, while 12T sram cells save more dynamic power, it will also consume 2x more leakage power, by virtue of having 2x the transistor count. Dynamic power is also only consumed when the transistor is toggling, so toggle rate is highly important. All this is to say that, Mediatek made a wonderful decision to use 12T cells in the L1$, but the complete power equation is so much more complex that a 60% - 70% reduction in power of a single component only equates to 4.6% in the grand scheme of things.
Also, thanks for pointing out that paper, lots to learn for people like me
That's fair - 2x transistors but half the SRAM cell density so, it works out neutral. (not quite vlbit you know what I mean) but you don't design SRAM size in a vacuum, so it's 'all things considered'
mostly the power reduction comes from reduction in nm and having better testing equipment to target certain voltage to operate at.
Great analysis! $25 for this kind of efficiency improvement is a steal.
It'd be interesting to see how Google's Tensor SoCs look like in the efficiency comparison. The raw geekbench scores certainly paint an unfortunate picture.
And it's crazy to see how strong off-the-shelve ARM cores an be in ST with the Dimensity 9400.
I'm looking forward to Google's next smartphone chip
Great that you've remembered and referenced my question from the Podcast regarding the innovation in the industry that needs to happen. :) By the way, you fit perfectly into that PC World studio.
Thanks! I'd love a space like that at home, but office space where I am is twice as expensive as my mortgage.
I live by the "never spend more than 200$ on a phone" mantra. Looking forward to experiencing these magical cores in 5 years.
you can get used pixel a series for that price. but you dont have to go that low. $300 will get you some used samsungs.
@@afif4738 Sure, but there's nothing a used 300$ pixel does that a brand new 200$ Xiaomi Redmi Pro can't do just as well. Talking for my use case, of course. The moment you don't care about the camera being more than good enough and don't use the phone for gaming, paying more than 200$ feels like throwing away perfectly good money.
Always buy old flagships. They always have better performance per dollar even years later.
@@enriquepadilla2542 exactly, i go a step further and then keep that used phone for a long time.
still rocking a used Galaxy S7 🤣
Fr got the galaxy m31 for 225£ cuse of it 6000mah in 2020 second hand still holds great battery, good screen good camera, I'll probably have it for another 4 years and it will still hold full day charge(I only charge up to 85% for health)
This discovery of the hidden power efficient cache detail is soo good, good investigative work Dr Cutress.
I was waiting for an 8gen4 chip phone (most likely motorola, I am stuck on their really nice gesture shortcuts and usb 3 display output support, apart from Samsung most other brand don't give a usb 3 port on even high end phones, and anything less than flagship level chips don't have support for multiple displays to begin with) to upgrade my still perfectly working fine 870 chip motorola phone ( I want AV1 decode).
Will consider the 9400 phones from now as well.
The S25 FE could get the 9400.
The X200 series from Vivo has just been released with the 9400, however it is likely that if a X200 Pro Plus comes later it will have an 8 gen 4 rather than a Dimensity, but who knows.
The only thing I would like to know more about is the modem efficiency of the 9400 vs the ones in the 8 gen 3 and upcoming 8 gen 4.
Xiaomi 15 will come with USB 3 as well.
@@petouser Let's hope so! Never tried Xiaomi - not a big fan of their OS customisation. But their price to benefit ratio is almost always insanely good.
Hi Ian, glad to have met you and your wife while you were in the UK. Great that you finally put out a video on your phone, was hoping you'd get around to explaining why efficiency cores aren't necessarily that efficient power wise! :D
It is mostly just a55 and a5xx being exceptionally bad, maybe it is because arm holdings keep these in order.
Its hilarious how much better the overall efficiency gets when they are ditched for the a7xxs, they are larger in scale but just better
The 9400 and 8 Gen4 competition is gonna be crazy, both look like absolute monsters while being the most efficient of their lines ever.
1x3x4 vs 2x6 comparison is going to be fascinating as well.
Yeah, especially since both are going for all "big" cores
Man, modern phones can play Crisis and weigh as much as a small cinderblock, what a time to be alive!
Production value slowly going up. Love it!!!
