Different performance of openssl speed on the same hardware with AES 256 (EVP and non EVP API)

In OpenSSL source code, the speed aes-256-cbc function calls AES_cbc_encrypt() which itself uses AES_encrypt(), a function from crypto/aes/aes_x86core.c. It is an obvious "classical" implementation with tables.

On the other hand, with EVP, you end up in the code in crypto/evp/e_aes.c which dynamically detects whether the current CPU supports the AES-NI instructions, a feature of recent x86 processors, which allow for vastly improved performance. In OpenSSL code, the AESNI_CAPABLE macro does the job (feeding on some flags which are set when the library is initialized, using CPUID).

Bottom-line: with EVP, you benefit from the automatic selection of the improved implementation, based on the current CPU model, whereas the non-EVP code directly uses the generic software implementation, which works everywhere, but is slower.

One more thing to notice:

$ ./openssl speed aes-256-cbc (i.e without EVP API)
Doing aes-256 cbc for 3s on 16 size blocks: 14388425 aes-256 cbc's in 3.00s
$ ./openssl speed -evp AES256
Doing aes-256-cbc for 3s on 16 size blocks: 71299827 aes-256-cbc's in 3.00s

Without EVP API, in 3.00s, processed 14,388,425 (~14M)
With EVP API, in 3.00s, processed 71,299,827 (~71M)

It's obviously processed faster in EVP mode.