Inbound Read Prefetching; Latency; Gxb Address Map; Bandwidth Estimates For Various Request Sizes - Intel 460GX Software Developer’s Manual

AGP Subsystem
Table 7-5. Bandwidth Estimates for Various Request Sizes
8
16
32
64
128
256
7.3.1

Inbound Read Prefetching

The PCI protocol has no transfer size explicitly spelled out. Reads begin and continue until the
device has the data it needs. For performance, a PCI bridge could prefetch data ahead of when the
PCI device requests it. AGP, as opposed to PCI, has a length field in its protocol. Therefore the
graphics bridge does know how much data to fetch. Also the GXB has 1 GB/s peak into it, which is
the AGP peak, so there is no extra bandwidth in case the bridge prefetched the wrong data. For this
reason, the 460GX chipset will have no prefetching of graphics data. The graphics card should
have enough buffering and outstanding transactions to keep its pipes full. With pipelining, the card
can fetch the data it needs in advance of using it. Knowing what it wants to do, the graphics card
can prefetch only the data it expects to use. This allows the best utilization of the graphics port.
7.4

Latency

Latency of reads from the graphics card is an important parameter. Unfortunately it is difficult to
specify usefully. For an idle system, the latency of a graphics read will be 450-500 ns. For a
moderately loaded system, the latency of a read should be 600-900 ns. Defining 'moderately
loaded' is the difficult part. If the GXB queues are backed up servicing many small AGP requests,
then the latency may be much higher.
7.5

GXB Address Map

The 'System Address Map' chapter contains a section describing the what the system address map
looks like from the perspective of an expander bridge. This section is included to explicitly state
which address ranges must be checked at the AGP interface when using PCI protocol and which
ranges must be checked after GART translation.
There are two address regions which must be checked at the AGP bus interface when the card is
using PCI protocol: the GXB's nx32M logical PCI space and the VGA space. This is required since
the AGP card may attempt to "talk to itself" using the PCI protocol in these regions, so the GXB
must not assert DEVSEL# to these accesses. The GXB's nx32M region is specified using its
MMBASE and MMTOP registers. The VGA gap is enabled as a memory hole using the VGASE
register in the GXB. All other accesses using PCI protocol will receive DEVSEL#. This includes
accesses above TOM, in a MARG, etc. that may cause XBINIT# after the GART (see below).
7-14
Request Size
(in bytes)
100-300
250-500
450-700
750-850
850-925
800-850
Sustainable Bandwidth
(MB/s)
Intel® 460GX Chipset Software Developer's Manual
Latency (in ns.) for First Data
Returned on AGP Bus
500
500
530
600
750
1000