Qualcomm Announces Snapdragon 768G: Higher-bin 765 up to 2.8GHz
by Andrei Frumusanu on May 11, 2020 3:00 AM EST- Posted in
- Mobile
- Qualcomm
- SoCs
- 915
- Snapdragon 768G
Today alongside with the launch of the Xiaomi Redmi K30 5G Racing Edition, Qualcomm is announcing the new Snapdragon 768G SoC which powers the device. The new SoC is a direct follow-up to the Snapdragon 765G announced last December, and the two chips are very likely the same silicon design, with the new variant increasing the clock frequencies.
Qualcomm Snapdragon Premium SoCs 2019-2020 | ||||
SoC | Snapdragon 768G | Snapdragon 765 Snapdragon 765G |
Snapdragon 730 | |
CPU | 1x Cortex A76 @ 2.8GHz 1x Cortex-A76 @ 2.4GHz 6x Cortex-A55 @ 1.8GHz |
1x Cortex A76 @ 2.3GHz (non-G) @ 2.4GHz (765G) 1x Cortex-A76 @ 2.2GHz 6x Cortex-A55 @ 1.8GHz |
2x Cortex-A76 @ 2.2GHz 6x Cortex-A55 @ 1.8GHz |
|
GPU | Adreno 620 +15% perf over 765G |
Adreno 620 +20% perf (non-G) +38% perf (765G) |
Adreno 618 | |
DSP / NPU | Hexagon 696 HVX + Tensor 5.4TOPS AI (Total CPU+GPU+HVX+Tensor) |
Hexagon 688 HVX + Tensor |
||
Memory Controller |
2x 16-bit CH @ 2133MHz LPDDR4X / 17.0GB/s |
2x 16-bit CH @ 1866MHz LPDDR4X 14.9GB/s |
||
ISP/Camera | Dual 14-bit Spectra 355 ISP 1x 192MP or 36MP with ZSL or 2x 22MP with ZSL |
Dual Spectra 350 ISP 1x 36MP with ZSL or 2x 22MP with ZSL |
||
Encode/ Decode |
2160p30, 1080p120 H.264 & H.265 10-bit HDR pipelines |
|||
Integrated Modem | Snapdragon X52 Integrated (LTE Category 24/22) DL = 1200 Mbps 4x20MHz CA, 256-QAM UL = 210 Mbps 2x20MHz CA, 256-QAM (5G NR Sub-6 4x4 100MHz + mmWave 2x2 400MHz) DL = 3700 Mbps UL = 1600 Mbps |
Snapdragon X15 LTE (Category 15/13) DL = 800Mbps 3x20MHz CA, 256-QAM UL = 150Mbps 2x20MHz CA, 64-QAM |
||
Mfc. Process | Samsung 7nm EUV (7LPP) |
Samsung 8nm (8LPP) |
The new chip features the same Cortex-A76 cores in a 1+1 configuration (one Prime high-clocked core, and one medium clocked core), alongside 6 Cortex-A55 cores. The difference in CPU performance lies in the frequencies of the big cores which are now at up to 2.8GHz and 2.4GHz for the Performance and Middle core – a more notable uplift from the 2.4 and 2.2GHz clocks of the Snapdragon 765G.
GPU clock frequencies have also been increased, resulting in at 15% performance boost over the Snapdragon 765.
The rest of the chip is seemingly identical to the Snapdragon 765 series. What’s interesting here is that Qualcomm does name it quite differently in its SKU line-up. While it very much shares the design of the Snapdragon 765, it’s also a possibility that it’s a silicon respin of the chip, the timelines certainly would make sense and it’s also not the first time that Qualcomm would have done this (Snapdragon 821 is an example of this). If the increased clocks come at a cost of higher power draw or loss of efficiency is anybody’s guess right now – there’s also the possibility that yields on Samsung’s 7LPP node has improved and thus enabled the higher frequencies.
