SHA-3 Hardware Implementations

1 Call for Contributions

Implementers (both submitters and non-submitters): You have results that complement this site? Let us know at sha3zoo-hardware@iaik.tugraz.at If you are making your HDL code available, please also provide us with according information.

2 Important Information

This page summarizes key properties of reported hardware implementations of those SHA-3 candidates, which are currently under consideration by NIST. This is work in progress. If you know of any implementations which should be mentioned on this page, refer to our call for contributions.

A list of hardware implementations of the round 1 candidates can be found here. Please note that the page for round 1 candidates is provided for reference and will not be updated.

The implementations are categorized into FPGA and standard-cell ASIC implementations. Note that the diversity of implementation scope, target technologies, and synthesis tools makes direct comparisions between different hardware implementation difficult. The more of these parameters agree, the more reasonable the comparison becomes.

The target technology should be as similar as possible. For FPGA implementation, it is desirable to compare implementations on the same target device (or at least on devices of the same FPGA family). For standard-cell ASIC implementation, at least the minimal gate length of the process (e.g., 0.13 µm) should agree. More ideally, the implementations use the same standard-cell library (which implies the use of the same process technology).

In order to facilitate the comparision of hardware modules with different implementation scopes, we classify them into three categories:

• Fully autonomous
• Using external memory
• Core functionality

For suggestions regarding the structure of this site, let us know at sha3zoo-hardware@iaik.tugraz.at

2.1 Fully Autonomous Implementation

Such hardware implementations include the complete functionality of a SHA-3 candidate (or a specific version thereof). That means the input message can be loaded piecewise into the hardware module and it delivers the message digest as output. All hash calculations happen exclusively within the hardware module. If integrated in a system, the achievable throughput of a fully autonomous implementation depends on the speed of the hardware module itself and the speed of the (system dependent) data interface delivering the input message.

2.2 Implementation with External Memory

These implementations use external memory to hold intermediate values during the hashing of a message. The implemented hardware itself normally consists of the core logic functionality of the hash function, some registers for short-lived temporary values, and possible a memory controller for access to the external memory. Such implementations can load the input message either over a dedicated interface (similar to a fully autonomous implementation) or from the external memory. In order to reach the maximal throughput of the hardware module, the external memory must be sufficiently fast.

2.3 Implementation of Core Functionality

Such implementations comprise only important parts of the hash function (e.g., the compression function), which normally allows to get a first-order estimate of the performance figures of full implementations.

3 Summary of All Results

This section includes four categories of implementations (high-speed, low-area, both for FPGA and ASIC) which include known published results. If the HDL sourcecode is available, a link is provided as well.

3.1 High-Speed Implementations (FPGA)

Important note: The size and functionality of slices varies between FPGA families. A direct comparision of the slice count of implementations on different FPGA families is therefore problematic.