Typing from my mediatek 9300+ device, picked that smartphone just because of this chip, amazing performance and incredible battery life. It's the redmi k70 ultra i imported, and it didnt cost me 1000 euros but 350. An amazing deal all around
Thanks for the tip. May I ask you where have you bought the phone for 350e?
Great stuff, the 9400 is looking incredible!
I actually just bought a Galaxy Tab S10 Ultra that has a 9300+ in it. Predictably, the internet is full of hate that it doesn't have the 8G3 in it, despite the 9300+ being the better chip, typical. The only real criticism you can leverage at Samsung is not delaying the launch by a couple of months to wait for the 9400, but I'm sure this chip will be speedy and efficient for years.
I think the bigger issue is that it simply didn't offer enough compelling reason to upgrade over the S9, so they should've delayed it and put either the 8Gen4 or the 9400 in. Or alternatively they could've introduced a "Pro" line that features the X Elite chips, especially if they could've offered a hybrid Android/Windows option. At least then it might would be worth the price they're asking (and yes, Apple's pricing on the iPad Pro line is also insane, but would potentially make sense if it was letting you get 2 devices in one, replacing say a Macbook Air and iPad by letting you run both tabletized iOS and MacOS).
This design details increased my respect for mediatech.
They're just doing what Apple did 7 years ago with the A10, ditching the useless A5x cores. At least they did it before Qualcomm
Mediatek, please release a source code for you GPU drivers.
They won't 🗿
@@yasirrakhurrafat1142 then it's no buy
@@Emoziga amd is the only manufacturer that may be willing.
They already kind of are.
Maybe not necessarily for their mobile counterparts though.
Which of course are the adreno gpus on the Snapdragon socs
@@yasirrakhurrafat1142 So it was AMD that told Qualcomm to release the source code?
@@Emoziga probably not.
The thing is, you may be able to find some adreno drivers floating around.
But never find mali drivers.
Particularly because mediatek used to release only MEDIocre processors paired with terrible gpus, and so no one paid any attention to reverse engineering their drivers. I mean, some people are trying with the panfrost/panfork but otherwise nah.
Whereas, tegra and adreno have consistently provided good performance in their release time.
And were popular for hardware accelerated applications, gaming, etc.
So, people gave a dayumm.
And added to that, partial open sourcy nature of amd, caused significantly more attention towards driver development of adreno gpus than mali.
Also, Qualcomm are probably collaborators with Amd and can't open source anything.
They're manufacturers, arm is not an open source architecture, they're just licensing it (the architecture) to sell processors based on that architecture.
And they're usually better than other arm soc manufacturers.
Except maybe apple.
Are you the guy from anand who made those graphs for them? If so, it's amazing that you're continuing this great work. Subbed
I spent 11 years at AnandTech reviewing CPUs. There's over a thousand articles there with my name on
I actually bought a Vivo X100 Pro earlier this year with the same SoC. The 9300 is a beast and I would've kept it as my main phone if it wasn't for the fact that I ended up with the Chinese model, which has a lot of OS issues compared to the International model.
That's my gripe with the Find X7. Still too much Chinese stuff I can't get rid of. Find X8 is getting an international model, so waiting patiently for that. Don't think Vivo has announced an international model for the X200 yet
@@TechTechPotatoI had to go back with my tail between my legs and purchase an S23 Ultra, so I think I'll be sticking with that for the foreseeable future, but yeah, I hope more Mediatek SoCs come out in the west because their designs look very impressive and is a welcome change from the regular Qualcomm/Apple duopoly we've seen at the high end of mobile SoCs.
Passionately spoken article Ian!
I'd be interested in the area increase of L1 cache with 12T and whether that area when used in additional 6T L1 cache capacity leads to an overall system power reduction and what that delta is between the two options
Great stuff. I'm definitely eyeing a Dimensity 9400 powered phone next. Currently running three with different variations of the 7300 and I feel it's time for me to try the top end unit.
Awesome Video, The cache part , I never knew and I REALLY REALLY Appreciate it now, I will search more on the topic .
I'm just saying chip makers shouldn't shy away from giving us the nitty gritty details, we love it
Thank YOU for this video 👍
I always wondered with the big little cores...say 4 big 4 little ..why not 6 big cores but with better power and frequency balancing ?
