1. FPGA Implementation of Multicore AES 128/192/256
By: William Whitehouse
Hi my name is William Whitehouse. I am presenting my project an FPGA implementation of Multi-Core AES 128/192/256.

2. Objectives
Design an FPGA that is capable of AES encryption and decryption with 128, 192, and 256 bit keys.
Optimize encryption/decryption throughput by implementing multiple cores
Synthesize multi-core design and generate a simulated throughput
The purpose of this project was to implement a field programmable gate array (FPGA) prototype that implements the Advance Encryption Standard (AES) with key sizes of 128-bits, 192-bits, and 256-bits. As a proof of concept, multiple AES encryption and decryption cores were used in the design to improve the data throughput. Finally, the key size selectable multi-core AES design was synthesized to generate a simulated throughput for comparison with a single core design.

3. AES Core Modules Used AES Core Modules developed by Jerzy Gbur
Contains:
key_expansion.vhd – 128, 192, or 256 bit key expansion module
aes_enc.vhd – AES encryption module
aes_dec.vhd – AES decryption module
I utilized the AES Core Modules developed by Jerzy Gbur that are available on opencores.org. The original modules had a generic parameter that determines the key to be used (0 – 128bit, 1 – 192bit, and 2 – 256bit) statically , so I modified the cores to make the key size a variable input so it could be changed when a new key was to be created. These cores were chosen because of the ability to use all three key sizes and to instantiate multiple cores easily.

4. Single AES Core Timing Key Size 128-bit 192-bit 256-bit
Key Valid (cycles) 16 24 32
Key Expansion (cycles)
133
123
157
Enc/Dec Data Valid (cycles)
Enc/Dec (cycles) 99
119
139
The AES core uses the 8-bit data bus and valid signals to capture the key data. The number of key valid clocks depends on the size of the key. After the entire key is received key expansion is performed at a rate dependent on the key size. Once the key expansion is complete the encryption and decryption cores are ready. The encryption/decryption cores accept one byte of data per cycle. Once all 128 bits are received the encryption/decryption begins. The time to encrypt/decrypt is also dependent on the key size. The cores output the encrypted or decrypted data one byte at a time.

5. Single Core Wave
This waveform illustrates the key expansion, encryption, and decryption timing for each of the key sizes for a single core AES design. Notice the enc_ready signal is de-asserted as soon as the first encryption data input is received, the next 128 bit data encryption can not start until this single core has completed the first encryption.
The encryption output is looped back into the decryption input and the decrypted output is checked to verify it matches the encryption input.
The 192-bit and 256-bit key expansion, encryption, and decryption are also verified with the loopback.

6. Top Level Diagram
This is a top level diagram of the design. The aes_top.vhd uses constants to specify how many AES Encryption and AES decryption cores are implemented. The encryption and decryption cores use the same key but are separated so encryption and decryption can be done simultaneously. Combinational and sequential logic in the aes_top module directs the incoming encryption and decryption data to cores that are not busy. Once a core drives the output the top level module sends that data on the output bus along with the output valid signal. After the last output data is driven the core becomes “not busy” and can then receive input data to encrypt/decrypt.

7. Multi-Core Timing Key Size 128-bit 192-bit 256-bit
Initial Enc/Dec (cycles) 99
119
139
Enc/Dec (cycles) 17
Max Cores Utilized 8 9 11
The multi-core design requires at least one idle clock between the 16-bytes of input data for encryption or decryption. The time to encrypt/decrypt the first 16 bytes is still the same as the single core design. After the initial encrypt/decrypt data will be output every 17 clock cycles (16 valid data and 1 idle).
Each 128 bits of encrypted or decrypted data output matches the order of the data input.
The maximum number of cores utilized depends on the current key size. Since 11 cores are needed to reach the maximum encryption/decryption rate for 256 bit key size, that is the number used to synthesize and generate simulation throughput.

8. Multi-core Enc/Dec Wave
This waveform illustrates the benefit of using multiple AES cores. The key expansion phase for any number of cores is constant. When using a single core the throughput is limited by the amount of time it takes for encryption/decryption. However, with 11 cores the throughput (after the initial encrypt/decrypt) is limited by the rate data is input (17 cycles if data is input continuously). Notice how the enc_ready and dec_ready signals stay asserted after the key expansion is completed for 128-bit and 192-bit key sizes. For the 256-bit key the ready signals are de-asserted for 1 clock cycle, but due to the idle clock necessary for data input, this does not slow down the encrypt/decrypt rate.

9. AES Design Synthesis Target FPGA: Xilinx Virtex 5 XC5VLX110
Single Core
11 Core
Slice Registers
693
7282
Slice LUTs
2416
26121
Block RAMs 4 44
Max Freq. (MHz)
297.95
297.42
Max Throughput (MBps)
39 (128-bit)
33 (192-bit)
29 (256-bit)
266
The target FPGA for design synthesis was the Xilinx Virtex-5 XC5VLX110 FPGA because it is optimized for high-performance logic and has enough flip-flops to only be 40% utilized by a 11 enc/dec core AES design. This allowed the max frequency (297MHz) for both the single and the multi-core design to be limited by the logic and not by size constraints of the target FPGA.
The throughput for the single core design is limited by the encryption/decryption timing of a single core. Because the key size determines the timing of a 128-bit data encryption or decryption the max throughput is also determined by the key size. The 11-core design has about a 7x improvement over the single core throughput.
Each encryption or decryption core adds about 330 slice registers, 1200 slice LUTs, and 2 block RAMs. So it is possible to sacrifice some of the throughput (by using fewer cores) to meet the resource constraints of a smaller FPGA. It is also possible to limit the design to encryption or decryption only and cut the full 11-core design size in half.

10. Growth Opportunities Add more Enc/Dec inputs and 11-core modules
Optimize the design for size
Allow the key to be different for each core
Use different AES Core
Some areas for improvement on this multi-core AES design are:
Increase the bandwidth by having multiple encryption/decryption IO and to support multiple 11-core modules (effectively multiplying the throughput, but also the size).
Each encryption and decryption core has its own key expansion module. That could be extracted and put in the top level to significantly decrease the resource utilization. However, this would require more logic in the top level so that each core could simultaneously access different addresses of the key expansion block RAM.
Each encrypt and decrypt core shares the same key and key size. The design could add inputs to select the key size and key data for each of the cores individually. This may not be a necessary use case, but the ability could be added to the current design.
This multi-core AES design could be improved by utilizing an AES core that uses fewer clock cycles for a single encryption/decryption, however, that may decrease the maximum frequency.

11. Questions Design is located at:
If you have any questions please me at:
If you have any questions feel free to me. I used Google code to maintain version control during the development of this project so all the VHDL design files are available online under an open source license.