Techniques For Implementing Multipliers in Stratix, Stratix GX & Cyclone Devices

Techniques for Implementing
Multipliers in Stratix,
Stratix GX & Cyclone Devices
August 2003, ver. 1.0 Application Note 306
Introduction Stratix™, Stratix GX, and Cyclone™ FPGAs have dedicated architectural
features that make it easy to implement high-performance multipliers.
Stratix and Stratix GX devices feature embedded high-performance
multiplier-accumulators (MACs) in dedicated digital signal processing
(DSP) blocks. DSP blocks can operate at data rates above 300 million
samples per second (MSPS), making Stratix and Stratix GX FPGAs ideal
for high-speed DSP applications. In addition to the dedicated DSP blocks,
designers can also use the devices’ TriMatrix™ memory blocks to
implement variable depth/width, high-performance soft multipliers. For
example, designers can implement TriMatrix memory blocks as look-up
tables (LUTs) that contain partial results from multiplication of input data
with coefficients. Cyclone devices have M4K memory blocks which can
be used as LUTs to implement variable depth/width high-performance
soft multipliers for low cost, high volume DSP applications.
Stratix, Stratix GX, and Cyclone FPGAs can implement the multiplier
types shown in Table 1.
Table 1. Supported Multiplier Implementations
Multiplier Type Description Stratix Stratix GX Cyclone

Soft multiplier These multipliers are implemented as LUTs in memory, v v v
which contains all possible partial results from
multiplication. There are five soft multiplier modes:
■ Parallel multiplication
■ Semi-parallel multiplication
■ Sum of multiplication
■ Hybrid multiplication
■ Fully variable multipliers
Multipliers using DSP These multipliers are implemented in dedicated DSP v v -
blocks or logic blocks or LEs using the lpm_mult, altmult_add, or
elements (LEs) altmult_accum megafunctions.
Hard multiplier These multipliers are implemented in a combination of v v -
DSP blocks and LEs.
Altera Corporation Quartus II Version 3.0 1

AN-306-1.0 Preliminary
Techniques for Implementing Multipliers in Stratix, Stratix GX & Cyclone Devices
Tables 2 and 3 show the total number of multipliers available in Stratix

and Stratix GX devices using DSP blocks and soft multipliers. Table 4
shows the total number of soft multipliers available in Cyclone devices.
Table 2. Number of Multipliers in Stratix Devices
DSP Blocks Soft Multipliers (16 × 16) Total Multipliers

Device
(18 × 18) Notes (1) Notes (2), (3)
EP1S10 24 81 105 (4.38)
EP1S20 40 132 172 (4.30)
EP1S25 40 190 230 (5.57)
EP1S30 48 241 289 (6.02)
EP1S40 56 280 336 (6.00)
EP1S60 72 434 506 (7.03)
EP1S80 88 558 646 (7.34)
Notes to Table 2:
(1) Soft multipliers implemented in sum of multiplication mode. RAM blocks
configured with 18-bit data widths and sum of coefficients up to 18 bits.
(2) The number in parentheses represents the increase factor, which is the total
number of multipliers with soft multipliers divided by the number of 18 × 18
multipliers supported by DSP blocks only.
(3) The total number of multipliers may vary according to the multiplier mode used.
Table 3. Number of Multipliers in Stratix GX Devices
DSP Blocks Soft Multipliers (16 × 16) Total Multipliers

Device
(18 × 18) Notes (1) Notes (2), (3)
EP1SGX10C 24 81 105 (4.38)
EP1SGX10D 24 81 105 (4.38)
EP1SGX25C 40 190 230 (5.57)
EP1SGX25D 40 190 230 (5.57)
EP1SGX25F 40 190 230 (5.57)
EP1SGX40D 56 280 336 (6.00)
EP1SGX40G 56 280 336 (6.00)
Notes to Table 3:
(2) The number in parentheses represents the increase factor, which is the total
number of multipliers with soft multipliers divided by the number of 18 × 18
multipliers supported by DSP blocks only.
2 Quartus II Version 3.0 Altera Corporation

Preliminary
Memory Blocks
Table 4. Number of Multipliers in Cyclone Devices
Device Soft Multipliers (16 × 16) Notes (1), (2)

EP1C3 11
EP1C4 14
EP1C6 17
EP1C12 45
EP1C20 56
Notes to Table 4:
This application note describes the dedicated memory and DSP blocks,
the supported multiplier types, and includes an example of each type.
Memory Blocks The Stratix and Stratix GX TriMatrix memories consist of three types of
RAM blocks: M512, M4K, and M-RAM. The M512 and M4K RAM blocks
are memory blocks with a maximum width of 18 and 36 bits, respectively,
and a maximum performance of approximately 300 MHz, which is ideal
for implementing soft multipliers.
Tables 5 and 6 show the available TriMatrix memory blocks in Stratix and
Stratix GX devices, respectively.
Table 5. Stratix Memory Blocks
M512 RAM M4K RAM M-RAM

Device Total RAM Bits
(32 × 18 Bits) (128 × 36 Bits) (4K × 144 Bits)
EP1S10 94 60 1 920,448
EP1S20 194 82 2 1,669,248
EP1S25 224 138 2 1,944,576
EP1S30 295 171 4 3,317,184
EP1S40 384 183 4 3,423,744
EP1S60 574 292 6 5,215,104
EP1S80 767 364 9 7,427,520

Preliminary
Table 6. Stratix GX TriMatrix Memory Blocks
M512 RAM M4K RAM M-RAM

Device Total RAM Bits
(32 × 18 Bits) (128 × 36 Bits) (4K × 144 Bits)
EP1SGX10C 94 60 1 920,448
EP1SGX10D 94 60 1 920,448
EP1SGX25C 224 138 2 1,944,576
EP1SGX25D 224 138 2 1,944,576
EP1SGX25F 224 138 2 1,944,576
EP1SGX40D 384 183 4 3,423,744
EP1SGX40G 384 183 4 3,423,744
The Cyclone M4K memory blocks have a maximum width of 36 bits and
a maximum performance of approximately 200 MHz. Table 7 shows the
number of Cyclone M4K memory blocks.
Table 7. Cyclone M4K Memory Blocks
Device M4K RAM (128 × 36 Bits)

EP1C3 13
EP1C4 17
EP1C6 20
EP1C12 52
EP1C20 64
Table 8 shows the possible configurations of the M512, M4K, and M-RAM
blocks found in Stratix, Stratix GX, and Cyclone devices.
Table 8. M512, M4K & M-RAM Memory Configurations (Part 1 of 2)
M512 RAM Block M4K RAM Block M-RAM Block