Die area
@TechTechPotato ahh cheers so would two big cores be bigger than 4 smaller cores ?
Also multithreaded performance-
benchmark scores per die area ( the Intel E core way )
@@AkilanKamarajan i don't think intels e-cores are a good example here as they are smaller, cheaper and more efficient.
I think with Metero-Lake it was about 4 e-cores had just about the same die-footprint as 1 p-core while offering about 50% more performance. They just have way worse latency.
@@EastyyBlogspot because "only" 6 cores is not good enough for marketing
This is good to hear! I like it when a product is released and runs on it's merit, not hype.
Love the video, but can you expand a little on why 12T SDRAM is more power efficient than 6T?
What I know about electronics would fill a very short and empty book indeed, but intuitively, I would have assumed "more circuits = more states to change = power consumption". Wondering what is actually happening to reduce the power consumption.
In a 6T design, read or write, youre using most of those transistors every op. In a 12 T design, you end up using fewer per read and fewer per write in a less dense arrangement, so fewer active and lower parasitics. You can also optimize the ones used in read only vs write only. In 6T, all have to do all, so there are restrictions.
If you think at the circuit level, it may sound counterintuitive, but at the materials and silicon level, everything is optimized.
The thing is that there are compromises. the 12T cell requires less energy for a state-change but has higher standby power. So with high read/write loads it is efficient, idling it sips power.
@@TechTechPotato @ABaumstumpf thank you both for the explanations. For someone in my industry (infrastructure & security) it is embarrassing how little I understand at the PHY level.
I wonder how much of the little/5-series cores doing so poorly is due to lack of investment into their design; IIRC before ARM v9 and the A510, the A55 had been several years old but many chips still used the A53 in part due to their lower licensing costs (and possibly Area). This could have disincentivized ARM from putting in as much work as they do into their more profitable X and Neoverse cores, resulting in an underwhelming improvement. However, I understand that this might also be the inherent limitations of in-order execution, kept for Area reasons.
in the beginning you said the A50 cores were efficient. Then you show the graph at 5:50 and say they're not efficient. Why not call them "low power" cores?
Been waiting and hoping Mediatek can break into the big league for a while now. Looks like it might finally be happening by embracing big dies.
At my Memory Technologies class we talked about the 6T, 8T and other types of SRAM cells. I've always wondered where they are actually used. I was under the impression that the default SRAM cell that Intel, AMD, Apple, Qualcomm and others use is still the 6T. Awesome to see that Mediatek actually managed to make a good use out of these more exotic SRAM cell designs.
It makes me think that there are really capable and innovative engineers at Mediatek.
I'm in the firm belief that the actual passionate and talented engineers go to the lesser companies and slowly raises them up until the cycle begins again. It's more interesting and lucrative to work for the underdog. It's happened with both AMD and Mediatek.
the audio! finally dope! love it ... btw: 07:14 and the trade-off is: Diarrhea :D
I've thought for a long time that chipmakers should forgo the A5x cores in favor of the even lower performance A3x cores - 4 * A76 + 2* A35 would sip power. All smartphones are fast enough now that I don't care about performance anymore. A midrange Snapdragon from 5 years ago is still fast enough. Everything else matters so much more, software, display, and camera quality being top of list, I hope MediaTek's ISP is as good as Qualcomm's.
I would consider Apple SoCs “all-big” as well. Unlike ARM’s A500-series cores, Apple’s efficiency cores are out-of-order and provide much better performance. Consequently, they are not solely the low-power standby cores, but also contribute significantly to the SoC’s multi-core performance.
p.s. I have spent some time with devices powered by both the D9300 and the S8G3 (from the same manufacturer as well). Surprisingly there isn’t much of a performance gap despite the D9300 having a huge advantage on paper, and the battery life seems ever so slightly better on the S8G3 device as well. It appears that while MTK invested more into higher performance IPs, Qualcomm’s experience in silicon optimization did a lot to narrow the gap.
Apple SOCs are closer to Intel hybrid CPUs, in the die size of cores.
after seeing the Geekbench ranking of chips graphed it makes me really appreciate the performance in a phone chip now just wish I could do desktop computing on it instead of needing to interface with tiny touchscreens and touch optimized app designs.
