Snapdragon 8G3 is the first mobile SoC based off ARMv9.2 (64-bit-only). It still uses TSMC’s N4P node like the 8+G1 and 8G2. Besides using ARM’s brand new core architecture (X4, A720, A520), it also now uses a 1+5+2 core configuration instead of the 8G2’s 1+4+3 or the more traditional 1+3+4 in the 8+G1.
ARM’s new cortex-X4 big core boasts some very decent improvements in ST performance (~20%) over the previous X3 in the likes of the Snapdragon 8G2. However, power consumption has also increased accordingly, so ST efficiency isn’t noticeably improved. This is mostly fine considering the big core is generally leveraged to handle short and bursty workloads in order to make the phone feel snappy and fast. Still noticeably behind Apple’s A16 /A17 Pro in terms of ST performance/efficiency though.
The A720 medium core and A520 small core are mostly incremental upgrades both in terms of performance and efficiency over their predecessors. A720 does get a pretty sweet efficiency improvement in floating point compute, but nothing crazy.
The swapping of yet another small core for a medium core results in some very nice MT performance improvements, now putting the Snapdragon 8G3 pretty much ahead of the A17 Pro’s perf/W in GB 5 and matching it in GB 6. Qualcomm has officially catched up to Apple in MT performance and efficiency. It does remain slightly less efficient at lower power levels though.
Apparently, this chipset is capable of partially turning of its L3 cache and SLC to save power, but we still don’t know much about how that would work and whether it requires software-based scheduling or not.
The GPU is yet again a noticeable upgrade over the SD 8G2, and seems to mostly come from scaling up the compute like 8G2 did over 8+G1. 8G2 was already on top of the competition when it came to GPU performance, so this upgrade pretty much puts 8G3 in a class of its own when it comes to both to performance and efficiency on the GPU side. It completely overshadows every other smartphone chipset and even seems to match the performance a Radeon 780M/GTX 1050Ti with an overclock.
Overall, 8G3 is a very solid upgrade over the 8G2 which was already a beast of an SoC. None of the performance or efficiency improvements are not groundbreaking by any means, but they are definitely there. 8G3 vs. 8G2 is a jump pretty much equivalent to 8G2 vs. 8+G1 both in terms of CPU and GPU performance/efficiency, which is great. This really puts to shame Google’s Tensor G3, which is barely a performance/efficiency improvement over the notoriously bad Tensor G2 (Samsung Foundries at it again). The amount of competition Qualcomm is putting up with Apple is really exciting to see.
ARM’s new cortex-X4 big core boasts some very decent improvements in ST performance (~20%) over the previous X3 in the likes of the Snapdragon 8G2. However, power consumption has also increased accordingly, so ST efficiency isn’t noticeably improved.
It’s actually considerably worse efficiency.
+17.5% ST perf for +40% power in SPEC2017Int
About as expected going from 6-wide to 10-wide decode and increasing clock speed…
Haven’t seen it all yet, but since the G2 GPU was already going blow for blow with the A17, I assume this one is comfortably ahead. Apple’s MT and GPU leads are gone, all they retain is single core and the margin is narrowing substantially.
I wonder if they can ever separate themselves from the pack again, or this is like the Zen moment of CPU convergence where AMD and Intel have just stuck close since.
TLDR:
Snapdragon 8G3 is the first mobile SoC based off ARMv9.2 (64-bit-only). It still uses TSMC’s N4P node like the 8+G1 and 8G2. Besides using ARM’s brand new core architecture (X4, A720, A520), it also now uses a 1+5+2 core configuration instead of the 8G2’s 1+4+3 or the more traditional 1+3+4 in the 8+G1.
ARM’s new cortex-X4 big core boasts some very decent improvements in ST performance (~20%) over the previous X3 in the likes of the Snapdragon 8G2. However, power consumption has also increased accordingly, so ST efficiency isn’t noticeably improved. This is mostly fine considering the big core is generally leveraged to handle short and bursty workloads in order to make the phone feel snappy and fast. Still noticeably behind Apple’s A16 /A17 Pro in terms of ST performance/efficiency though.
The A720 medium core and A520 small core are mostly incremental upgrades both in terms of performance and efficiency over their predecessors. A720 does get a pretty sweet efficiency improvement in floating point compute, but nothing crazy.
The swapping of yet another small core for a medium core results in some very nice MT performance improvements, now putting the Snapdragon 8G3 pretty much ahead of the A17 Pro’s perf/W in GB 5 and matching it in GB 6. Qualcomm has officially catched up to Apple in MT performance and efficiency. It does remain slightly less efficient at lower power levels though.
Apparently, this chipset is capable of partially turning of its L3 cache and SLC to save power, but we still don’t know much about how that would work and whether it requires software-based scheduling or not.
The GPU is yet again a noticeable upgrade over the SD 8G2, and seems to mostly come from scaling up the compute like 8G2 did over 8+G1. 8G2 was already on top of the competition when it came to GPU performance, so this upgrade pretty much puts 8G3 in a class of its own when it comes to both to performance and efficiency on the GPU side. It completely overshadows every other smartphone chipset and even seems to match the performance a Radeon 780M/GTX 1050Ti with an overclock.
Overall, 8G3 is a very solid upgrade over the 8G2 which was already a beast of an SoC. None of the performance or efficiency improvements are not groundbreaking by any means, but they are definitely there. 8G3 vs. 8G2 is a jump pretty much equivalent to 8G2 vs. 8+G1 both in terms of CPU and GPU performance/efficiency, which is great. This really puts to shame Google’s Tensor G3, which is barely a performance/efficiency improvement over the notoriously bad Tensor G2 (Samsung Foundries at it again). The amount of competition Qualcomm is putting up with Apple is really exciting to see.
It’s actually considerably worse efficiency.
+17.5% ST perf for +40% power in SPEC2017Int
About as expected going from 6-wide to 10-wide decode and increasing clock speed…
Haven’t seen it all yet, but since the G2 GPU was already going blow for blow with the A17, I assume this one is comfortably ahead. Apple’s MT and GPU leads are gone, all they retain is single core and the margin is narrowing substantially.
I wonder if they can ever separate themselves from the pack again, or this is like the Zen moment of CPU convergence where AMD and Intel have just stuck close since.