(32 × 18 Bits) (128 × 36 Bits) (4K × 144 Bits)
512 × 1 4K × 1 64K × 8
256 × 2 2K × 2 64K × 9
128 × 4 1K × 4 32K × 16
64 × 8 512 × 8 32K × 18
64 × 9 512 × 9 16K × 32
32 × 16 256 × 16 16K × 36
32 × 18 256 × 18 8K × 64

Preliminary
DSP Blocks
Table 8. M512, M4K & M-RAM Memory Configurations (Part 2 of 2)
M512 RAM Block M4K RAM Block M-RAM Block

(32 × 18 Bits) (128 × 36 Bits) (4K × 144 Bits)
- 128 × 32 8K × 72
- 128 × 36 4K × 128
- - 4K × 144
DSP Blocks Stratix and Stratix GX devices contain dedicated DSP blocks for
implementing high-speed multiplication functions within the FPGA.
Tables 9 and 10 show the number of DSP blocks in Stratix and Stratix GX
respectively.
Table 9. Number of DSP Blocks in Stratix Devices Note (1)
Device DSP Blocks 9 × 9 Multipliers 18 × 18 Multipliers 36 × 36 Multipliers

EP1S10 6 48 24 6
EP1S20 10 80 40 10
EP1S25 10 80 40 10
EP1S30 12 96 48 12
EP1S40 14 112 56 14
EP1S60 18 144 72 18
EP1S80 22 176 88 22
Note to Table 9:
(1) Each device has either the number of 9 × 9-, 18 × 18-, or 36 × 36-bit multipliers shown. The total number of
multipliers for each device is not the sum of all the multipliers.
Table 10. Number of DSP Blocks in Stratix GX Devices (Part 1 of 2) Note (1)

EP1SGX10C 6 48 24 6
EP1SGX10D 6 48 24 6
EP1SGX25C 10 80 40 10
EP1SGX25D 10 80 40 10
EP1SGX25F 10 80 40 10
EP1SGX40D 14 112 56 14

Preliminary
Table 10. Number of DSP Blocks in Stratix GX Devices (Part 2 of 2) Note (1)

EP1SGX40G 14 112 56 14
Note to Table 10:

(1) Each device has either the number of 9 × 9-, 18 × 18-, or 36 × 36-bit multipliers shown. The total number of
multipliers for each device is not the sum of all the multipliers.
DSP Arithmetic DSP is a multiplication-intensive technology and to achieve high speeds,

these multiplication operations must be accelerated. This section
Basics provides basic information on the mathematical theory and algorithms
behind common DSP arithmetic implementations.
Multiplication
The base of many DSP algorithms is multiplication in which a multiplier
is multiplied to a multiplicand. In this operation, each element of the
multiplier is multiplied by each bit of the multiplicand. Then, the partial
product of each multiplication is accumulated according to the weight of
the partial product, where the weight indicates the location of a bit
corresponding to other bits. For example, if a partial product of bits 4
through 7 is added to a partial product of bits 0 through 3, the partial
product of 4 through 7 is shifted according to their weight and then
accumulated to the partial product of previous stages. Figure 1 shows a
simple 2 × 2 multiplication of multiplier a1a0 to multiplicand b1b0.

Preliminary
DSP Arithmetic Basics
Figure 1. Multiplication of Two 2-Bit Numbers

a0
b1
b0
a1
b1 b0
b1 b0
x a1 a0
a0b1 a0b0
+ a1b1 a1b0
c3 c2 c1 c0
carry_in
Half Adder Half Adder
carry_out sum carry_out sum
c3 c2 c1 c0
Distributed Arithmetic
Distributed arithmetic is a method of performing multiplication by
distributing the operation over many LUTs. Figure 2 shows a four-
product MAC function that uses sequential shift and add to multiply four
pairs, and then sums their partial product to obtain a final result. Each
multiplier forms partial products by multiplying the multiplicand by one
bit of the input data (multiplier) at a time, using an AND gate.

Preliminary
Figure 2. Distributed Arithmetic with Four Constant Multiplicands
c0 Scaling Accumulator
w SREG
>> 1
c1
x SREG
D Q
c2
y SREG CLK
c3
wc0 + xc1 + yc2 + zc3
z SREG
At the end of the process, each partial product result of each input bit is
summed prior to the final scaling accumulator stage, which performs a
shift-accumulate.
The distributed-arithmetic circuit simultaneously performs four

multiplications and sums the results when all of the products are
completed. The scaling accumulator shifts the sums of partial products
according to the appropriate number of bits and accumulates the result to
provide the final multiplier output.
Distributed Arithmetic in LUTs

Figure 3 shows how to implement distributed arithmetic using LUTs: the
combined product and adder tree are reduced for the LUT
implementation. In this example, the LUT contains the sums of constant
coefficients for all possible input combinations to the LUT. The sums of
the bits from the LUTs are added together in the scaling accumulator and
shifted by the appropriate weights.

Preliminary
Implementing Soft Multipliers Using Memory Blocks
Figure 3. Four-Bit Multiplication with Constant Coefficients Note (1)
c0
w
Addr Data
0000 0
0001 c0
c1
x 0010 c1
0011 c0 + c 1
c2
y
1110 c1 + c 2 + c 3
1111 c0 + c 1 + c 2 + c 3
c3
z
Note to Figure 3:
(1) c0 to c3 are constant coefficients.
The addressing method and data values stored in the LUT in Figure 3
apply to the sum of multiplication operation mode. The addressing
method and LUT data values vary depending on the multiplier
implementation mode.
Implementing You can use the Stratix and Stratix GX M512 or M4K RAM memory blocks
and Cyclone M4K RAM memory blocks as LUTs to implement
Soft Multipliers multiplication for DSP applications. Combinations of the coefficient
Using Memory results are pre-calculated and stored in the M512 or M4K RAM blocks as
a LUT. The address port of the RAM block represents one of the
Blocks multiplication operands. The content of the RAM block at each address
represents a unique multiplication result calculated between the input
operand and a known coefficient value based on the multiplier mode
implemented.
The five soft multiplier modes supported by Stratix and Stratix GX

devices are:
■ Parallel Multiplication—Multiple memories produce one

multiplication result every clock cycle. This mode is useful for high-
speed data scaling.
■ Semi-Parallel Multiplication—Each memory produces one

multiplication with multi-cycle operation. This mode is useful for
coefficient update of least mean squares (LMSs) and coefficient
update of equalizers.