33 Comments
View All Comments
Wilco1 - Monday, May 11, 2020 - link
A shorter pipeline is better for IPC. Most Arm designs have 64KB L1 caches and L2 at 512KB or 1MB, so 2-4 times as large as Skylake.Also Skylake can execute up to 4 instructions per cycle (plus 1 with branch fusion) and has 8 execution pipelines. Cortex-A76 does also 4 instructions per cycle and has 8 execution pipelines.
jospoortvliet - Monday, May 11, 2020 - link
Note that a shorter pipeline means better performance (but lower clocks), all other things being equal (which they are not, of course!).But you are right that it is near impossible to imagine those 22 execution units not kicking the a** of the A76. But keep in mind that Skylake can only shoot 8 ops into these units per clock, that is beaten by Apple's a12 at 9, and yes, equaled by the A76 with 8. Im no cpu designer but it seems to ignorant me that if the execution units are fully pipelined it doesn't matter if you say there are 20, but you can feed only 8 per clock. Given all the other differences they are probably hard to compare but given they can pump a similar nr of instructions in the execution units i would guess they aren't THAT far off in IPC... not 22 vs 8, at least.
We do know Apple with its 9 is far (30-40%?) ahead of the A76 so the 8 vs 9 doesn't say everything - it isn't 40% more...
But I agree Intel is probably still ahead of the A76 in IPC, apple might have surpassed them, maybe the A76 is close but it sure isn't 40% ahead of intel which it would need to compensate for lower frequency..
Wilco1 - Monday, May 11, 2020 - link
You're quite right it's all about feeding execution units. But let's compare using actual measured SPEC results rather than meaningless execution unit numbers.According to AnandTech's SPECINT results, Cortex-A76 has ~12% higher IPC than the 9900K - see https://images.anandtech.com/doci/15776/SPEC-D1000...
Similarly Neoverse N1 (which uses the same microarchitecture as Cortex-A76) has ~16% higher IPC than Cascade Lake, see https://images.anandtech.com/graphs/graph15578/111... from the Graviton 2 review: "It’s a great showcase of the Neoverse N1’s IPC capabilities, as the cores are only running at 2.5GHz compared to ~2.9GHz for the AMD system and ~3.2GHz for the Intel system."
The latter result is more accurate since memory systems, cache sizes and frequencies are very similar for these servers, unlike any mobile vs desktop comparison.
The A13 has not just surpassed Intel and AMD in IPC, it has 1.8 times the IPC! The reason it can execute twice as many instructions per cycle as A76 or Skylake.
So I repeat, aiming for 5+GHz is as stupid as Pentium 4 or Bulldozer was.
KarlKastor - Tuesday, May 12, 2020 - link
Where do you see there a higher IPC for the A76?Wilco1 - Tuesday, May 12, 2020 - link
It's basic math:Neoverse N1 scores 32.34 / 2.5GHz = 12.9 SPECINT/GHz
Platinum 8529 scores 35.74 / 3.2GHz = 11.2 SPECINT/GHz
philehidiot - Wednesday, May 13, 2020 - link
Excuse me, the P4 was NOT stupid. You talk about parallel processing, blah blah blah but what other CPU would allow you to fry an egg whilst SIMULTANEOUSLY weeping at the shoddy performance? We called it Simultaneous Multi Egging. It was ahead of it's time and AMD have clearly just ripped off the nomenclature.Okay, fine. I'll be quiet and go away.
SaolDan - Monday, May 11, 2020 - link
Amenliquid_c - Monday, May 11, 2020 - link
I think i found the AMD stockholder. Was wondering when you’d pop up. How the f*ck can you be so retarded as to compare an ARM cpu against a fully fledged desktop cpu and even go as far as to claim that it “beats it”? Can the A13 execute x86 instruction sets? Do you think that, if it were that good, Apple wouldn’t use them in their line of Macs and iPads and instead, choose to pay Intel for their CPUs? God damn it, the internet’s wikipedia has turned all you dropouts into PhDs.Wilco1 - Monday, May 11, 2020 - link
LOL - A13 actually beats the 4.7GHz Ryzen 3950X on SPECINT: https://images.anandtech.com/doci/15776/SPEC-D1000...You'd have to be retarded to believe that the next generation will not beat the fastest Intel and AMD desktop CPUs. And: https://www.bloomberg.com/news/articles/2020-04-23...
MarcGP - Monday, May 11, 2020 - link
Hold your horses, no need for name calling. He was talking about the Graviton 2 and the Ampere ARM server processors, not the Apple A13. They were reviewed and benchmarked by Anandtech and they did a pretty good job keeping up and even besting the top x86 server processors.https://www.anandtech.com/show/15578/cloud-clash-a...
By the way, everybody seems to agree that this time is going to happen, that the next generation of Macbooks will use their own Apple ARM SOCs.
https://www.bloomberg.com/news/articles/2020-04-23...