Hash Function Name	Reference / HDL	Impl. Scope	Impl. Details	Technology	Size	Throughput	Clock Frequency
BLAKE-32	Submission doc. [1] / Submission webpage	Core functionality	Compression function with 8 G function units	Xilinx Virtex-II Pro	3091 slices	1724 Mbit/s	37.0 MHz
BLAKE-32	Submission doc. [1] / Submission webpage	Core functionality	Compression function with 8 G function units	Xilinx Virtex 4	3087 slices	2235 Mbit/s	48.0 MHz
BLAKE-32	Submission doc. [1] / Submission webpage	Core functionality	Compression function with 8 G function units	Xilinx Virtex 5	1694 slices	3103 Mbit/s	67.0 MHz
BLAKE-32	Namin and Hasan [2] / N/A	Core functionality	Compression function with 8 G function units and I/O registers	Altera Stratix III	5435 ALUTs	2186.2 Mbit/s	46.97 MHz
BLAKE-32	Kobayashi et al. [3] / RCIS webpage	Fully autonomous		Xilinx Virtex 5	1660 slices	2676 Mbit/s	115 MHz
BLAKE-64	Submission doc. [1] / Submission webpage	Core functionality	Compression function with 8 G function units	Xilinx Virtex-II Pro	11122 slices	1177 Mbit/s	17.0 MHz
BLAKE-64	Submission doc. [1] / Submission webpage	Core functionality	Compression function with 8 G function units	Xilinx Virtex 4	11483 slices	1707 Mbit/s	25.0 MHz
BLAKE-64	Submission doc. [1] / Submission webpage	Core functionality	Compression function with 8 G function units	Xilinx Virtex 5	4329 slices	2389 Mbit/s	35.0 MHz
Blue Midnight Wish-256	Namin and Hasan [2] / N/A	Core functionality	Compression function with f0, f1, and f2 unrolled in sequence and I/O registers	Altera Stratix III	12917 ALUTs	4889.6 Mbit/s	9.55 MHz
CubeHash8/1-256(***)	Baldwin et al. [4] / N/A	Core functionality	2 compression functions unrolled	Xilinx Spartan 3	3268 slices	70 Mbit/s	37.9 MHz
CubeHash8/1-256(***)	Baldwin et al. [4] / N/A	Core functionality	1 iterated compression function	Xilinx Virtex 5	1178 slices	160 Mbit/s	166.8 MHz
CubeHash16/32-256	Kobayashi et al. [3] / RCIS webpage	Fully autonomous		Xilinx Virtex 5	590 slices	2960 Mbit/s	185 MHz
ECHO-224/256	Lu et al. [5] / N/A	Fully autonomous		Xilinx Virtex 5	9333 slices	14860 Mbit/s	87.1 MHz
ECHO-224/256	Kinsy and Uhler [21] / N/A	Fully autonomous	273 cycles per block	Altera Cyclone II	39091 LEs	397 Mbit/s(*)	70.6 MHz
ECHO-256	Ramakers and Narinx [25] / Hosted by SHA-3 zoo	Core functionality	4 x 2 AES round instances with pipeline register in BigSubWords	Xilinx Virtex 5	15006 slices	23860 Mbit/s	139 MHz
ECHO-256	Kobayashi et al. [3] / RCIS webpage	Fully autonomous		Xilinx Virtex 5	3556 slices	1614 Mbit/s	104 MHz
ECHO-256	Mabrouk and Benadjila [28] / Implementer's webpage	Fully autonomous	Fully parallel iterations of Compress512	Xilinx Virtex 5	10407 slices	26390 Mbit/s	154.6 MHz
ECHO-256	Mabrouk and Benadjila [28] / Implementer's webpage	Fully autonomous	Fully parallel iterations of Compress512	Xilinx Virtex 6	8071 slices	29457 Mbit/s	172.6 MHz
ECHO-384/512	Lu et al. [5] / N/A	Fully autonomous		Xilinx Virtex 5	9097 slices	7810 Mbit/s	83.9 MHz
ECHO-384/512	Kinsy and Uhler [21] / N/A	Fully autonomous	341 cycles per block	Altera Cyclone II	39091 LEs	212 Mbit/s(**)	70.6 MHz
Grøstl-224/256	Jungk et al. [6] / N/A	Fully autonomous	P & Q permutation in parallel	Xilinx Spartan 3	6136 slices	4520 Mbit/s	88.3 MHz
Grøstl-224/256	Submission doc. [7] / N/A	Fully autonomous	P & Q permutation in parallel	Xilinx Virtex 5	1722 slices	10276 Mbit/s	200.7 MHz
Grøstl-224/256	Baldwin et al. [4] / N/A	Core functionality	P & Q permutation in parallel, S-box in BRAM	Xilinx Spartan 3	4827 slices	3660 Mbit/s	71.53 MHz
Grøstl-224/256	Baldwin et al. [4] / N/A	Core functionality	P & Q permutation in parallel, S-box in BRAM	Xilinx Virtex 5	4516 slices	7310 Mbit/s	142.87 MHz
Grøstl-256	Kobayashi et al. [3] / RCIS webpage	Fully autonomous		Xilinx Virtex 5	4057 slices	5171 Mbit/s	101 MHz
Grøstl-384/512	Submission doc. [7] / N/A	Fully autonomous	P & Q permutation in parallel	Xilinx Spartan 3	20233 slices	5901 Mbit/s	80.7 MHz
Grøstl-384/512	Baldwin et al. [4] / N/A	Core functionality	P & Q permutation parallel, S-box in LUTs	Xilinx Spartan 3	17452 slices	3180 Mbit/s	79.61 MHz
Grøstl-384/512	Baldwin et al. [4] / N/A	Core functionality	P & Q permutation parallel, S-box in LUTs	Xilinx Virtex 5	19161 slices	6090 Mbit/s	83.33 MHz
Grøstl-384/512	Submission doc. [7] / N/A	Fully autonomous	P & Q permutation in parallel	Xilinx Virtex 5	5419 slices	15395 Mbit/s	210.5 MHz
Grøstl-384/512	Jungk and Reith [22] / N/A	Fully autonomous	Shared P & Q permutation	Xilinx Spartan 3	8308 slices	3474 Mbit/s	95 MHz
Hamsi-256	Kobayashi et al. [3] / RCIS webpage	Fully autonomous		Xilinx Virtex 5	718 slices	1680 Mbit/s	210 MHz
Hamsi-256	Ramakers and Narinx [25] / Hosted by SHA-3 zoo	Core functionality	Three P and Pf instances each in pipeline	Xilinx Virtex 5	4664 slices	6620 Mbit/s	207 MHz
Keccak	Updated spec. (v1.2) [8] / Submission webpage	Fully autonomous	Core (round function, state register) & IO buffer	Altera Cyclone III	5776 LEs	7500 Mbit/s	133 MHz
Keccak	Updated spec. (v1.2) [8] / Submission webpage	Fully autonomous	Core (round function, state register) & IO buffer	Altera Stratix III	4713 ALUTs	12400 Mbit/s	218 MHz
Keccak	J. Strömbergson [9] / Submission webpage	Fully autonomous	Core (round function, state register) only	Xilinx Spartan 3A	3393 slices	4800 Mbit/s	85 MHz
Keccak	Updated spec. (v1.2) [8] / Submission webpage	Fully autonomous	Core (round function, state register) & IO buffer	Xilinx Virtex 5	1412 slices	6900 Mbit/s	122 MHz
Luffa-256	Namin and Hasan [2] / N/A	Core functionality	Compression function (1 cycle latency) and I/O registers	Altera Stratix III	16552 ALUTs	12042.2 Mbit/s	47.04 MHz
Luffa-256	Kobayashi et al. [3] / RCIS webpage	Fully autonomous		Xilinx Virtex 5	1048 slices	6343 Mbit/s	223 MHz
Luffa-256	Ramakers and Narinx [25] / Hosted by SHA-3 zoo	Core functionality		Xilinx Virtex 5	9611 slices	12290 Mbit/s	48.2 MHz
Shabal	Feron and Francq [10] / N/A	Fully autonomous	36 adders in permutation	Xilinx Virtex 5	1171 slices	2588 Mbit/s	126 MHz
Shabal	Francq and Thuillet [26] / Shabal webpage	Fully autonomous	4 iterations of the permutation unrolled	Xilinx Virtex 5	1715 slices	3242 Mbit/s	76 MHz
Shabal	Baldwin et al. [4] / N/A	Core functionality	36 adders in permutation	Xilinx Spartan 3	2223 slices	740 Mbit/s	71.48 MHz
Shabal	Baldwin et al. [4] / N/A	Core functionality	36 adders in permutation	Xilinx Virtex 5	2768 slices	1450 Mbit/s	138.87 MHz
Shabal-256	Namin and Hasan [2] / N/A	Core functionality	Compression function with I/O registers (latency of 16 clock cycles)	Altera Stratix III	1440 ALUTs	3125.6 Mbit/s	195.35 MHz
Shabal-256	Kobayashi et al. [3] / RCIS webpage	Fully autonomous		Xilinx Virtex 5	1251 slices	1739 Mbit/s	214 MHz
Shabal-512	Detrey et al. [23] / INRIA webpage (see SCM tree)	Fully autonomous	Exploiting SRL16 primitive	Xilinx Virtex 5	153 slices	2051 Mbit/s	256 MHz
Shabal-512	Detrey et al. [23] / INRIA webpage (see SCM tree)	Fully autonomous	Exploiting SRL16 primitive	Xilinx Spartan 3	499 slices	800 Mbit/s	100 MHz
Skein-256-h	Men Long [11] / N/A	Core functionality	UBI component	Xilinx Virtex 5	1001 slices	408.7 Mbit/s	114.9 MHz
Skein-256-256	Stefan Tillich [12] / On request	Fully autonomous	8 Threefish rounds unrolled	Xilinx Virtex 5	937 slices	1751 Mbit/s	68.4 MHz
Skein-256-256	Stefan Tillich [12] / On request	Fully autonomous	8 Threefish rounds unrolled	Xilinx Spartan 3	2421 slices	669 Mbit/s	26.14 MHz
Skein-256-256	Kobayashi et al. [3] / RCIS webpage	Fully autonomous		Xilinx Virtex 5	854 slices	1482 Mbit/s	115 MHz
Skein-512-h	Men Long [11] / N/A	Core functionality	UBI component	Xilinx Virtex 5	1877 slices	817.4 Mbit/s	114.9 MHz
Skein-512-512	Stefan Tillich [12] / On request	Fully autonomous	8 Threefish rounds unrolled	Xilinx Virtex 5	1632 slices	3535 Mbit/s	69.04 MHz
Skein-512-512	Stefan Tillich [12] / On request	Fully autonomous	8 Threefish rounds unrolled	Xilinx Spartan 3	4273 slices	1365 Mbit/s	26.66 MHz