Preliminary
■ Sum of Multiplication—One memory or group of memories produces

the sum of multiplication results. This mode is useful in applications
such as finite impulse response (FIR) filtering and discrete cosine
transforms (DCTs).
■ Hybrid Multiplication—Combination and optimization of semi-

parallel and sum of multiplication modes of operation. This mode is
ideal for a complex number of multiplications in complex fast
Fourier transforms (FFTs) and infinite impulse response (IIR) filters.
■ Fully Variable Multiplication—This mode is useful for a soft multiplier

implementations in which both the input data and coefficients are
varying. This mode is ideal for low-resolution multiplication
functions.
The following sections describe each of these modes and provide

examples.
Parallel Multiplication
Parallel multiplication involves multiplying all sections of a single input
bus or multiplier value with a single multiplicand or coefficient and
summing the partial product of each multiplication to obtain the final
result. All of the input bits are parallel-loaded into the RAM block address
port registers and a new multiplication is completed each clock cycle. For
example, a 16-bit input bus can be separated into two groups of eight bits
(one group of eight LSB bits and another group of eight MSB bits) and
simultaneously shifted into the address ports of two RAM blocks. The
output of the RAM blocks indicate the multiplication result for the
particular set of bits with the coefficient. Figure 4 represents the
decomposition of a 16-bit data input, 10-bit constant coefficient parallel
multiplier.

Preliminary
Figure 4. Decomposition of a 16-Bit Input, 10-Bit Coefficient Parallel Multiplier

Input[15..8] Input[7..0]
Signed (MSB) Unsigned (LSB)
Input[15..0]
Coefficient[9..0]
Sign Extend
LSB Partial Product[18..0]
Shift 8 Bits MSB Partial Product[21..4]
Mult_Result[25..0]
Sum MSB & LSB

Partial Product Results
Figure 5 shows the RAM LUT implementation of the parallel multiplier

decomposition shown in Figure 4. Because a parallel multiplier accepts a
new input every clock cycle, this implementation takes three clock cycles
(one to load the input values into the RAM block address ports and two
pipeline delays) to compute the final multiplication result. New partial
products are obtained from the RAM blocks every clock cycle and the
partial products are summed according to their weights. Each partial
product multiplication generates an output of 18 bits. At the end of the
partial product accumulation, the multiplier generates a 26-bit output.

Preliminary
Figure 5. 16-Bit Input, 10-Bit Coefficient Parallel Multiplication Implementation Using M4K RAM Blocks as
LUTs Note (1)
ADDRESS MULT_RESULT
00000000 0
00000001 C
00000010 2*C
00000011 3*C
11111110 -2*C
11111111 -1*C
M4K RAM (1)
16 8 18
Block (LUT)
Input[15:0] 256 x 18
MSB (MSB)
<< 8
26
Output[25..0]
8 M4K RAM (1) 18

Block (LUT)
256 x 18
LSB
(LSB)
ADDRESS MULT_RESULT
00000000 0
00000001 C
00000010 2*C
00000011 3*C
11111110 254*C
11111111 255*C
Note to Figure 5:
(1) Optional pipeline register to increase system performance.
Figure 5 shows an implementation for a 16-bit data input, split into two
8-bit sections implemented using two M4K RAM blocks, one for the MSB
section and the other for the LSB section. For signed input buses, the M4K
RAM block that accepts the MSB bits must contain precalculated
coefficient values for signed inputs because the eight MSB bits that feed
this RAM block are treated as signed values. The M4K RAM block that
accepts the LSB bits must contain precalculated coefficient values for
unsigned inputs because the eight LSB bits that feed this RAM blocks are
unsigned values.

Preliminary
Because the size for M4K RAM blocks is 256 × 18 bits, the maximum
number of bits per section for each M4K RAM block for this coefficient
size is eight (28 = 256 addresses). The input bus and coefficient size
directly affects the number and configuration of RAM blocks used to
implement the multiplier. The parallel multiplication mode ensures
maximum data throughput (i.e., a new data value every clock cycle).
You can also implement the parallel fixed-coefficient multiplier using the
altmemmult Quartus II megafunction. You can use the MegaWizard®
Plug-In Manager to customize the altmemmult megafunction to specify
a parallel, fixed coefficient soft multiplier in your design. The input and
coefficient bit width settings as well as RAM block selection type
determine if the altmemmult function implements a semi-parallel or
parallel mode soft multiplier, whichever is more efficient. Figures 6 and 7
show the appropriate settings required to implement the both the MSB
and LSB M4K RAM blocks respectively, for the 16-bit input, 10-bit parallel
multiplier example shown in Figure 14. The coefficient implemented in
this example is a constant value of five.
Figure 6. altmemmult MegaWizard Settings for the MSB RAM Block 16-Bit
Input, 10-Bit Constant Coefficient Parallel Multiplier

Preliminary
Figure 7. altmemmult MegaWizard Settings for the LSB RAM Block for a 16-Bit
Input, 10-Bit Constant Coefficient Parallel Multiplier
The sload_data signal and the message located at the bottom right
hand corner of the MegaWizard window indicates whether the
altmemmult function chose to implement a semi-parallel or parallel
mode soft multiplier. A parallel soft multiplier does not have the
sload_data signal and the megafunction can accept a new input every
clock cycle. The altmemmult megafunction can only implement small
parallel mode soft multipliers (i.e., 8-bit input, 10-bit coefficient
multipliers). Larger parallel multipliers require multiple altmemmult
megafunctions to generate partial product results. To obtain the final
multiplication result, these partial products must be summed in an end-
stage adder implemented externally to the altmemmult function.
Fixed-Coefficient Multiplication
Figure 8 shows the simulation results for the example shown in Figure 5.
This example multiplies the input, which has a decimal value of 297, with
a coefficient, which has a value of 5.

Preliminary
Figure 8. Parallel Multiplication Simulation Results
Input Data Sent in on Partial Products Final Result

Clock Cycle 1 (Held Available on Clock Avail ale on Clock
for One Clock Cycle) Cycle 3 Cycle 4
Table 11 shows the implementation result for the parallel fixed coefficient
multiplication example shown in Figure 5. The example is implemented
using the altmemmult megafunction.
Table 11. 16-Bit Input, 10-Bit Constant Coefficient Parallel Multiplication

Implementation Results
Device EP1S10F484C5
Utilization Logic cells: 26/10,570 (1%)
M4K RAM blocks: 2/60 (3%)
Latency (1) 3 clock cycles
Throughput 291 megasamples per second
Performance 291.0 MHz
Note to Table 11
(1) Latency is the number of clock cycles required to complete a single multiplication
computation.
f You can download the files (parallel_fixed.zip) for the design described
in Table 11 from the Design Examples section of the Altera web site at
www.altera.com.
Variable Coefficient Multiplication

To perform constant coefficient multiplication, you can implement the
Stratix, Stratix GX, and Cyclone memory blocks as ROM. For variable
coefficient multiplication, these memory blocks must be implemented as
RAM blocks, which allow you to rewrite blocks with new precalculated
coefficients. Figure 9 shows an implementation for variable coefficient
parallel multiplication implementation using M4K single-port RAM
blocks. Using the method shown in Figure 9, the multiplier function is

