AWS T2 vs T3 vs T3a Performance
Amazon just came out with the new 3rd generation of burst-able EC2 servers and to keep things short, you should shut down your t2 servers and restart them as t3.
What is the difference between AWS t2 and t3 and t3a series?
1 Core from an Intel Xeon E5-2676 v3 @ 2.40Ghz (released 2013)
3227 Geekbench 4 Single-Core
3025 Geekbench 4 Multi-Core
15.5 GB/sec Memory Bandwidth
2 Core from an Intel Xeon Undisclosed @ 2.50Ghz
3437 Geekbench 4 Single-Core
4571 Geekbench 4 Multi-Core
9.31 GB/sec Memory Bandwidth
From its looks, t3 clock speed is only marginally higher, but that isn’t the impressive part. On the t2.micro, it was not unexpected to see a single core machine running a multi-threaded test to have a score lower than its single-core. But now that t3.micro has 2 core, it’s able to get a huge 1,500 point boost to its score, which helps multi-threaded processes like web servers.
(burstable generation3 AMD processor micro instance)
1 Core from an AMD EPYC 7571 @ 2.20Ghz (released 2017)
3239 Geekbench 4 Single-Core
4044 Geekbench 4 Multi-Core
24.8 GB/sec Memory Bandwidth
EBS write speed 18.1 MB/sec
speedtest-cli 382mbps download | 234 mbps upload
At first glance, the lower clock speed and single-core were predicting to be similar to a t2 but its newer chip generation made it jump ahead. The memory bandwidth is much higher than both the t2 and t3 intel variants.
Geekbench 5 Chart
The test were done with Geekbench 5.4.5 on August 2022. There is a bit of a lottery of what instance CPU you will get, so don’t read too much into a 890 score vs a 907 score.
WIP – work in progress aug 2022
??? = t2.micro (1 cores, 1gb ram)
727 = t2.small (1 cores on Intel Xeon E5-2676 v3, 2gb ram)
??? = t2.medium ()
??? = t2.large ()
??? = t2.xlarge ()
??? = t3.micro (1 core, 2 vCPU, 1gb ram)
907 = t3.small (1 core, 2 vCPU on Intel Xeon Platinum 8180M, 2gb ram)
956 = t3.medium (1 core, 2vCPU on Intel Xeon Platinum 8259CL, 4gb ram)
896 = t3.large (1 core, 2vCPU on Intel Xeon Platinum 8259CL, 8gb ram)
1952 = t3.xlarge (2 cores, 4 vCPU, 16gb ram)
??? = t3.2xlarge (4 cores, 8 vCPU, 32gb ram)
An interesting footnote, the original benchmarks are in Geekbench 4 because when I try to run Geekbench 5 it crashes after single-core
Running Gaussian Blur test without any message. Poking around, I got Geekbench to show these warnings which might expect some graphic libraries loaded that a stock Amazon Ubuntu doesn’t have. It looks like newer versions of GB5 have fixed this.
[1201/050810:WARNING:src/halogen/cuda/cuda_library.cpp(1465)] Cannot find or load CUDA library. [1201/050810:WARNING:src/halogen/cl/opencl_library.cpp(691)] Cannot find or load OpenCL library. [1201/050810:WARNING:src/halogen/vulkan/vulkan_library.cpp(1342)] Failed to load vkGetInstanceProcAddr:
Steve Zazeski ( )
For EBS write test I used DD with a 2GB file, this test is definitely synthetic and bias towards doing well, so it’s surprising to see such low numbers.
dd if=/dev/zero of=./test bs=512k count=4048 oflag=direct
Steve Zazeski ( )
m5.large with a 20GB EBS volume got 143 MB/s