There are attempts at it, the problem is android isn't really a great OS in general for general productivity.
High end Samsung and Motorola phones have a desktop mode (dex and ready for respectively) that enables a desktop like experience if you connect an external monitor, mouse and keyboard.
Can do web based tasks from the mode quite easily like email and research. Rest of the stuff however isn't so good.
Vivo X200 pro is already out in China, we should see some English reviews probably soon
This sounds like a great approach to chip design, targeting real performance rather than what seems almost like an exercise in on-paper min-maxing for most conventional designs these days. That L1$ thing seems like an absolute no-brainer regardless of the core design chosen though, given how relatively small those caches are. Doubling the transistor count for 32-48k of cache per core doesn't sound like it'll matter much on the scope of a whole SoC with billions of transistors.
Ironically all cores have a small die area nowadays.
Back in the day a desktop quad core would be 250-300mm^2, now an octa-core zen4/zen5 is tiny 71mm^2 with a 100mm^2 or so I/O die. Or intel's 8p+16e is 257mm^2 with an iGPU being part of it.
While compared to old ARM chips, all cores are big.
That's because of node improvements... The die area rougly halves on every major node (except for cache).
@@panjak323 That's not exactly true. They could have made the cores wider and kept the overall mm^2 per core the same. But they wanted to maximize profits by lowering both R&D and production costs.
@@rattlehead999 But making the core mindlessly wider has diminishing returns.
There many more factors at play... Especially the limits of decoding, caching, and superscalar efficiency.
Also you don't get higher efficiency with having simply wider cores, because for it to increase any performance, the larger chip area has to be active - drawing more power.
@@panjak323 yes, true, but they always said that there are diminishing returns, until suddenly there weren't, for ages people said that sandy bridge to skylake were the limit of how wide a core can go, then boom Apple and AMD have a wider core and suddenly intel can also have a wider core... oops.
And yes, simply making the core wider doesn't make it more efficient, but making it wider while lowering clock speed does, because frequency doesn't scale linearly with power and vice versa. There is a sweet spot for each node where it achieves the best clock per watt and in the past 5-6 years AMD, intel and Nvidia have pushed way beyond it.
@@rattlehead999 UA-cam deleted my previous 2 comments, just because I tried to share link to imgur with measured perf/power curves of 8G2 cores. Either way look up geekervan.
The frequency scaling is only true, when the leakage current is negligible.. X3 for example uses 0.6W at 600MHz and is 30% slower vs A715 at 1.2 Ghz with the same power draw.
The large cores are only more power efficient within some optimal efficiency point and much less efficient everywhere else. Even then the efficiency gains are minimal, because performance scaling is barely linear (you can only do so much to speed up inherently sequential tasks), whilst the power scales with area, not just with frequency.
GPUs make much more sense, because of the smaller cores, and much better parallelism.
From a die area perspective it looks like you could fit about 6 to 8 A510 cores in the space of those 4 A720 cores. For a multithreaded load doesn't that mean you're going to get more performance from a given die area out of the A510 cores? Does it make sense to look at the cores on an individual basis? Maybe the better way to do an efficiency comparison should be on performance and energy use over a certain die area rather than on a core basis.
Die size matters because it's not always about cost saving, it can be about fitting more cores into the product. And generally speaking from performance perspective, many small die size E cores wins out against a fewer count of big cores.
"For a multithreaded load doesn't that mean you're going to get more performance from a given die area out of the A510 cores?"
No. 2 A510 still offer worse performance than 1 A720.
"Does it make sense to look at the cores on an individual basis? "
Yes - cause the little cores are NOT efficient cores but rather the opposite. a single A510 needs the same or more power than an A720.
"Die size matters because it's not always about cost saving, it can be about fitting more cores into the product."
That is absolutely only about cost. "more cores" is an irrelevant metric. They dont give higher performance not better efficiency. it is only about having a slightly smaller die - so lower cost (and the licensing-cost for the A5 is also lower).
"And generally speaking from performance perspective, many small die size E cores wins out against a fewer count of big cores."
That is the case for Intels e-core, but absolutely not for the ARM architecture.