Preliminary
stalled while the coefficients are updated. But, by implementing multiple

sets of RAM blocks for storing different precalculated coefficient sets, you
can switch multiplication between two different sets of coefficients in a
single clock cycle. One way of doing this is to partition the RAM block to
store two unique sets of coefficients and to use the MSB address bit to
select which coefficient set to use. Also, with the use of dual-port RAM
blocks, you can write/update the values of a set of coefficients in a
partition while simultaneously using a different set of coefficients in
another partition to perform multiplication.
Figure 9. 16-Bit Input, 10-Bit Variable Coefficient Parallel Multiplication Implementation Using M4K Single-
Port RAM Blocks as LUTs Note (1)
ADDRESS MULT_RESULT
00000000 0
00000001 C
00000010 2*C
00000011 3*C
16 8
Input[15:0]
MSB
Coefficient 8 8 M4K RAM Block (LUT) (1)
256 x 18 18 11111110 -2*C
Address [7:0] 11111111 -1*C
(MSB)
Coefficient
Write Enable
18 << 8
MSB Coefficient
Input [17:0]
26
Output[25..0]
LSB
8 M4K RAM Block (LUT) (1)
256 x 18 18
ADDRESS MULT_RESULT
(LSB)
00000000 0
18 00000001 C
LSB Coefficient 00000010 2*C
Input [17:0] 00000011 3*C
11111110 254*C
11111111 255*C
Note to Figure 9:
Table 12 shows the implementation results for a parallel variable

coefficient multiplication example.

Preliminary
Table 12. 16-Bit Input, 14-Bit Variable Coefficient Parallel Multiplication

Device EP1S10F484C5
Utilization Logic cells: 43/10,570 (<1%)
Note to Table 12:

computation.
f You can download the files (parallel_var.zip) for the design described in
Table 12 from the Design Examples section of the Altera web site at
www.altera.com.
Semi-Parallel Multiplication
Semi-parallel multiplication involves multiplying sections of a single
input bus or multiplier value with a single multiplicand or coefficient and
shift accumulating the partial product of each multiplication to obtain the
final result. For example, a 16-bit input bus can be separated into four
groups of four bits that are consecutively shifted into the address port of
the RAM block once every clock cycle, beginning with the first four LSB
bits. The output of the RAM block indicates the multiplication result for
a particular set of bits with the coefficient, every clock cycle. Figure 10
shows the decomposition of a 16-bit data input, 14-bit coefficient semi-
parallel multiplier.

Preliminary
Figure 10. Decomposition of a 16-Bit Input, 14-Bit Coefficient Semi-Parallel Multiplier

Input[15..12] Input[11..8] Input[7..4] Input[3..0]
Input[15..0]
Coefficient[13..0]
Sign Extend
Partial Product[17..0]
Sign Extend
Shift 4 Bits Partial Product[21..4]
Sign Extend
Mult_Result[29..0]
Accumulate Results
from Each Multiply
Figure 11 shows the RAM LUT implementation of the semi-parallel

multiplier decomposition shown in Figure 10. This implementation loads
the input data four bits every clock cycle, taking six clock cycles (four to
load the input values into the RAM block plus two pipeline delays) to
complete the multiplication operation by shift-accumulating the partial
products obtained from the RAM block once per clock cycle, according to
their weights. Each shift-accumulation of a partial product generates four
extra bits. At the end of the fourth partial product accumulation, the
multiplier generates a 30-bit output.

Preliminary
Figure 11. 16-Bit Input, 14-Bit Coefficient Semi-Parallel Multiplication Implementation Using M512 RAM
Blocks as LUTs Note (1)
30
>> 4
30
4 4 M512 RAM (1) 18 30

Input[15..0] Block (LUT) Output[29..0]
16 x 18
Semi-Parallel Multiplications Table

ADDRESS MULT_RESULT
0000 0
0001 C
0010 2*C
0011 3*C
1110 14*C
1111 15*C
Note to Figure 11:

Figure 11 shows an implementation for a 16-bit data input, split into four
4-bit sections implemented using a single M512 RAM block. In this
example, for the same memory block utilization, factors like the input bus
size help determine the output bit width and the latency of the multiplier.
Increasing the bit width of the sections (i.e., implementing more than
4-bit sections in this case) can reduce the latency of the multiplier. This
implementation may require more M512 RAM blocks or that you use
M4K RAM blocks.
You can also implement the semi-parallel fixed coefficient multiplier

using the altmemmult Quartus II megafunction. You can use the
MegaWizard Plug-In Manager to customize the altmemmult
megafunction to specify a semi-parallel, fixed coefficient soft multiplier in
your design. The input and coefficient bit width settings as well as RAM
block selection type determine whether the altmemmult function
implements a semi-parallel or parallel mode soft multiplier; it
implements whichever is more efficient. Figure 12 shows the settings
required to implement the 16-bit input, 14-bit semi-parallel multiplier
example shown in Figure 11. The coefficient implemented in this example
is a constant value of two.

Preliminary
Figure 12. altmemmult MegaWizard Settings for a 16-Bit Input, 14-Bit

Constant Coefficient Semi-Parallel Multiplier
The sload_data signal and the message located at the bottom right
hand corner of the MegaWizard window indicate whether the
altmemmult function chose to implement a semi-parallel or parallel
mode soft multiplier. A semi-parallel soft multiplier has an sload_data
signal and can only accept a new input after more than one clock cycle.
The semi-parallel multiplier in Figure 11 indicates that the 16-bit input is
split into four groups of four bits each. Because it takes four clock cycles
to load the entire 16-bits into the RAM block, the current input must
remain stable for four clock cycles prior to loading the new input. A high
signal on sload_data for one clock cycle indicates the start of a new
block of input data.
f For information on implementing variable coefficient soft multipliers,

refer to the “Variable Coefficient Multiplication” on page 15.
Figure 13 shows the simulation results for the example shown in

Figure 11. This example multiplies the input, which has a decimal value
of 10, with a coefficient, which has a value of 2.

