Uploaded March 2026 | Updated September 2026, 1 week ago
Thanks to PCBWay.com/?from=atomic for the fantastic PCBs and SMT assembly.
I was trying to squeeze more performance out of an ESP32-S3 and ended up with a result I didn’t expect: optimizing for size was faster than optimizing for speed.
In this video, I benchmark different ESP-IDF settings and show what actually made the biggest difference.
The ESP32-S3 is a seriously capable little CPU, but getting the most out of it isn’t always obvious.
I started with a completely stock ESP-IDF configuration and used a simple but CPU-heavy benchmark: decoding an embedded animated GIF from flash. From there, I measured the impact of a number of common performance tweaks:
• CPU clock speed (160 MHz vs 240 MHz)
• Compiler optimization level (-Os vs -O2)
• Flash SPI mode (DIO vs QIO)
• Instruction and data cache sizes
What surprised me most was that “optimize for size” consistently beat “optimize for performance” in this workload - even before touching the cache settings. Once the caches were increased, the results got even more interesting.
If you’re working on performance-critical ESP32 projects (graphics, codecs, emulation, signal processing), hopefully this gives you a few useful ideas - and maybe challenges a couple of assumptions.
Let me know your findings in the comments!
Thanks for watching!
00:00 The ESP32-S3 and why performance matters
00:36 The benchmark: decoding an embedded GIF
01:03 Baseline results (stock ESP-IDF)
01:43 CPU frequency: 160 MHz vs 240 MHz
02:14 Compiler optimizations: -Os vs -O2
03:32 Flash SPI mode: DIO vs QIO
04:10 Cache configuration (instruction & data cache)
05:08 Turning everything up to 11
05:55 Revisiting -O2 with larger caches
06:14 The winning configuration
06:22 Final thoughts
---
If you'd like to help support the channel and help me buy more nonsense - I have a Patreon page: patreon.com/atomic14
We've also got a shop now! shop.atomic14.com
And if you're shopping on AliExpress or Amazon then if you click on these links I will get a little bit of affiliate money:
Aliexpress - https://s.click.aliexpress.com/e/_Ac05mF
Amazon - amazon.com/?tag=atomic14-20
You can also join us on Discord here: discord.gg/h2DDnDdjD2
There's also the Atomic14 newsletter here: atomic14.substack.com
And don't forget about the MakerNews newsletter - makernews.substack.com
Thanks to PCBWay.com/?from=atomic for the fantastic PCBs and SMT assembly.
I was trying to squeeze more performance out of an ESP32-S3 and ended up with a result I didn’t expect: optimizing for size was faster than optimizing for speed.
In this video, I benchmark different ESP-IDF settings and show what actually made the biggest difference.
The ESP32-S3 is a seriously capable little CPU, but getting the most out of it isn’t always obvious.
I started with a completely stock ESP-IDF configuration and used a simple but CPU-heavy benchmark: decoding an embedded animated GIF from flash. From there, I measured the impact of a number of common performance tweaks:
• CPU clock speed (160 MHz vs 240 MHz)
• Compiler optimization level (-Os vs -O2)
• Flash SPI mode (DIO vs QIO)
• Instruction and data cache sizes
What surprised me most was that “optimize for size” consistently beat “optimize for performance” in this workload - even before touching the cache settings. Once the caches were increased, the results got even more interesting.
If you’re working on performance-critical ESP32 projects (graphics, codecs, emulation, signal processing), hopefully this gives you a few useful ideas - and maybe challenges a couple of assumptions.
Let me know your findings in the comments!
Thanks for watching!
00:00 The ESP32-S3 and why performance matters
00:36 The benchmark: decoding an embedded GIF
01:03 Baseline results (stock ESP-IDF)
01:43 CPU frequency: 160 MHz vs 240 MHz
02:14 Compiler optimizations: -Os vs -O2
03:32 Flash SPI mode: DIO vs QIO
04:10 Cache configuration (instruction & data cache)
05:08 Turning everything up to 11
05:55 Revisiting -O2 with larger caches
06:14 The winning configuration
06:22 Final thoughts
---
If you'd like to help support the channel and help me buy more nonsense - I have a Patreon page: patreon.com/atomic14
We've also got a shop now! shop.atomic14.com
And if you're shopping on AliExpress or Amazon then if you click on these links I will get a little bit of affiliate money:
Aliexpress - https://s.click.aliexpress.com/e/_Ac05mF
Amazon - amazon.com/?tag=atomic14-20
You can also join us on Discord here: discord.gg/h2DDnDdjD2
There's also the Atomic14 newsletter here: atomic14.substack.com
And don't forget about the MakerNews newsletter - makernews.substack.com