@@ABaumstumpf Hmm yes it appears in Intel terminology a A510 would be more comparable to a LPE core infact worse thatn that as it lacks OOOE and is only 3 decode wide - I am quite surprised to find it's actual numbers are about 30% the performance of a A720. It is far worse than I assumed.
Have you even compared it to the battery life of newer Snapdragon chips? In theory it all sounds great, but I have never seen an Android phone really outlasting the newer Samsung phones with new Snapdragons in terms of battery life. They also do a ton of Software optimization and chinese phones don't last as long with the same chips. Older Samsung phones are really shit because of using Exynos chips.
And in terms of gpu performance Qualcomm is completely unmatched.
I don't want to rant here, but it is always people comparing older phones to these newer chips and being impressed, while it was all because Exynos was just so bad in the past. I would not be shocked if this Mediatek chips in fact is not more efficient than the newer Snapdragon chips, because believe me they are completely insane while delivering the highest performance possible and even competing with Apple chips and sometimes beating them.
It is still impressive what Mediatek has achieved and is likely already far better than even the newer Exynos chips.
incredibly useful video in understanding what really matters in the new generation of mobile phone chips.
I switched from a tensor G3 to an a18 pro and the difference in power and efficieny was outstanding
G3 is really bad.
It's about time the smartphone SoC market moved off of those tiny A5xx cores. For fun I ran a prime number generator I coded on each of the cores of my SD 8G2 phone (through Termux). The X3 finished generating in 110 ms, the A710/A715 in 180 ms, and the A510 in 550...
Hi, the efficiency curves only show the efficiency of each core at the peak frequency of their operating range. It would be interesting to see the efficiency of the a5xx vs A7xx vs Apple E cores at their lowest operating frequencies. I would imagine this would be relevant for a lot of "screen-off" use cases like say receiving notifications or audio playback.
Yup, and I'd love it if android vendors were more flexible in letting the use set the frequency on the DVFS without having to do a ton of crap behind the scenes.
I know Geekerwan presents data on the efficiency curve for the cores in their videos. But they never clarify if they go all the way down to the lowest frequencies.
You also don't know how close the dvfs is to the Vmin guard rails, or if they're decoupled.
Is it possible to run SPEC on any Android phone (rooted if needed), or is it something done by the manufacturer or integrator? A very quick search and look at their website seems to integrate the latter, but I did see a forum post or two that implied it may be possible to compile and run on Android if one has access to a/the terminal.
You dropped this video two days after I was looking for a benchmark suite that would report the amount of power consumed to do a set amount of work because I enjoy fiddling and want to see if my modifications to my phone's CPU behavior is actually improving efficiency, or if it's all in my head.
What are those phones like when it comes to storage space, rootability, compatibility with third-party ROMs, types of sensors, cameras etc?
go to review channels, akk we care about here is chips :)
And now we have Samsung Galaxy Tab S10 Ultra with Mediatek SOC instead of Qualcomm.
the bar graphs for the Manhattan test are kinda hilarious
An "undertested" part of smartphones SoCs is the modem, I am curious to know whether the Qualcomm or MTK or Exynos modem is more efficient when on Wi-Fi and when on mobile data and on idle
The annoying thing is that it also depends on the antenna and tuning. One of my previous phones had a modem that had severe US issues, for example, which made travel really frustrating (and I go to the US a lot).
@@TechTechPotato That is something I wasn't thinking about
19:21 I agree with you, well I am following you so I am tech minded to begin with. Yet, even if I know about the Qualcomm IP mafia style, I still thought of Qualcomm as creating better products than MediaTek.
If MediaTek wants to really make a dent in the high-end marker they ought to give real innovations juice to the marketing and show they are a real player not just a provider of good enough chips for subpar phones.
HI Ian, I really, really appreciated this video. I'm an EE undergrad (3rd year) really interested in compute and semiconductors and I love your content generally, but this video has helped pushed me even more to do my best to try and attend ISSCC 2025 next year!!! Like that is so cool!!!!! The devil is in the details and you never know what sort of nugget of information may come up and have you like whoa hold on!! I'm in Toronto so a trip to San Fran may not be the most economical, and a lot of the highly technical stuff will fly over my head, but I think it will be awesome and I will find a way to get there!!!! And when I do, I hope to run into you :)
I actually will also need a replacement for my Galaxy S10 next year (i think I have about that left in it) and I am now convinced, at least for now, that a phone with a 9400 may be the way to go!! (Especially as a bit of a power user)
Ping me if you make it - if you reach out early to the conference and ask to be a student volunteer, you might get free entry or assistance.