Preliminary
Figure 13. Semi-Parallel Simulation Results

Start of Input Sequence First Partial Product
Indicated by Pulse of Available on Clock Cycle 4
Input Data Held for Final Result Available on
sload_data on Clock Cycle 1
Four Clock Cycles Clock Cycle 8
Table 13 shows the implementation result for the semi-parallel fixed

coefficient multiplication example shown in Figure 11.
Table 13. 16-Bit Input, 14-Bit Constant Coefficient Semi-Parallel

Multiplication Implementation Results
Device EP1S10F484C5
M512 RAM blocks: 1/94 (2%)
Note to Table 13:

computation.
f You can download the files (semi_prl_fixed.zip) for the design

described in Table 13 from the Design Examples section of the Altera
web site at www.altera.com.
Table 14 shows the implementation results for a semi-parallel variable


Preliminary
Table 14. 16-Bit Input, 14-Bit Variable Coefficient Semi-Parallel

Multiplication Implementation Results
Device EP1S10F484C5
M512 RAM blocks: 1/94 (2%)
Note to Table 14:

computation.
f You can download the files (semi_prl_var.zip) for the design described
www.altera.com.
Sum of Multiplication
The sum of multiplication mode result is the weighted summation of
results produced by multiplying a set of input data (multiplier) to a set of
multiplicands. This sum forms the basis of a MAC function that is useful
in functions such as FIR filters, where each input data (multiplier) value
is multiplied with a particular coefficient (or multiplicand) and summed
to provide the final result.
In the sum of multiplication mode, each input bus shifts into the address
port of the memory block one bit per clock cycle, starting with the LSB. If
there are four inputs (called A, B, C, and D) to the multiplier block, at the
first clock cycle, the LSB of inputs A, B, C, and D forms the 4-bit address
value to the RAM block. The next clock cycle, the second LSB bit for each
input forms the next address value to the RAM block, and so on. For an
n-bit input data width, it takes n clock cycles to load in all of the data bits
required to compute the multiplication result. The RAM block output
indicates the multiplication result for a specific bit position at each clock
cycle.
Figure 14 shows the RAM LUT implementation of four 4-bit data inputs
and up to 16-bit constant coefficients. This fixed coefficient
implementation takes six clock cycles (four to load the input values into
the RAM block plus two pipeline delays) to complete the multiplication
operation by shift-accumulating the partial products obtained from the
RAM block once per clock cycle, according to their weights. Each shift-
accumulation of a partial product generates an extra carry bit. At the end

Preliminary
of the fourth partial product accumulation, the multiplier generates a

22-bit output. The size of the input data helps determine the output bit
width and the latency of the multiplier.
Figure 14. 4-Input Sum of Multiplication Implementation Using M512 RAM Blocks as LUTs
22
>> 1
A 22
M512 RAM (1) 18

B 22
Block (LUT)
Output[21..0]
C 16 x 18
D Sum of Multiplications Table

ADDRESS MULT_RESULT Equivalent Circuit:
0000 0
A B C D
0001 c0
0010 c1
c0 c1 c2 c3
0011 c0 + c 1
1110 c1 + c 2 + c 3
1111 c0 + c 1 + c 2 + c 3 Output
Note to Figure 14:

Figure 14 shows an implementation for four 4-bit data inputs. Because

M512 RAM blocks are 32 × 18 bits, the maximum number of inputs for
each M512 RAM block for this coefficient size is five (25 = 32 addresses).
Depending on the number of inputs, size and number of coefficients, and
the required operating speed, the number of RAM blocks used varies. The
example shown in Figure 14 requires only one M512 RAM block.

refer to “Variable Coefficient Multiplication” on page 15.
Figure 15 shows the simulation result for an example based on Figure 14.
This example has additional pipeline stages and multiplies input A,
which has a binary value of 0001, with the c0 coefficient, which has a
value of -3.
1 You can choose to reduce the number of pipeline stages to

reduce the latency, but your design may have reduced fMAX as a
result.

Preliminary
Figure 15. Sum of Multiplication Simulation Results
Input D LSB Bits Sent First Partial Product Final Result

Input C on Clock Cycle 1 Available on Clock Available on
Input B Cycle 3 Clock Cycle 8
Input A
Table 15 shows the implementation results of the four input, 16-bit fixed
coefficient sum of multiplication example shown in Figure 14.
Table 15. 4-Input, 16-Bit Fixed Coefficient Sum of Multiplication

Device EP1S10F484C5
M512 RAM blocks: 1/94 (2%)
Note to Table 15:

(1) Latency is the number of clock cycles required to complete an entire sum of
multiplication computation.
f You can download the files (sum_mult_fixed.zip) for the design

Table 16 shows the implementation results of a four input, 16-bit variable

coefficient sum of multiplication example.

Preliminary
Table 16. Four Input, 16-Bit Variable Coefficient Sum of Multiplication

Device EP1S10F484C5
M512 RAM blocks: 1/94 (1%)
Note to Table 16:

(1) Latency is the number of clock cycles required to complete an entire sum of
multiplication computation.
f You can download the files (sum_mult_var.zip) for the design described
www.altera.com.
You can combine multiple M512 blocks and/or M4K blocks to create
larger multiplier structures that are capable of multiplying more data
inputs and coefficients simultaneously. Figure 16 shows the
multiplication of eight 4-bit data inputs to eight 16-bit constant
coefficients in two M512 RAM blocks.
Figure 16. Using Multiple M512 RAM Blocks for an 8-Coefficient Multiplier
M512 RAM (1) 18

B
Block (LUT) 23
16 x 18 >> 1
C
D 23
1 19 23
Output[22..0]
M512 RAM (1) 18

F
Block (LUT)
16 x 18
G
Note to Figure 16

Preliminary

You can also create similar implementations using M4K RAM blocks,
particularly if the coefficients are larger than 16 bits. Figure 17 shows
multiplication of seven 16-bit data inputs to a 20-bit constant coefficient
in one M4K RAM block. The 128 addressed lines correspond to seven data
inputs or unique coefficients in a M4K RAM block. Performing seven 16
× 20-bit multiplications generates a 23-bit output from a M4K RAM block.
It takes 18 clock cycles to complete accumulation of the partial products
(16 clock cycles to shift the input values into the address port of the RAM
block plus two pipeline delays). After each partial product accumulation,
one bit is added to the total number of output bits, making the final
output 39 bits wide.
Figure 17. Using a M4K RAM Block for a 7-Coefficient Multiplier
39
>> 1
A 39
B
C M4K RAM (1) 23 39
D Block (LUT) Output[38..0]
E 128 x 23
F
G
Note to Figure 17:


Hybrid Multiplication
The hybrid multiplication mode is a combination of the semi-parallel and
sum of multiplication modes where bit sections from two unique input
streams are multiplied with two different coefficients values. This mode
is useful in applications that require complex multiplication like fast
Fourier transforms (FFTs) where each signal generally has a real and
imaginary component that could be multiplied by two unique coefficient
values. The partial products obtained from each bit section within the
components are shift accumulated to obtain the final result.
In the hybrid multiplication mode, an equal number of bits from each

input is concatenated and shifted into the address port of the RAM block
every clock cycle, starting with the LSB. If the address port to the RAM