(*) Estimated peak throughput ignoring I/O bottleneck resulting from specific interface: (1536 bits/block) * (70.6 * 10^6 cycles/s) / (273 cycles/block) = 397.22 * 10^6 bits/s.
(**) Estimated peak throughput ignoring I/O bottleneck resulting from specific interface: (1024 bits/block) * (70.6 * 10^6 cycles/s) / (341 cycles/block) = 212.01 * 10^6 bits/s.
(***) CubeHash16/32-h implemented in a similar fashion can be expected to have throughput increased by a factor of about 16.

3.2 Low-Area Implementations (FPGA)

Hash Function Name	Reference / HDL	Impl. Scope	Implementation Details	Technology	Size	Throughput	Clock Frequency
BLAKE-32	Beuchat et al. [13] / N/A	Fully autonomous	Rescheduled G function	Xilinx Spartan-3	124 slices	115 Mbit/s	190.0 MHz
BLAKE-32	Beuchat et al. [13] / N/A	Fully autonomous	Rescheduled G function	Xilinx Virtex-4	124 slices	216 Mbit/s	357.0 MHz
BLAKE-32	Beuchat et al. [13] / N/A	Fully autonomous	Rescheduled G function	Xilinx Virtex-5	56 slices	225 Mbit/s	372.0 MHz
BLAKE-32	Beuchat et al. [13] / N/A	Fully autonomous	Rescheduled G function	Altera Cyclone III	285 LEs	116 Mbit/s	192.0 MHz
BLAKE-32	Submission doc. [1] / Submission webpage	Core functionality	Compression function with 1 G function unit	Xilinx Virtex-II Pro	958 slices	371 Mbit/s	59.0 MHz
BLAKE-32	Submission doc. [1] / Submission webpage	Core functionality	Compression function with 1 G function unit	Xilinx Virtex 4	960 slices	430 Mbit/s	68.0 MHz
BLAKE-32	Submission doc. [1] / Submission webpage	Core functionality	Compression function with 1 G function unit	Xilinx Virtex 5	390 slices	575 Mbit/s	91.0 MHz
BLAKE-64	Beuchat et al. [13] / N/A	Fully autonomous	Rescheduled G function	Xilinx Spartan-3	229 slices	138 Mbit/s	158.0 MHz
BLAKE-64	Beuchat et al. [13] / N/A	Fully autonomous	Rescheduled G function	Xilinx Virtex-4	230 slices	219 Mbit/s	250.0 MHz
BLAKE-64	Beuchat et al. [13] / N/A	Fully autonomous	Rescheduled G function	Xilinx Virtex-5	108 slices	314 Mbit/s	358.0 MHz
BLAKE-64	Beuchat et al. [13] / N/A	Fully autonomous	Rescheduled G function	Altera Cyclone III	542 LEs	123 Mbit/s	140.0 MHz
BLAKE-64	Submission doc. [1] / Submission webpage	Core functionality	Compression function with 1 G function unit	Xilinx Virtex-II Pro	1802 slices	326 Mbit/s	36.0 MHz
BLAKE-64	Submission doc. [1] / Submission webpage	Core functionality	Compression function with 1 G function unit	Xilinx Virtex 4	1856 slices	381 Mbit/s	42.0 MHz
BLAKE-64	Submission doc. [1] / Submission webpage	Core functionality	Compression function with 1 G function unit	Xilinx Virtex 5	939 slices	533 Mbit/s	59.0 MHz
ECHO	Beuchat et al. [24] / N/A	Fully autonomous	Adapted towards FPGA implementation (127 slices and 1 memory block)	Xilinx Virtex 5	127 slices	72 Mbit/s	352.0 MHz
Grøstl-224/256	Jungk et al. [6] / N/A	Fully autonomous	64-bit datapath, P & Q permutation in parallel	Xilinx Spartan 3	2486 slices	404 Mbit/s	63.2 MHz
Grøstl-224/256	Jungk et al. [6] / N/A	Fully autonomous	64-bit datapath, P & Q permutation in parallel	Xilinx Virtex 2 Pro	2754 slices	512 Mbit/s	81.5 MHz
Grøstl-224/256	Jungk and Reith [22] / N/A	Fully autonomous	Shared P & Q permutation, S-Box based on composite field arithmetic	Xilinx Spartan 3	1276 slices	192 Mbit/s	60 MHz
Grøstl-384/512	Jungk and Reith [22] / N/A	Fully autonomous	Shared P & Q permutation, S-Box based on composite field arithmetic	Xilinx Spartan 3	2110 slices	144 Mbit/s	63 MHz
Keccak	Updated spec. (v1.2) [8] / Submission webpage	Using external memory	Small core using system memory	Altera Stratix III	855 ALUTs	96.8 Mbit/s	366 MHz
Keccak	Updated spec. (v1.2) [8] / Submission webpage	Using external memory	Small core using system memory	Altera Cyclone III	1559 LEs	47.8 Mbit/s	181 MHz
Keccak	Updated spec. (v1.2) [8] / Submission webpage	Using external memory	Small core using system memory	Xilinx Virtex 5	444 slices	70.1 Mbit/s	265 MHz
Shabal	Feron and Francq [10] / N/A	Fully autonomous	36 adders in permutation	Xilinx Virtex 5	596 slices (+ 40 DSP blocks)	1142 Mbit/s	109 MHz
Shabal	Baldwin et al. [4] / N/A	Core functionality	1 adder in permutation	Xilinx Spartan 3	1933 slices	540 Mbit/s	89.71 MHz
Shabal	Baldwin et al. [4] / N/A	Core functionality	1 adder in permutation	Xilinx Virtex 5	2307 slices	1330 Mbit/s	222.22 MHz
Shabal-512	Detrey et al. [23] / INRIA webpage (see SCM tree)	Fully autonomous	Exploiting SRL16 primitive	Xilinx Virtex 5	153 slices	2051 Mbit/s	256 MHz
Shabal-512	Detrey et al. [23] / INRIA webpage (see SCM tree)	Fully autonomous	Exploiting SRL16 primitive	Xilinx Spartan 3	499 slices	800 Mbit/s	100 MHz
Skein-256-256	Namin and Hasan [2] / N/A	Core functionality	One round of Threefish iterated	Altera Stratix III	1385 ALUTs	573.9 Mbit/s	161.42 MHz