Want to see that Spec graph colored by process node and foundry.
Looking forward to see your coverage from Hawaii.
Intel E-cores showed pretty much the same thing, these kind of "effficient" cores are not really that efficient, but they're small. Could AMD have purposedly just cut a bit from Zen to make "c" versions, not making new tiny cores, in order to avoid this situation?
"this phone is something I've been carrying for the best part of the year now"
Pulls out a turquoise purse lol
It is still boils down on how good the kernel support these cores and effeciently switch between modes.
One day we might get scheduling between cores purely on the microcode level, we can only dream.
is the 12 transistor cells implemented in all caches L1 L2 L3 or just the L1?
On one side, I really don't need more power in my smartphone. On the other side, if we want to be able to use our smartphone as a computer with usb-c dock, I will definetly not say no to more performance. I still like that concept.
edit: I also don't say no to better battery life :D
Ian, how is the modem on your dimensity? Are they as good now as Qualcomm modems?
I'd also be interested in the power consumption of the different cores at idle/ low utilization
Hi Techtechpotato,
Could you make a similar showcase/deep dive on what parts of the Qualcomm Snapdragon 8 Elite that has made it able to equal the Mediatek Dimensity 9400 in (GeekBench 6) CPU power efficiency? :)
Is it mostly that a large shared L2 cache is better than separate L2 $ pr core and a shared L3 $?
If you could do the same for the GPU's that would be awesome! But I'm most curious about the CPU's and the strategies of Qualcomm vs Mediatek.
Nice background. I can almost see Gordon and Adam beside you.
edit: 21:21 - oh, you're actually there...
I think the title should have been, 'Go big (cores) or go home."
IMPRESSIVE! Mediatek's achievments here are quite the thing - hopefully Samsung and google will follow suit with similar designs soon.
Glad to see that Taiwan has some very capable CPU-core designers, working for a home-grown business like Mediatek.
Chinesium phones are on a different level
Mine's got a dimesnsity 8300. Quad a715's + quad a510's. One of the 715's run at 3350 while the other 3 run at 3200. Performance wise, couldn't have asked for anything better but I find the efficiency a bit lacking...
Are you saying that the arm 500 cores take more energy for the same calculation as the 700 cores?
yes, and they're slower
Cortex-A700 series cores running at around 1.00 GHz are better than Cortex-A500 series cores at around 1.25 GHz in both performance and power consumption. Implementation by OEMs may differ but generally it's like that.
You have to consider one thing in that graph is that compared to Apple SOC which only has 6 cores almost all android phones have 8 cores so when you compare those individual cores they might seem a little slower.
So as to compensate for that I suggest you should watch the multicore scores of the SOC. And if possible multicore scores of big and little cores separately is that's possible.
Could this be used in laptops or tablets? I saw there are laptops and tablets on the site.
WOW that was a really awesome video. Hope more companies adopt mediatek instead of sticking with Qualcomm. And if mediatek also joins the PC world, oh god it's gonna be fun
It's still big.little despite the fact that the A720 is a "big/medium/little" core design depending on use case.
Dimensity 9300 has it at 2.0GHz efficiency mode making it in essence a little core, particularly given it's matched with 4 X4 cores which are considerably more capable.
It's great that SOCs are being configured without traditional 50/500 little cores, but given the difference between X cores and 70/700 cores, you can't consider them anything but little in these configurations without 50/500 cores.
The L1 cache power saving is really cool. For the "all big cores" approach, it just seems to copy Apple's strategy all along: use more die size and the most expensive silicon on the latest process to get the best efficiency
yes but also have more leakage, its a trade off
Dr. Ian
Thank you for your insights
Hadn’t really considered, but I update every 4 years to my wifes old phone which she updates every 4 years. So i’m constantly rocking a 4 year old phone until it’s 8 years old. I replace the battery but aside from that it’s always plenty usable the entire span
I wish there was a way to capture, at what frequency do these cores run the most efficient or at how much load.