Preliminary
block is four bits wide, each input contributes two bits to the partial
product calculation every clock cycle until the entire bit width of the
inputs have completely shifted into the RAM block. In this case, for an
input bus of 16-bits, it takes 8 clock cycles to shift in all of the data bits of
that particular input. The output of the RAM block indicates the sum of
multiplication result for a particular set of bits with the coefficients, every
clock cycle.
Figure 18 shows the RAM LUT implementation of two 16-bit inputs, each
labeled I Input and Q Input, respectively, and up to 15-bit constant
coefficients. This implementation takes 11 clock cycles (eight to load the
input values into the RAM block plus three pipeline delays) to complete
the multiplication operation by shift-accumulating the partial products
obtained from the RAM once per clock cycle, according to their weights.
Each shift-accumulation of a partial product generates two extra bits. At
the end of the last (eighth) partial product accumulation, the multiplier
generates a 32-bit output. The size of the input data helps determine the
output bit width and the latency of the multiplier.
Figure 18. Two-Input Hybrid Multiplication Implementation Using M512 RAM Blocks as LUTs
32
>> 2
MSB LSB
2 2 32
Input Q [15..0]
M512 RAM (1) 18 32
Block (LUT)
Output[31..0]
32 x 18
2 2
Input I [15..0]
Hybrid Multiplications Table
ADDRESS MULT_RESULT
0000 0
0001 Ci
0010 2*Ci
0011 3*Ci
1110 3*Cq + 2*Ci Ci - I Coefficient

1111 3*Cq + 3*Ci Cq - Q Coefficient
Note to Figure 18:

Figure 18 shows an implementation for two 16-bit data inputs. Even

though the 32 × 18-bit configured M512 RAM block can accept five
address bits (25 = 32 addresses), the maximum number of bits equally
contributed by each input is two bits (totaling four bits). In this example,
for the same memory block utilization, factors such as the input bus size
help determine the output bit width and the latency of the multiplier.

Preliminary
Increasing the number of M512 RAM blocks used or moving to larger

memory blocks like M4K RAM blocks can reduce the latency of the
multiplier an support larger coefficient bit widths.

Figure 19 shows the simulation results for an example based on Figure 18.
This example has additional pipeline stages and multiplies the I and Q
inputs, which have values of 300 and 55, respectively, with coefficients Ci
and Cq, which have values of 10 and 25, respectively (result = (input_I ×
Ci) + (input_Q × Cq) = (300 × 10) + (55 × 25) = 4375).
1 You can choose to reduce the number of pipeline stages to

reduce the latency, but your design may have reduced fMAX as a
result.
Figure 19. Hybrid Multiplication Simulation Results
Start of Input Data Sequence Input Data Held First Partial Product Final Result
Indicated by Pulse of sload_data for 8 Clock Cycles Available on Clock Available on Clock
on Clock Cycle 1 Cycle 5 Cycle 13
Table 17 shows the implementation results of the two 16-bit input, 15-bit
constant coefficient hybrid multiplication example shown in Figure 18.

Preliminary
Table 17. Two Input, 15-Bit Constant Coefficient Hybrid Multiplication

Device EP1S10F484C5
M512 RAM blocks: 1/94 (1%)
Note to Table 17:

computation.
f You can download the files (hybrid_fixed.zip) for the design described
www.altera.com.
Table 18 shows the implementation results for a hybrid variable

Table 18. Two Input, 15-Bit Variable Coefficient Hybrid Multiplication

Device EP1S10F484C5
M512 RAM blocks: 1/94 (1%)
Note to Table 18:

computation.
f You can download the files (hybrid_var.zip) for the design described in
www.altera.com.

Preliminary
Fully Variable Multipliers

The fully variable multiplier mode allows you to implement a soft
multiplier in which both the input and the coefficient can vary every clock
cycle. The partial product values, which are stored in the RAM blocks, are
calculated based on the algebraic expansion of the following equation:
(a + b)2 - (a - b)2 = a2 + 2ab + b2 - (a2 - 2ab + b2)
= 4ab
therefore:
ab = ((a + b)2 / 4) - ((a - b)2 / 4)
Where a and b are both variable inputs to the multiplier
Figure 20 shows the RAM LUT implementation of the fully variable

multiplier calculated using these equations. Two unique RAM blocks are
required, to store the (a + b)2/4 and (a - b)2/4 precalculated values,
respectively. The address inputs of (a + b) for the former and (a - b) for the
latter RAM block are precalculated in logic prior to the RAM block. The
final result of the multiplication is obtained by subtracting the result of
the (a - b) RAM block by the result from the (a + b) RAM block. The fully
variable multiplier can accept a new input every clock cycle, and takes
three clock cycles to compute the final multiplication result.

Preliminary
Figure 20. 8-Bit Fully Variable Multiplier Implementation Using M4K RAM Blocks as LUTs
2
((a + b) )/4
8
Input A [7..0]
(1) M4K RAM (1)

9 (a + b)[8..0] Block (LUT) x 2 16
256 x 16 x 2
(512 x 16)
Output[15..0]
2
((a - b) )/4
(1) M4K RAM (1)

8 9 (a - b)[8..0] 16
Block (LUT) x 2
Input B [7..0]
256 x 16 x 2
(512 x 16)
Note to Figure 20:

Figure 20 shows an implementation for two 8-bit data inputs. 8-bit inputs
result in 16-bit outputs and 9-bit addresses per partial product RAM
block. Therefore, for each partial product, two M4K RAM blocks are
required in a 256 × 16 configuration (29 = 512 addresses). In this multiplier
mode, the size of the inputs directly affects the total number of RAM
blocks required.

Figure 20.

Preliminary
Figure 21. Fully Variable Multiplier Simulation Results
Input Sent in on Partial Product Final Result Available

Clock Cycle 1 (Held Available on Clock on Clock Cycle 5
for 1 Clock Cycle) Cycle 4
Table 19 shows the implementation results of the 8-bit fully variable

multiplier example shown in Figure 20. The fully variable multiplication
mode is ideal for low-resolution multiplication in which the input and
coefficient bit widths are not too large. Larger input and coefficient bit
widths require a significant amount of memory block resources
compared to other variable soft multiplier modes of the same size.
Table 19. 8-Bit Fully Variable Multiplier Implementation Results

Device EP1S10F484C5
Note to Table 19:

computation.
f You can download the files (fully_var.zip) for the design described in
www.altera.com.