3.3 High-Speed Implementations (ASIC)

A comparison of implementations of all 14 round 2 candidates has been presented informally at IAIK (Graz University of Technology) on Sept. 16, 2009. The updated presentation slides can be found here.

Hash Function Name	Reference / HDL	Impl. Scope	Implementation Details	Technology	Size	Throughput	Clock Frequency
BLAKE-32	Submission doc. [1] / Submission webpage	Core functionality	Compression function with 8 G function units	UMC 0.18 µm	58.30 kGates	5295 Mbit/s	114 MHz
BLAKE-32	Submission doc. [1] / Submission webpage	Core functionality	Compression function with 4 G function units	UMC 0.18 µm	41.31 kGates	4153 Mbit/s	170 MHz
BLAKE-32	Namin and Hasan [2] / N/A	Core functionality	Compression function with 8 G function units and I/O registers	STM 90 nm	53 kGates	4475 Mbit/s(*)	96.15 MHz
BLAKE-32	Tillich et al. [14] / On request	Fully autonomous	Compression function with 4 G function units with CSAs	UMC 0.18 µm	45.64 kGates	3971 Mbit/s	170.64 MHz
BLAKE-64	Submission doc. [1] / Submission webpage	Core functionality	Compression function with 8 G function units	UMC 0.18 µm	132.47 kGates	5910 Mbit/s	87 MHz
BLAKE-64	Submission doc. [1] / Submission webpage	Core functionality	Compression function with 4 G function units	UMC 0.18 µm	82.73 kGates	4810 Mbit/s	136 MHz
Blue Midnight Wish-256	Namin and Hasan [2] / N/A	Core functionality	Compression function with f0, f1, and f2 unrolled in sequence and I/O registers	STM 90 nm	164 kGates	26665 Mbit/s(*)	52.08 MHz
Blue Midnight Wish-256	Tillich et al. [14] / On request	Fully autonomous	Compression function with f0, f1, and f2 unrolled	UMC 0.18 µm	169.74 kGates	5358 Mbit/s	10.46 MHz
CubeHash16/32-h	Tillich et al. [14] / On request	Fully autonomous	Dynamically reconfigurable r and b parameters, two rounds unrolled	UMC 0.18 µm	58.87 kGates	4665 Mbit/s	145.77 MHz
CubeHash16/32-h	Bernet et al. [20] / N/A	Fully autonomous	One round per cycle	0.13 µm	34.33 kGates	9248 Mbit/s(***)	578 MHz
CubeHash16/32-h	Bernet et al. [20] / N/A	Fully autonomous	Half a round per cycle	0.13 µm	21.54 kGates	8000 Mbit/s(***)	1000 MHz
ECHO-224/256	Lu et al. [5] / N/A	Fully autonomous		0.13 µm	521.1 kGates	14850 Mbit/s	87.1 MHz
ECHO-256	Tillich et al. [14] / On request	Fully autonomous	Four parallel AES rounds, 16 AES MixColumns 32-bit column multipliers	UMC 0.18 µm	141.49 kGates	2246 Mbit/s	141.84 MHz
ECHO-384/512	Lu et al. [5] / N/A	Fully autonomous		0.13 µm	516.8 kGates	7750 Mbit/s	83.3 MHz
Fugue-256	Submission doc. [15] / N/A	Fully autonomous	Four columns of SMIX transformation in parallel (SUPER4_P)	IBM 90 nm	109.85 kGates	13913 Mbit/s	869.5 MHz
Fugue-256	Tillich et al. [14] / On request	Fully autonomous	Four columns of SMIX transformation in parallel	UMC 0.18 µm	46.26 kGates	4092 Mbit/s	255.75 MHz
Grøstl-256	Tillich et al. [14] / On request	Fully autonomous	One shared permutation for P & Q, one pipeline stage	UMC 0.18 µm	58.40 kGates	6290 Mbit/s	270.27 MHz
Grøstl-384/512	Submission doc. [7] / N/A	Fully autonomous	P & Q permutation in parallel	UMC 0.18 µm	341 kGates	6225 Mbit/s	85.1 MHz
Hamsi-256	Junfeng Fan (Hamsi website) [16] / N/A	Fully autonomous		0.13 µm	22 kGates	4940 Mbit/s	1080 MHz
Hamsi-256	Tillich et al. [14] / On request	Fully autonomous	Three instances of P/Pf function unrolled	UMC 0.18 µm	58.66 kGates	5565 Mbit/s	173.91 MHz
Hamsi-512	Junfeng Fan (Hamsi website) [16] / N/A	Fully autonomous		0.13 µm	50 kGates	3970 Mbit/s	820 MHz
JH-256	Tillich et al. [14] / On request	Fully autonomous	320 S-boxes, one round of R8 per cycle	UMC 0.18 µm	58.83 kGates	4991 Mbit/s	380.22 MHz
Keccak	Updated spec. (v1.2) [8] / Submission webpage	Fully autonomous	Core (round function, state register) & IO buffer	ST 0.13 µm	48 kGates	29900 Mbit/s	526 MHz
Keccak	Submission doc. [8] / Submission webpage	Fully autonomous	Core (round function, state register) only	ST 0.13 µm	40 kGates	15000 Mbit/s	500 MHz
Keccak(-256)	Tillich et al. [14] / On request	Fully autonomous	One instance of Keccak-f round	UMC 0.18 µm	56.32 kGates	21229 Mbit/s	487.80 MHz
Luffa-224/256	Knežević and Verbauwhede [17] / Author's webpage	Fully autonomous	Three permutation blocks in parallel (64 S-boxes, 4 MixWord blocks each)	UMC 0.13 µm	30.83 kGates	31960 Mbit/s	1124 MHz
Luffa-256	Namin and Hasan [2] / N/A	Core functionality	Compression function (1 cycle latency) and I/O registers	STM 90 nm	122 kGates	25702 Mbit/s(*)	100.4 MHz
Luffa-224/256	Tillich et al. [14] / On request	Fully autonomous	Three permutation blocks in parallel (64 S-boxes, 4 MixWord blocks each)	UMC 0.18 µm	44.97 kGates	13741 Mbit/s	483.09 MHz
Luffa-384	Knežević and Verbauwhede [17] / Author's webpage	Fully autonomous	Four permutation blocks in parallel (64 S-boxes, 4 MixWord blocks each)	UMC 0.13 µm	50.07 kGates	23126 Mbit/s	813 MHz
Luffa-512	Knežević and Verbauwhede [17] / Author's webpage	Fully autonomous	Five permutation blocks in parallel (64 S-boxes, 4 MixWord blocks each)	UMC 0.13 µm	65.1 kGates	19617 Mbit/s	690 MHz
Shabal-256	Namin and Hasan [2] / N/A	Core functionality	Compression function with I/O registers (latency of 16 clock cycles)	STM 90 nm	20 kGates	4408 Mbit/s(*)	413.22 MHz
Shabal-256	Tillich et al. [14] / On request	Fully autonomous	One word rotation per cycle, 50 cycles per block	UMC 0.18 µm	54.19 kGates	3282 Mbit/s	320.51 MHz
Shabal	Bernet et al. [20] / N/A	Fully autonomous	One word rotation per cycle, 52 cycles per block	0.13 µm	41.32 kGates	6351 Mbit/s(***)	645 MHz
SHAvite-3256	Tillich et al. [14] / On request	Fully autonomous	Four AES rounds (two for compression, two for message expansion)	UMC 0.18 µm	57.39 kGates	3152 Mbit/s	227.79 MHz
SIMD-256(**)	Tillich et al. [14] / On request	Fully autonomous	Two FFT-64 with two FFT-8 and 16 multipliers (8x8 bit) each	UMC 0.18 µm	104.17 kGates	924 Mbit/s	64.93 MHz
Skein-256-256	Stefan Tillich [12] / On request	Fully autonomous	8 Threefish rounds unrolled	UMC 0.18 µm	53.87 kGates	1762 Mbit/s	68.8 MHz
Skein-256-256	Namin and Hasan [2] / N/A	Core functionality	All 72 Threefish rounds unrolled	STM 90 nm	369 kGates	3126 Mbit/s(*)	12.21 MHz
Skein-256-256	Tillich et al. [14] / On request	Fully autonomous	8 Threefish rounds unrolled	UMC 0.18 µm	58.61 kGates	1882 Mbit/s	73.52 MHz
Skein-512-512	Tillich et al. [14] / On request	Fully autonomous	8 Threefish rounds unrolled	UMC 0.18 µm	102.04 kGates	2502 Mbit/s	48.87 MHz

