A couple of days in the past, ARM launched their new lineup for 2026 that begins out with the Lumex rename/model. SemiAccurate likes the brand new C1 CPUs, G1 GPUs, and different IP.
Lets begin out with the painful and get higher from there, rebranding. Certain ARM got here out with 5 manufacturers, 4 new, a few days earlier than Computex, however this time round each IP block has been renamed. Certain there are excuses as to why this can be a good concept however, nicely, ENOUGH ALREADY. Prior to now few years, ARM CPUs alone have gone from Cortex-Axx two digits to a few digits, then to Xy, and now to C1-suffix. GPUs have completed the identical 2 to three digit shuffle, then to Immortalis which simply died in an ironic twist, and now G1-suffix. And all of this ignores the identify modifications to the overarching platforms, interconnects, and different IP.
The issue we see just isn’t that the IP is unhealthy, it’s truly excellent, simply that there is no such thing as a continuity. What’s the trendy model of an Cortex-A78? Is an Immortalis-G720 akin to this technology’s Mali G1-Extremely or G1-Professional? We may go on however you most likely get the image, consistency isn’t simply missing, it appears to be deliberately disregarded. This time round ARM factors to the letter/quantity scheme and says they are going to be constant from right here on out. Whereas that is what they stated the final seemingly 17 instances, we hope they’re proper. Replace in a 12 months, however lets get again to the good things.
The brand new ARM Lumex Platform
Right here is the guts of the brand new stuff, the brand new Lumex platform which is actually not this. Lumex is the brand new CSS for Consumer which was the brand new TCS final 12 months if that parses proper. In any case that is the default excessive finish shopper platform ARM says is what the premium gadgets ought to use subsequent 12 months. On a technical entrance it’s onerous to argue a 2+6 platform with a excessive finish GPU, all of the goodies enabled, and AI is something however a premium system. Lets check out the excessive stage modifications, the deep tech will are available later articles.
The 4 new Cortex-C1 strains
The most important change is the CPUs with the addition of a fourth core sort. Huge/Little went to Huge/Medium/Little just a few generations in the past and now has change into Huge/LessThanThatLastOne/Medium/Little, when you don’t depend variants like a number of varieties of smaller cores optimized for efficiency or effectivity in the identical SoC. Qualcomm, we’re taking a look at you right here. In any case these are ARMv9.3-A cores with just a few v9.4-A options pulled in, this can be a good factor.
Higher but is the bit about all cores having the identical ISA, all will execute the identical code with out modifications, albeit at very completely different speeds. There is no such thing as a AVX-512 flag lottery right here, it’s all only one ISA, once more, this can be a excellent factor. The cores from excessive finish to low are known as Extremely, Premium, Professional, and Nano, abbreviated as C1-U, C1-P, C1-P…oh wait that doesn’t work. How about C1-U, C1-Pr, C1-Pr…rattling it! At the very least Professional is brief sufficient however this one will annoy for years.
C1 cluster comparability
The massive change right here is that the fourth CPU core class, Premium, is rather less than the Extremely which was the X# cores of the previous few years. If you’re questioning why anybody would hassle with yet one more core sort, aside from to prematurely age the wizards who write schedulers, there may be truly a very good purpose. The slide above factors out that going from 2U+6Pro to 2Pre+6Pro configs saves 35% die space for less than 25% much less efficiency. Certain the core clusters are a minority of the SoC die however this can be a substantial price financial savings for a claimed 25% efficiency loss. Do word that is for a 2MB Extremely vs 1MB Premium core.
ARM C1 cluster lineup
So how a lot financial savings are you able to get with the official ARM configurations? As you may see going from the Lumix Lumex 6+2 platform to a 2x Nano cluster will get you a 25x space financial savings. That isn’t %, it’s instances, and the ISA stays the identical. Efficiency goes up a ‘mere’ 17x so is it price it? If you’re making a ‘good’ TV, then the realm financial savings are seemingly nicely price it, however a excessive finish smartphone might even see issues in another way. In any case you get an concept of the vary of the 2026 CPU clusters scalability for space and efficiency, it’s fairly extensive.
The brand new Lumex Interconnects
On the system plumbing entrance we’ve got a significant change that doesn’t sound like a lot however can have huge effectivity advantages. The earlier CI-700 and NI-700 interconnects are changed by a single cloth known as SI L1 or System Interconnect (presumably) Lumex 1. The previous MMU is now changed by the brand new MMU L1 and all of it ties collectively extra tightly. For those who take a look at the price of shifting bits in a contemporary SoC vs the price of computing them, you’ll perceive why these modifications matter quite a bit.
Don’t overlook the brand new G1 GPU line
That brings us to the fanatic’s favourite subject, GPUs and the brand new Mali G1 household of gadgets. At a excessive stage there isn’t practically as a lot change right here as there was on the CPU facet, three configurations which can be extra unit counts than considerably completely different IP blocks. The G1-Extremely is a ten+ core beast with necessary raytracing, G1-Premium goes from 6-9 cores and doesn’t should have raytracing, and the G1-Professional is 1-5 cores. Inside the cores issues are very completely different, however once more that’s for an additional article.
In previous iterations of Mali you can configure the cores themselves to have completely different ratios of items in them, particularly on the low finish the place 3D was much less of a precedence than pixel pushing, assume TV/video controllers vs gaming gadgets. That function appears to be a factor of the previous however many of the low finish gadgets there’ll most likely stick to older GPU households for space and price financial savings, if the brand new options are usually not going for use, why pay for them?
The modifications to the Mali G1 ranges are usually not merely unit counts although, the plumbing is modified between ranges. Why? The much less items you’ve gotten the much less bandwidth and interconnects you could feed them, the much less I/O ports you want and so forth. Every stage of Mali matches the plumbing to the obtainable compute functionality for better effectivity and decrease price. This will appear apparent as a result of, nicely it’s. We are going to get into the modifications to the GPU cluster itself intimately in later articles.
If these new CPUs and GPUs sound iterative, they type of are at a excessive stage, aside from the brand new C1-Premium they’re all updates to present IP blocks. This isn’t to rattling them with faint reward, there may be plenty of good in them with many new options we are going to cowl later. The most important bang is the uncore, the SI L1 is actually new and completely different as are the modifications to the DSU itself. If the brand new naming scheme sticks round, I believe ARM is on to an excellent factor with this technology of IP.S|A