Doing the most work they can while consuming as little power as possible is far more useful than raw performance at the expense of energy and heat.
@15:49 - what's the performance diff between X3 @3.35GHz-1mb-L2 vs X4 @2.85GHz-512kb-L2 ???
it's interesting to me that the A18 Pro is ~20% smaller than the 9400 and yet its cache sizes are massive by comparison
Smaller because they don't integrate the 5G, 4G and wifi modems into the SoC and instead rely on external connectivity chips. If integrated, then it could take up as much as 15% to 20% of the total area of the SoC. Sometimes, its size can even be bigger than the GPU or CPU.
Good call I wasn't considering that. So using your estimation, adding 20% die area for a modem would put it right in line with the 9400. In other words the point stands that it's interesting Apple's cache sizes are so much bigger than competitors while not having larger die area, especially given that SRAM scaling has slowed to a near standstill.
@@JKHYT they are full node ahead.
@@panjak323the node doesn't matter at this size.
Not integrating connectivity saves space for more cache.
@@panjak323 Which is why I mentioned the lack of SRAM scaling. Being a node ahead doesn't mean your cache is denser, not meaningfully.
I very much enjoyed this video, I would be interesting to see a little more content about the smartphone side of silicon from time to time
Does mediatek still violate gpl or do they publish sources now for custom roms to be made? That was a major reason to avoid mtk
This has been a known solution for a long time for those who have been following Linux scheduler development. Doing things fast means less power usage.
Oppo Find X is great but they don't sell worldwide. Thank you for sharing this neat information
It would be cool if the next Google Tensor takes this approach. I wonder if Intel is doing it with the the continous improvement of the e cores (I think AMD has done that though they did not go bigLITTLE)
I bought the iPhone 11 largely based on the SoC. And I’m still using the thing to this day.
shouldn't those efficieny cores be tested with like a lot of idle time?
Wait was that a 9900k losing by quite a bit to phone cpus?? Had no idea they were getting that good
Wow, that was a plot twist! I'm with you, 12T SRAM is where it's at.
Still hankering for a 5G OLED respin of 2010s HTC Desire with a nicer camera. If you can't make it work with 2 cores, is it even worth doing?
The S24 FE launched with the Exynos 2400e SoC (X4 core is lower clocked and it might be a worse binned version of the normal version) worldwide. Would it be possible to get your hands on it? 👀
Appreciate the energy efficiency comparisons, as only power efficiency is measured by everyone else!
The new snapdragon 8 elite has no little cores its Prime and Performance cores!
2 Class is really enough at this power envelope, I see necessity for smaller cores but mainly for wearables.
Can you list out the phones with the new chip?
With the D9300, there's the Find X7 and X100 Pro, and a couple others? For D9400, only the Find X8 has been announced so far.
@15:23 - why does the A715 @3GHz-512kb-L2 seems faster than the A720 @2GHz-256kb-L2 ?
3 GHz vs 2 GHz.
Awesome stuff, I wonder if HTC wouldve been on board....
Beyond the CPU improvements, it's really hard to care about anything else. It really doesn't do any good to have a great GPU that is phenomenal in synthetic benchmarks but then has subpar driver support that makes it perform as if it is several generations older than it is.
I would love to see some kind of 3d V-cache in the ARM CPU's of smartphone just like what AMD did with there Ryzen CPU's in the PC segment.
other tradeoff:bigger die area is more heat more heat is more cooling needed or depending on design more wear on battery
die area has no effect on power consumption. I can make a 600mm2 chip use 10 watts or a 20mm2 chip use 500 watts.
@@TechTechPotato if you constrain yoursef to power yes, but why do you use bigger die area? to use for more performance even if overal power usage is the same the shorter but higher burst of power generate more heat
Death by a thousand cuts in CPU design (Arrow Lake mentioned)
imagine a raspberry pi type device with something like this
Still sticking to that Qualcomm Snpapdragon drug. I need that fast modem that comes with Qualcomm SoCs.
Can your phone cook for you?
Can your phone do your dirty clothes?
Can your phone travel to the Sirius solar system (in and all by itself)?