Preliminary
Implementing Multipliers Using DSP Blocks or LEs
Implementing Altera provides three Quartus II megafunctions for implementing

various multiply, multiply-accumulate, and multiply-add functions
Multipliers using DSP blocks or LEs:
Using DSP ■ lpm_mult—Performs
multiply functions only
Blocks or LEs ■ altmult_add—Performs multiply or multiply-add functions
■ altmult_accum—Performs multiply-accumulate functions only
f For more information on using these megafunctions to implement

multipliers, refer to AN 214: Using the DSP Blocks in Stratix & Stratix GX
Devices.
Firm Multipliers Firm multipliers use a combination of DSP blocks and LEs, enabling you
to increase the utilization efficiency of the DSP blocks within your Stratix
or Stratix GX device. Stratix and Stratix GX DSP blocks support 9 × 9,
18 × 18, and 36 × 36 multipliers. If you implement a multiplier of a
different size, some DSP blocks may be partially used. For example, a
12 × 9 multiplier uses two 9 × 9 DSP blocks because the 12-bit input
exceeds the maximum requirement of a single 9 × 9 multiplier. The first
9 × 9 DSP block is fully utilized but the second 9 × 9 multiplier is partially
used. Instead of using the partially utilized DSP block for the remaining
logic, you can use a firm multiplier to implement it, freeing the DSP block
for other use. This method is particularly useful if your design requires a
lot of DSP blocks but has LE resources available.
To implement a firm 12 × 9 multiplier, split up the 12-bit input and

decompose the multiplication into smaller, partial products that can be
implemented in DSP blocks and LEs. To maximize DSP block usage, split
the 12-bit input into two sections: a 9-bit section that is multiplied using
the DSP blocks and a 3-bit section that is multiplied using LEs. If the 9-bit
section consists of LSBs, it becomes an unsigned value while the 3-bit
section becomes a signed value and vice versa.
When deciding whether to select the 3-bit section from the MSB or the
LSB of the 12-bit input, keep in mind that an LE multiplier is more
resource efficient when implemented as a signed multiplier than as an
unsigned multiplier. If the 9-bit input is unsigned, the 3-bit section is
chosen from the MSB so that the LE multiplier performs signed
multiplication. If the 9-bit input is signed, you can choose the 3-bit section
from the MSB or LSB because either implementation results in a signed
multiplier implemented in LEs.
Figure 22 shows the decomposition of the 12 × 9 firm multiplier.

Preliminary
Figure 22. Decomposition of the 12 × 9 Multiplier

Input A [11..9] Input A [8..0]
(signed) unsigned
Input A [11..0]
Input B [8..0]
Sign Extend
Mult_Result[20..0]
Accumulate Results
from Each Multiply
Based on this decomposition, you can build the circuit for the firm
multiplier using three main blocks:
■ DSP block multiplier—Built using either the lpm_mult or

altmult_add megafunctions
■ LE-based multiplier—Built using either the lpm_mult or
■ End-stage adder—Built using the lpm_add_sub megafunction
The DSP block multiplier multiplies the 9-bit input by the 9-bit LSB
section of the 12-bit input. The LE-based multiplier multiplies the 9-bit
input with the 3-bit MSB section of the 12-bit input. The result of both
multipliers is the partial products of the decomposition. The results of the
partial products are weighted prior to being summed in the end-stage
adder. This weighting and addition restores the bit-alignment of the
partial products to ensure proper result values. Based on Figure 22, the
9 × 3 multiplication partial product is weighted by a shift to the left of
nine bits. The 12-bit end-stage adder has to accommodate the 12-bit result
of the 9 × 3 multiplication and the nine MSBs of the 9 × 9 multiplication,
sign extended.
Figure 23 shows the circuit of the 12 × 9 firm multiplier.

Preliminary
Firm Multipliers
Figure 23. 12 × 9 Firm Multiplier Circuit

3
12 Input A [11..9]
Input A [11..0] (1) 12 LE Mult [11..0]
<< 9
9 9
Input B [8..0]
Input B [8..0]
LE Multiplier
12
Output [20..0]
9
Input A [8..0]
Unsigned (1) 18 DSP Mult [17..9] 9 9
9
Input B [8..0] DSP Mult [8..0]
DSP Block Multiplier
Notes to Figure 23:

(2) Using the altmult_add megafunction to implement the multipliers allows you to mix signed and unsigned
inputs.

Figure 23.
Figure 24. 12 × 9 Firm Multiplier Simulation Results

Input Sent in on Final Result Available
Clock Cycle 1 (Held on Clock Cycle 3
for 1 Clock Cycle)
Table 20 shows the implementation results for the 12 × 9 firm multiplier

circuit example shown in Figure 23.

Preliminary
Table 20. 12 × 9 Firm Multiplier Implementation Results Note (2)

Device EP1S10F484C5
DSP block 9-bit elements: 1/48 (2%)
Note to Table 20:

computation
(2) The altmult_add megafunction implements both the LE and DSP block
multipliers.
f You can download the files (12x9_firm_mult.zip) for the design

The example shown in Figure 23 is suitable when only one of the

multiplier inputs exceeds the 9-bit input width of a single DSP block.
When both multiplier inputs exceed 9-bits, as in the case of a 12 × 12
multiplier, the multiplication must be decomposed into three partial
products instead of two. The 12-bit inputs must be sectioned to maximize
the use of the 9 × 9 DSP blocks and the utilization efficiency of
implementing signed multiplication in LEs. Therefore, both inputs
should be sectioned into a 3-bit MSB section and a 9-bit LSB section.
Figure 25 shows the decomposition of the 12 × 12 multiplier.

Preliminary
Firm Multipliers
Figure 25. Decomposition of the 12 × 12 Multiplier

Input A [11..9] Input A [8..0]
(signed) unsigned
Input A [11..0]
Input B [11..0]
Input B [11..9] Input B [8..0]

(signed) unsigned
Sign Extend
Mult_Result[23..0]
Accumulate Results
from Each Multiply
The circuit for the firm multiplier can now be extracted from the
decomposition. The firm multiplier circuit consists of five main blocks:
■ One DSP block multiplier—Built using either the lpm_mult or

■ Two LE-based multipliers—Built using either the lpm_mult or
■ Two adders—Built using the lpm_add_sub megafunction
The DSP block multiplier multiplies the two 9-bit LSB sections of the
12-bit inputs. The first LE-based multiplier multiplies the 9-bit LSB
section of one 12-bit input with the 3-bit MSB section of the other 12-bit
input. The other LE-based multiplier multiplies the 3-bit MSB of one 12-
bit input with the entire 12-bits of the other input. The results of these
three multipliers are the three partial products of the decomposition. The
results of these partial products are summed in two stages (using two
adders) prior to producing the final output.
Figure 26 shows the two adder stages within the final circuit of the 12 × 12
firm multiplier.

Preliminary
Figure 26. 12 × 12 Firm Multiplier Circuit
12 12
Input A [11..0]
Input A [11..0]
(1) 15 LE Mult2 [14..0] 15
P3 [23..9]
<< 9
3
12 Input B [11..9]
Input B [11..0]
LE Multiplier
15
P4 [23..9]
3 Output [23..0]
Input A [11..9]
(1) 12 P1[20..9]
12 LE Mult1 [11..0]
<< 9
9
Input B [8..0]
LE Multiplier 12 P2 [20..9]
9
Input A [8..0]
Unsigned (1) 9 P0 [17..9]
18
9 P0 [8..0]
9
Input B [8..0]
Unsigned DSP Block Multiplier
Notes to Figure 26:

(2) Using the altmult_add megafunction to implement the multipliers allows you to mix signed and unsigned
inputs.

