Texas Instruments SPRU938B manual Read Performance, Throughput, Latency μsec

www.ti.com

Read Performance

B.3 Read Performance

Since reads must complete a transmit-remote read-receive cycle before starting another read transaction, the data throughput is lower as compared to writes. There is latency involved in reading the data from the remote device; and in some cases, a local latency in writing the returned data before the next read can start.

The max read rate is calculated the same way as the max write rate. The packet overhead is as shown below:

Read32 - caaaaT

ReadBurst - claaaaT

ReadReturn - clddddddddT

Where

T - EndOfPacket

d - data, dddd represents additional 32-bit words in burst, up to 16 words.

a - address

c - command

l - length

There are 6 bytes of overhead for a single read, 7 bytes for burst reads, and 3 bytes for read returns. The time required for a read is the total of the time for the read request, remote latency, read return, and local latency. Thus, the throughput can be calculated as data bytes/total transaction time, where the latency of both local and remote devices is combined.

Read Throughput = data/ (((Read + ReadReturn + data)/max read rate) + Latency

=(data × max read rate)/((Read + ReadReturn + data) + Latency × max read rate)

For example, with a 4 pin, 99 MHZ VLYNQ connection, for a single 32-bit word read:

Read Throughput = 32 bits × 316.8 Mbps/ (6 × 8 + 3 × 8 + 4 × 8 + Latency × 316.8Mbps)

= 10137.6/(104 + Latency × 316.8 Mbps)

Similarly, for a burst read of sixteen 32-bit words, with a 4 pin, 99 MHZ VLYNQ connection

Read Throughput = 16 × 32 bits × 316.8Mbps/(6 × 8 + 3 × 8 + 16 × 4 × 8 + Latency × 316.8Mbps) = 162201.6/(584 + Latency × 316.8Mbps)

Using the formula above, the relative performance with various latencies is illustrated for a 4 pin, 99 MHZ VLYNQ clock, burst read (sixteen 32-bit words) throughput rate, as shown in Table B-3.

Table B-3. Relative Performance with Various Latencies

Number of VLYNQ	Burst Size in		Throughput
		Latency (μsec)
Pins (99 MHZ)	32-bit Words		Mbits/sec	Mbytes/sec
4	16	0	277.74	34.72
		1	179.70	22.46
		10	43.02	5.38
		100	5.00	0.62

To efficiently use VLYNQ bandwidth, it is desirable for each VLYNQ device to write from the local device to the remote device. Burst transactions are more efficient than single read/write transactions.

46 Write/Read Performance SPRU938B –September 2007

Submit Documentation Feedback