(*) Estimated peak throughput for the minimal delay of compression function: 1000 * (Input Size in bits) / [(Compression Function Delay in ns) * (Number of Cycles)] = Throughput in Mbit/s.
(**) Implementation of round-one variant.
(***) Estimated peak throughput: Throughput for CubeHash8/1-h implementation * 16.

3.4 Low-Area Implementations (ASIC)

Hash Function Name	Reference / HDL	Impl. Scope	Implementation Details	Technology	Size	Throughput	Clock Frequency
BLAKE-32	Tillich et al. [18] / N/A	Fully autonomous	One G function in 11 cycles	AMS 0.35 µm	25.57 kGates	15.4 Mbit/s	31.25 MHz
BLAKE-32	Submission doc. [1] / Submission webpage	Core functionality	Compression function with a single G function unit	UMC 0.18 µm	10.54 kGates	253 Mbit/s	40 MHz
BLAKE-32	Submission doc. [1] / Submission webpage	Core functionality	Compression function with a half G function unit	UMC 0.18 µm	9.89 kGates	127 Mbit/s	40 MHz
BLAKE-64	Submission doc. [1] / Submission webpage	Core functionality	Compression function with a single G function unit	UMC 0.18 µm	20.61 kGates	181 Mbit/s	20 MHz
BLAKE-64	Submission doc. [1] / Submission webpage	Core functionality	Compression function with a half G function unit	UMC 0.18 µm	19.46 kGates	91 Mbit/s	20 MHz
CubeHash16/32-h	Bernet et al. [20] / N/A	Fully autonomous	Process two 32-bit words per cycle, 64 cycles per round	0.13 µm	7.63 kGates	32 Mbit/s(****)	100 MHz
ECHO-224/256	Lu et al. [5] / N/A	Fully autonomous		0.13 µm	82.8 kGates	373 Mbit/s	66.6 MHz
Fugue-256	Submission doc. [15] / N/A	Fully autonomous	One SMIX transformation (SUPER1_L)	IBM 90 nm	59.22 kGates	2000 Mbit/s	500 MHz
Grøstl-224/256	Tillich et al. [18] / N/A	Fully autonomous	64-bit datapath, P & Q permutation shared	AMS 0.35 µm	14.62 kGates	145.9 Mbit/s	55.87 MHz
Grøstl-224/256	Grøstl website [19] / N/A	Fully autonomous	64-bit datapath, P & Q permutation shared	UMC 0.18 µm	17 kGates	645 Mbit/s	246.9 MHz
Keccak	Updated spec. (v1.2) [8] / Submission webpage	Using external memory	Small core using system memory	ST 0.13 µm	6.5 kGates	176.4 Mbit/s(*)	666.7 MHz
Keccak	Updated spec. (v1.2) [8] / Submission webpage	Using external memory	Small core using system memory, clock freq. limited to 200 MHz	ST 0.13 µm	5 kGates	52.9 Mbit/s(**)	200 MHz
Luffa-224/256	Knežević and Verbauwhede [17] / Author's webpage	Fully autonomous	One permutation block (64 S-boxes, 4 MixWord blocks)	UMC 0.13 µm	18.26 kGates	2461 Mbit/s	250 MHz
Luffa-256	Mikami et al. [27] / N/A	Fully autonomous	One permutation block (64 S-boxes, 4 MixWord blocks)	UMC 0.13 µm	10.34 kGates	538 Mbit/s	806 MHz
Luffa-384	Knežević and Verbauwhede [17] / Author's webpage	Fully autonomous	6 S-boxes, 1 MixWord	TSMC 90 nm	27.13 kGates	1882 Mbit/s	250 MHz
Luffa-512	Knežević and Verbauwhede [17] / Author's webpage	Fully autonomous	One permutation block (64 S-boxes, 4 MixWord blocks)	UMC 0.13 µm	37.35 kGates	1524 Mbit/s	250 MHz
Shabal	Bernet et al. [20] / N/A	Fully autonomous	One adder, one subtractor, one incrementer. 165 cycles per block	0.13 µm	23.32 kGates	310 Mbit/s	100 MHz
Skein-256-256	Tillich et al. [18] / N/A	Fully autonomous	64-bit datapath	AMS 0.35 µm	12.89 kGates	19.8 Mbit/s	80 MHz
Skein-256-256	Namin and Hasan [2] / N/A	Core functionality	One round of Threefish iterated	STM 90 nm	21 kGates	1018.8 Mbit/s(***)	286.53 MHz