Figure 26.
Figure 27. 12 × 12 Firm Multiplier Simulation Results

Input Sent in on Final Result Available
Clock Cycle 1 (Held on Clock Cycle 3
for 1 Clock Cycle)
Table 21 shows the implementation results for the 12 × 12 firm multiplier

example shown in Figure 26.

Preliminary
Conclusion
Table 21. 12 × 12Firm Multiplier Implementation Results Note (2)

Device EP1S10F484C5
DSP block 9-bit elements: 1/48 (2%)
Note to Table 21:

computation
(2) The altmult_add megafunction implements both the LE and DSP block
multipliers.
f You can download the files (12x12_firm_mult.zip) for the design

Conclusion Although Stratix and Stratix GX DSP blocks are useful for implementing
DSP applications, you can also use Stratix and Stratix GX TriMatrix blocks
(M512 or M4K RAM blocks) or Cyclone M4K RAM blocks for designs that
need more multipliers than are available using DSP blocks alone. For
example, using soft multipliers, you can increase the number of 16 × 16
multipliers in a Stratix E1S80 device by a factor of more than 7 see Table 9
on page 5). Another example, the fully variable soft multiplier is an ideal
implementation for applications requiring smaller multipliers with
frequently varying coefficients. Other soft multiplier modes are more
resource efficient and better suited for applications that do not require
frequent coefficient updates. The firm multiplier allows you to balance
the use of DSP block multipliers with LE-based multipliers, allowing
more efficient use of the Stratix and Stratix GX DSP blocks.

Preliminary
Copyright © 2003 Altera Corporation. All rights reserved. Altera, The Programmable Solutions Company,
the stylized Altera logo, specific device designations, and all other words and logos that are identified as
trademarks and/or service marks are, unless noted otherwise, the trademarks and service marks of Altera
Corporation in the U.S. and other countries. All other product or service names are the property of their re-
spective holders. Altera products are protected under numerous U.S. and foreign patents and pending
101 Innovation Drive applications, maskwork rights, and copyrights. Altera warrants performance of its semiconductor products
San Jose, CA 95134 to current specifications in accordance with Altera's standard warranty, but reserves the right to make chang-
(408) 544-7000 es to any products and services at any time without notice. Altera assumes no responsibility or liability
arising out of the application or use of any information, product, or service described
www.altera.com herein except as expressly agreed to in writing by Altera Corporation. Altera customers
Applications Hotline: are advised to obtain the latest version of device specifications before relying on any pub-
(800) 800-EPLD lished information and before placing orders for products or services.
Literature Services:
Printed on recycled paper
[email protected]

Preliminary

Techniques For Implementing Multipliers in Stratix, Stratix GX & Cyclone Devices

Uploaded by

Copyright:

Available Formats

Techniques For Implementing Multipliers in Stratix, Stratix GX & Cyclone Devices

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Techniques For Implementing Multipliers in Stratix, Stratix GX & Cyclone Devices

Uploaded by

Copyright:

Available Formats

Techniques for Implementing

Table 1. Supported Multiplier Implementations

Multiplier Type Description Stratix Stratix GX Cyclone

Altera Corporation Quartus II Version 3.0 1

Tables 2 and 3 show the total number of multipliers available in Stratix

Table 2. Number of Multipliers in Stratix Devices

DSP Blocks Soft Multipliers (16 × 16) Total Multipliers

Table 3. Number of Multipliers in Stratix GX Devices

DSP Blocks Soft Multipliers (16 × 16) Total Multipliers

2 Quartus II Version 3.0 Altera Corporation

Table 4. Number of Multipliers in Cyclone Devices

Device Soft Multipliers (16 × 16) Notes (1), (2)

Table 5. Stratix Memory Blocks

M512 RAM M4K RAM M-RAM

Altera Corporation Quartus II Version 3.0 3

Table 6. Stratix GX TriMatrix Memory Blocks

M512 RAM M4K RAM M-RAM

Table 7. Cyclone M4K Memory Blocks

Device M4K RAM (128 × 36 Bits)

Table 8. M512, M4K & M-RAM Memory Configurations (Part 1 of 2)

M512 RAM Block M4K RAM Block M-RAM Block

4 Quartus II Version 3.0 Altera Corporation

Table 8. M512, M4K & M-RAM Memory Configurations (Part 2 of 2)

M512 RAM Block M4K RAM Block M-RAM Block

Table 9. Number of DSP Blocks in Stratix Devices Note (1)

Device DSP Blocks 9 × 9 Multipliers 18 × 18 Multipliers 36 × 36 Multipliers

Device DSP Blocks 9 × 9 Multipliers 18 × 18 Multipliers 36 × 36 Multipliers

Altera Corporation Quartus II Version 3.0 5

Device DSP Blocks 9 × 9 Multipliers 18 × 18 Multipliers 36 × 36 Multipliers

Note to Table 10:

DSP Arithmetic DSP is a multiplication-intensive technology and to achieve high speeds,

6 Quartus II Version 3.0 Altera Corporation

Figure 1. Multiplication of Two 2-Bit Numbers

Half Adder Half Adder

carry_out sum carry_out sum

Altera Corporation Quartus II Version 3.0 7

Figure 2. Distributed Arithmetic with Four Constant Multiplicands

The distributed-arithmetic circuit simultaneously performs four

Distributed Arithmetic in LUTs

8 Quartus II Version 3.0 Altera Corporation

Figure 3. Four-Bit Multiplication with Constant Coefficients Note (1)

The five soft multiplier modes supported by Stratix and Stratix GX

■ Parallel Multiplication—Multiple memories produce one

■ Semi-Parallel Multiplication—Each memory produces one

Altera Corporation Quartus II Version 3.0 9

■ Sum of Multiplication—One memory or group of memories produces

■ Hybrid Multiplication—Combination and optimization of semi-

■ Fully Variable Multiplication—This mode is useful for a soft multiplier

The following sections describe each of these modes and provide

10 Quartus II Version 3.0 Altera Corporation

Figure 4. Decomposition of a 16-Bit Input, 10-Bit Coefficient Parallel Multiplier

Shift 8 Bits MSB Partial Product[21..4]

Sum MSB & LSB

Figure 5 shows the RAM LUT implementation of the parallel multiplier

Altera Corporation Quartus II Version 3.0 11

8 M4K RAM (1) 18

12 Quartus II Version 3.0 Altera Corporation

Altera Corporation Quartus II Version 3.0 13

14 Quartus II Version 3.0 Altera Corporation

1110 3Cq + 2Ci Ci - I Coefficient