ARM is introducing a series of new chip designs for the coming year that the company says will bring stronger CPU, graphics, and AI performance and reduced power consumption.
The new ARM Cortex-A78 CPU, Mali-G78 and Mali-G68 GPU, and Ethos-N78 neural processing unit are likely to show up in devices shipping next year, if not sooner.
ARM Cortex-A78
ARM says a 5nm ARM Cortex-A78 CPU offers 20-percent better sustained performance than a previous-gen Cortex-A77 core while consuming the same amount of power. In this case, “sustained” performance indicates that you’re less likely to see speeds drop at the processor gets hot.
The company says Cortex-A78 chips can also offer up to a 30-percent boost in peak single-core performance when used with ARM’s new Cortex-X1 custom program. (Anandtech takes a deeper dive into Cortex-X1, which is basically a high-performance implementation of Cortex-A78 technology that isn’t bound by the same power constraints).
And as usual, these high-performance CPU cores can be combined with lower-performance, more energy-efficient CPU cores in the same chip. In this case, ARM says you may see big.LITTLE chips that combine Cortex-A78 and Cortex-A55 cores.
ARM Mali-G78 and Mali-G68
ARM is promising up to a 25-percent performance boost in its new flagship-class Mali-G78 GPU, while using less power than its predecessors. The GPU also supports between 7 and 24 graphics cores.
The chip designer is also introducing a “sub-flagship” version called Mali-G68 which has many of the same features, but tops out at 6 graphics cores. ARM says we’ll see “sub-flagship” devices with Mali-G68 graphics in 2021.
Ethos-N78
According to ARM, compared with the previous-gen Ethos-N77 NPU, we can expect:
- 100-percent boost in peak performance
- 25-percent boost in performance efficiency
- 40-percent better bandwidth efficiency
I was expecting more. Usually ARM over-promises/under-delivers on these announcements. And I’ve noticed a tick-tock cadence in their architecture development. Where they’ll iterate initially on a new fabrication node… then follow-up later with an optimised architecture the next round, with notable boost to efficiency, performance, and area.
Example: A72 and A75 weren’t great.
Whereas: A73 and A76 were much more successful.
Here’s their current lineup:
$2, 32nm, Cortex A7, Mali-400 (Obsolete)
$3, 28nm, Cortex A35, Mali-G31 (tiny)
$5, 20nm, Cortex A53, Mali-G51
$10, 16nm, Cortex A55, Mali-G52
$20, 14nm, Cortex A72, Mali-G71
$25, 12nm, Cortex A73, Mali-G57
$35, 10nm, Cortex A75, Mali-G72
$40, 8nm, Cortex A76, Mali-G76
$60, 7nm, Cortex A77, Mali-G77
$60, 5nm, Cortex A78, Mali-G78 (upcoming)
ARM was really playing safe this time. I was hoping they would diverge a little so they can scale all the way from weak wearables to powerful desktop servers. So they should’ve introduced a new high-end platform.
Remember back in 2016 ARM had a clean three tier offering:
Cortex A35 – Ultra-low power, same perf as A7
Cortex A53 – Low power, same perf as A9
Cortex A73 – Medium power, beyond Apple A8 perf
I was hoping a similar thing to happen. Maybe it will come on a new architecture-branch, ergo ARMv9, maybe in 2022 on a 4nm node. For instance:
ARM v9 Cortex A41- Ultra-low power, perf. as A55
ARM v9 Cortex A61 – Low power, same perf. as A73
ARM v9 Cortex A81 – Medium power, beyond the perf. of Apple A14
Although, a big part will be optimisations and implementing a new (InfinityFabric, big.LITTLE, DynamIQ style) platform that scales well from wearables to desktop.
What they needs is to get these into TV boxes with good fast GPU. Not high end phones only.
“ARM Cortex-A78 CPU offers 20-percent better sustained performance than a previous-gen Cortex-A77 core while consuming the same amount of power. In other words, you’re 20-percent less likely to see speeds drop at the processor gets hot.”
Erm… I’m no math genius, but I’m pretty sure that’s not how either math or physics works…
Yeah, that was poorly worded, at best. I’ve updated the article. Thanks!
Hope I did not come across as hostile. I proof-read professionally 😐
So basically, what the original sentence says: when the chip has heated up and it thermal throttles, it runs 20% faster in that stable thermal throttled state. There is no information about how much more or less likely thermal throttling is now than it was before.