(*) Estimation for 64-bit memory interface: (1024 bits/permutation) * (666.7 * 10^6 cycles/s) / (3870 cycles/permutation) = 176.41 * 10^6 bits/s
(**) Estimation for 64-bit memory interface: (1024 bits/permutation) * (200 * 10^6 cycles/s) / (3870 cycles/permutation) = 52.92 * 10^6 bits/s
(***) Estimated peak throughput for the minimal delay of compression function: 1000 * (Input Size in bits) / [(Compression Function Delay in ns) * (Number of Cycles)] = Throughput in Mbit/s
(****) Estimated peak throughput: Throughput for CubeHash8/1-h implementation * 16.

4 Comparative Studies

This section summarizes the reported results of publications which examined more than one round-two candidate in a similar setup.

4.1 Blake, BMW, Luffa, Shabal, Skein

Reference	HDL	Category	Impl. Scope	Technology
Namin and Hasan [2]	N/A	High-speed FPGA	Core functionality	Altera Stratix III

Hash Function Name	Impl. Details	Size	Throughput	Clock Frequency
BLAKE-32	Compression function with 8 G function units and I/O registers	5435 ALUTs	2186.2 Mbit/s	46.97 MHz
Blue Midnight Wish-256	Compression function with f0, f1, and f2 unrolled in sequence and I/O registers	12917 ALUTs	4889.6 Mbit/s	9.55 MHz
Luffa-256	Compression function (1 cycle latency) and I/O registers	16552 ALUTs	12042.2 Mbit/s	47.04 MHz
Shabal-256	Compression function with I/O registers (latency of 16 clock cycles)	1440 ALUTs	3125.6 Mbit/s	195.35 MHz
Skein-256-256	All 72 Threefish rounds unrolled (device too small)	N/A	N/A	N/A

---

Reference	HDL	Category	Impl. Scope	Technology
Namin and Hasan [2]	N/A	High-speed ASIC	Core functionality	STM 90 nm

Hash Function Name	Impl. Details	Size	Throughput	Clock Frequency
BLAKE-32	Compression function with 8 G function units and I/O registers	53 kGates	4475 Mbit/s(*)	96.15 MHz
Blue Midnight Wish-256	Compression function with f0, f1, and f2 unrolled in sequence and I/O registers	164 kGates	26665 Mbit/s(*)	52.08 MHz
Luffa-256	Compression function (1 cycle latency) and I/O registers	122 kGates	25702 Mbit/s(*)	100.4 MHz
Shabal-256	Compression function with I/O registers (latency of 16 clock cycles)	20 kGates	4408 Mbit/s(*)	413.22 MHz
Skein-256-256	All 72 Threefish rounds unrolled	369 kGates	3126 Mbit/s(*)	12.21 MHz

(*) Estimated peak throughput for the minimal delay of compression function: 1000 * (Input Size in bits) / [(Compression Function Delay in ns) * (Number of Cycles)] = Throughput in Mbit/s.

4.2 Blake, CubeHash, ECHO, Grøstl, Hamsi, Luffa, Shabal, Skein

Reference	HDL	Category	Impl. Scope	Technology
Kobayashi et al. [3]	RCIS webpage	High-speed FPGA	Fully autonomous	Xilinx Virtex 5

Hash Function Name	Impl. Details	Size	Throughput	Clock Frequency
BLAKE-32		1660 slices	2676 Mbit/s	115 MHz
CubeHash16/32-256		590 slices	2960 Mbit/s	185 MHz
ECHO-256		3556 slices	1614 Mbit/s	104 MHz
Grøstl-256		4057 slices	5171 Mbit/s	101 MHz
Hamsi-256		718 slices	1680 Mbit/s	210 MHz
Luffa-256		1048 slices	6343 Mbit/s	223 MHz
Shabal-256		1251 slices	1739 Mbit/s	214 MHz
Skein-256		854 slices	1482 Mbit/s	115 MHz

4.3 CubeHash, Grøstl, Shabal

Reference	HDL	Category	Impl. Scope	Technology
Baldwin et al. [4]	N/A	High-speed FPGA	Core functionality	Xilinx Spartan 3

Hash Function Name	Impl. Details	Size	Throughput	Clock Frequency
CubeHash8/1-256(*)	2 compression functions unrolled	3268 slices	70 Mbit/s	37.9 MHz
Grøstl-224/256	P & Q permutation in parallel, S-box in BRAM	4827 slices	3660 Mbit/s	71.53 MHz
Grøstl-384/512	P & Q permutation parallel, S-box in LUTs	17452 slices	3180 Mbit/s	79.61 MHz
Shabal	36 adders in permutation	2223 slices	740 Mbit/s	71.48 MHz

(*) CubeHash16/32-h implemented in a similar fashion can be expected to have throughput increased by a factor of about 16.

---

Reference	HDL	Category	Impl. Scope	Technology
Baldwin et al. [4]	N/A	High-speed FPGA	Core functionality	Xilinx Virtex 5

Hash Function Name	Impl. Details	Size	Throughput	Clock Frequency
CubeHash8/1-256(*)	1 iterated compression function	1178 slices	160 Mbit/s	166.8 MHz
Grøstl-224/256	P & Q permutation in parallel, S-box in BRAM	4516 slices	7310 Mbit/s	142.87 MHz
Grøstl-384/512	P & Q permutation parallel, S-box in LUTs	19161 slices	6090 Mbit/s	83.33 MHz
Shabal	36 adders in permutation	2768 slices	1450 Mbit/s	138.87 MHz

(*) CubeHash16/32-h implemented in a similar fashion can be expected to have throughput increased by a factor of about 16.

4.4 All 14 Round-Two Candidates

An interactive graphical comparison of various area-performance tradeoffs of this study can be found here.

Reference	HDL	Category	Impl. Scope	Technology
Tillich et al. [14]	On request	High-speed ASIC	Fully autonomous	UMC 0.18 µm

Hash Function Name	Impl. Details	Size	Throughput	Clock Frequency
BLAKE-32	Compression function with 4 G function units with CSAs	45.64 kGates	3971 Mbit/s	170.64 MHz
Blue Midnight Wish-256	Compression function with f0, f1, and f2 unrolled	169.74 kGates	5358 Mbit/s	10.46 MHz
CubeHash16/32-h	Dynamically reconfigurable r and b parameters, two rounds unrolled	58.87 kGates	4665 Mbit/s	145.77 MHz
ECHO-256	Four parallel AES rounds, 16 AES MixColumns 32-bit column multipliers	141.49 kGates	2246 Mbit/s	141.84 MHz
Fugue-256	Four columns of SMIX transformation in parallel	46.26 kGates	4092 Mbit/s	255.75 MHz
Grøstl-256	One shared permutation for P & Q, one pipeline stage	58.40 kGates	6290 Mbit/s	270.27 MHz
Hamsi-256	Three instances of P/Pf function unrolled	58.66 kGates	5565 Mbit/s	173.91 MHz
JH-256	320 S-boxes, one round of R8 per cycle	58.83 kGates	4991 Mbit/s	380.22 MHz
Keccak(-256)	One instance of Keccak-f round	56.32 kGates	21229 Mbit/s	487.80 MHz
Luffa-224/256	Three permutation blocks in parallel (64 S-boxes, 4 MixWord blocks each)	44.97 kGates	13741 Mbit/s	483.09 MHz
Shabal-256	One word rotation per cycle, 50 cycles per block	54.19 kGates	3282 Mbit/s	320.51 MHz
SHAvite-3256	Four AES rounds (two for compression, two for message expansion)	57.39 kGates	3152 Mbit/s	227.79 MHz
SIMD-256(*)	Two FFT-64 with two FFT-8 and 16 multipliers (8x8 bit) each	104.17 kGates	924 Mbit/s	64.93 MHz
Skein-256-256	8 Threefish rounds unrolled	58.61 kGates	1882 Mbit/s	73.52 MHz
Skein-512-512	8 Threefish rounds unrolled	102.04 kGates	2502 Mbit/s	48.87 MHz

(*) Implementation of round-one variant.

4.5 BLAKE, Grøstl, Skein

Reference	HDL	Category	Impl. Scope	Technology
Tillich et al. [18]	N/A	Low-area ASIC	Fully autonomous	AMS 0.35 µm

Hash Function Name	Impl. Details	Size	Throughput	Clock Frequency
BLAKE-32	One G function in 11 cycles	25.57 kGates	15.4 Mbit/s	31.25 MHz
Grøstl-224/256	64-bit datapath, P & Q permutation shared	14.62 kGates	145.9 Mbit/s	55.87 MHz
Skein-256-256	64-bit datapath	12.89 kGates	19.8 Mbit/s	80 MHz

4.6 ECHO, Hamsi, Luffa

Reference	HDL	Category	Impl. Scope	Technology
Ramakers and Narinx [25]	Hosted by SHA-3 zoo	High-speed FPGA	Core functionality	Xilinx Virtex 5

Hash Function Name	Impl. Details	Size	Throughput	Clock Frequency
ECHO-256	4 x 2 AES round instances with pipeline register in BigSubWords	15006 slices	23860 Mbit/s	139 MHz
Hamsi-256	Three P and Pf instances each in pipeline	4664 slices	6620 Mbit/s	207 MHz
Luffa-256		9611 slices	12290 Mbit/s	48.2 MHz

5 References

[1] Jean-Philippe Aumasson, Luca Henzen, Willi Meier, and Raphael C.-W. Phan. SHA-3 proposal BLAKE (version 1.3). Available online at http://131002.net/blake/blake.pdf.

[2] A. H. Namin and M. A. Hasan. Hardware Implementation of the Compression Function for Selected SHA-3 Candidates. Available online at http://www.vlsi.uwaterloo.ca/~ahasan/hasan_report.html.

[3] Kazuyuki Kobayashi, Jun Ikegami, Shin'ichiro Matsuo, Kazuo Sakiyama, and Kazuo Ohta. Evaluation of Hardware Performance for the SHA-3 Candidates Using SASEBO-GII. IACR Eprint report 2010/010. Available online at http://eprint.iacr.org/2010/010.pdf.

[4] Brian Baldwin, Andrew Byrne, Mark Hamilton, Neil Hanley, Robert P. McEvoy, Weibo Pan, and William P. Marnane. FPGA Implementations of SHA-3 Candidates: CubeHash, Grøstl, LANE, Shabal and Spectral Hash. IACR Eprint report 2009/342. Available online at http://eprint.iacr.org/2009/342.pdf.

[5] Liang Lu, Maire O'Neil, and Earl Swartzlander. Hardware Evaluation of SHA-3 Hash Function Candidate ECHO. Presentation at the Clauce Shannon Institute Workshop on Coding and Cryptography 2009. Slides available online at http://www.ucc.ie/en/crypto/CodingandCryptographyWorkshop/TheClaudeShannonWorkshoponCodingCryptography2009/DocumentFile,75649,en.pdf.

[6] Bernhard Jungk, Steffen Reith, and Jürgen Apfelbeck. On Optimized FPGA Implementations of the SHA-3 Candidate Grøstl. IACR Eprint report 2009/206. Available online at http://eprint.iacr.org/2009/206.pdf.

[7] Praveen Gauravaram, Lars R. Knudsen, Krystian Matusievicz, Florian Mendel, Christian Rechberger, Martin Schläffer, and Søren S. Thomsen. Grøstl - a SHA-3 candidate (October 31, 2008). Available online at http://www.groestl.info/Groestl.pdf.

[8] Guido Bertoni, Joan Daemen, Michaël Peeters, and Gilles van Assche. KECCAK sponge function family main document (Version 1.2, April 23, 2009). Available online at http://keccak.noekeon.org/Keccak-main-1.2.pdf.

[9] Joachim Strömbergson. Implementation of the Keccak Hash Function in FPGA Devices. Available online at http://www.strombergson.com/files/Keccak_in_FPGAs.pdf.

[10] Romain Feron and Julien Francq. FPGA Implementation of Shabal: Our First Results (Version 2.0, February 19, 2010). Available online at http://www.shabal.com/wp-content/uploads/2010/03/FPGA-Implementation-of-Shabal-First-ResultsV2.0.pdf.

[11] Men Long. Implementing Skein Hash Function on Xilinx Virtex-5 FPGA Platform (Version 0.7, February 2, 2009). Available online at http://www.skein-hash.info/sites/default/files/skein_fpga.pdf.

[12] Stefan Tillich. Hardware Implementation of the SHA-3 Candidate Skein. IACR Eprint report 2009/159. Available online at http://eprint.iacr.org/2009/159.pdf.

[13] Jean-Luc Beuchat, Eiji Okamoto, and Teppei Yamazaki. Compact Implementations of BLAKE-32 and BLAKE-64 on FPGA. IACR Eprint report 2010/173. Available online at http://eprint.iacr.org/2010/173.pdf.

[14] Stefan Tillich, Martin Feldhofer, Mario Kirschbaum, Thomas Plos, Jörn-Marc Schmidt, and Alexander Szekely. High-Speed Hardware Implementations of BLAKE, Blue Midnight Wish, CubeHash, ECHO, Fugue, Grøstl, Hamsi, JH, Keccak, Luffa, Shabal, SHAvite-3, SIMD, and Skein. IACR Eprint report 2009/510. Available online at http://eprint.iacr.org/2009/510.pdf.

[15] Shai Halevi, William E. Hall, and Charanjit S. Jutla. The Hash Function Fugue (October 30, 2008). Available online at http://domino.research.ibm.com/comm/research_projects.nsf/pages/fugue.index.html/$FILE/NIST-submission-Oct08-fugue.pdf.

[16] Junfeng Fan. Hardware Evaluation of The Hash Function Hamsi. Available online at http://homes.esat.kuleuven.be/~okucuk/hamsi/implementations.html.

[17] Miroslav Knezevic and Ingrid Verbeiwhede. Hardware Evaluation of the Luffa Hash Family. 4th Workshop on Embedded Systems Security 2009. Available online at http://www.cosic.esat.kuleuven.be/publications/article-1282.pdf.

[18] Stefan Tillich, Martin Feldhofer, Wolfgang Issovits, Thomas Kern, Hermann Kureck, Michael Mühlberghuber, Georg Neubauer, Andreas Reiter, Armin Köfler, and Mathias Mayrhofer. Compact Hardware Implementations of the SHA-3 Candidates ARIRANG, BLAKE, Grøstl, and Skein. IACR Eprint report 2009/349. Available online at http://eprint.iacr.org/2009/349.pdf.

[19] Grøstl website. http://www.groestl.info/.

[20] Markus Bernet, Luca Henzen, Hubert Kaeslin, Norbert Felber, and Wolfgang Fichtner. Hardware Implementations of the SHA-3 Candidates Shabal and CubeHash. 52nd IEEE International Midwest Symposium on Circuits and Systems, 2009. Available online at http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=5236043.

[21] Michel Kinsy and Richard Uhler. SHA-3: FPGA Implementation of ESSENCE and ECHO Hash Algorithm Candidates Using Bluespec. Available online at http://csg.csail.mit.edu/6.375/6_375_2009_www/projects/group1_report.pdf.

[22] Bernhard Jungk and Steffen Reith. On FPGA-based implementations of Grøstl. IACR Eprint report 2010/260. Available online at http://eprint.iacr.org/2010/260.pdf.

[23] Jérémie Detrey, Pierre Gaudry, and Karim Khalfallah. A Low-Area yet Performant FPGA Implementation of Shabal. IACR Eprint report 2010/292. Available online at http://eprint.iacr.org/2010/292.pdf.

[24] Jean-Luc Beuchat, Eiji Okamoto, and Teppei Yamazaki. A Compact FPGA Implementation of the SHA-3 Candidate ECHO. IACR Eprint report 2010/364. Available online at http://eprint.iacr.org/2010/364.pdf.

[25] Wim Ramakers and Hans Narinx. Implementation and evaluation of SHA-3 candidates on FPGA. Extended abstract of Master Thesis "Implementatie en Evaluatie van SHA-3-Kandidaten op FPGA" (Dutch). Extended abstract available online at http://ehash.iaik.tugraz.at/uploads/1/12/Ramakers_Narinx2010ECHO-Hamsi-Luffa_ExtendedAbstract_ENGLISH.pdf. Full thesis available online at http://ehash.iaik.tugraz.at/uploads/6/62/Ramakers_Narinx2010ECHO-Hamsi-Luffa_Thesis_DUTCH.pdf.

[26] Julien Francq and Céline Thuillet. Unfolding Method for Shabal on Virtex-5 FPGAs: Concrete Results. IACR Eprint report 2010/406. Available online at http://eprint.iacr.org/2010/406.pdf.

[27] Shugo Mikami, Nagamasa Mizushima, Setsuko Nakamura, and Dai Watanabe. A Compact Hardware Implementation of SHA-3 Candidate Luffa. Available online at http://www.sdl.hitachi.co.jp/crypto/luffa/ACompactHardwareImplementationOfSHA-3CandidateLuffa_20100810.pdf.