[Home] [This version is outdated, a new version is here]
*Title...
*Search:...
*Read Me/FAQ/General Info...
*_IBM...
*ACC Micro...
*ALD...
*ALi...
*AMD . . . . . . . [no datasheets, some info]...
*Chips & Technologies...
*Contaq . . . . . [no datasheets, some info]...
*Efar Microsystems [no datasheets, some info]...
*ETEQ...
*Faraday...
*Forex . . . . . . [List only, no datasheets found]...
*Intel...
**82395SX Smart Cache 12/17/90
***Notes:...
**82396SX Smart Cache 12/17/90
***Notes:...
***Info:
The 82396SX Smart Cache (part number 82396SX) is a low cost, single
chip, 16-bit peripheral for Intel's i386 SX Microprocessor. By
storing frequently accessed code or data from main memory the 82396SX
Smart Cache enables the i386 SX Microprocessor to run at near zero
wait states. The dual bus architecture allows another bus master to
access the System Bus while the i386 SX Microprocessor operates out of
the 82396SX Smart Cache on the Local Bus. The 82396SX Smart Cache has
a snooping mechanism which maintains cache coherency with main memory
during these cycles.
The 823968X Smart Cache is completely software transparent, protecting
the integrity of system software. The advanced architectural features
of the 82596SX Smart Cache offer high performance with a cache data
RAM size that can be integrated on a single chip, offering the board
space and cost savings needed in an i386 SX Microprocessor based
system.
1.0 823968X SMART CACHE FUNCTIONAL OVERVIEW
1.1 Introduction
The primary function of a cache is to provide local storage for freq-
uently accessed memory locations. The cache intercepts memory
references and handles them directly without transferring the request
to the System Bus. This results in lower traffic on the System Bus and
decreases latency on the Local Bus. This leads to improved performance
for a processor on the Local Bus. It also increases potential system
performance by reducing each processor's demand for System Bus band-
width, thus allowing more processors or system masters in the system.
By providing fast access to frequently used code and data the cache is
able to reduce the average memory access time of the i386 SX
Microprocessor based system.
The 82396SX Smart Cache is a single chip cache subsystem specifically
designed for use with the i386 SX Microprocessor. The 82396SX Smart
Cache integrates 16KB cache, the Cache Directory and the cache control
logic onto one chip. The cache is unified for code and data and is
transparent to application software. The 82396SX Smart Cache provides
a cache consistency mechanism which guarantees that the cache has the
most recently updated version of the main memory. Consistency sup-
port has no performance impact on the i386 SX Microprocessor. Section
1.2 covers all the 82396SX Smart Cache features.
The 82396SX Smart Cache architecture is similar to the i486 SX
Microprocessor's on-chip cache. The cache is four Way SET associative
with Pseudo LRU (Least Recently Used) replacement algorithm. The line
size is 16B and a full line is retrieved from the memory for every
cache miss. A TAG is associated with every 16B line. The 82396SX Smart
Cache architecture allows for cache read hit cycles to run on the
Local Bus even when the System Bus is not available. 82396SX Smart
Cache incorporates a new write buffer cache architecture, which allows
the i386 SX Microprocessor to continue operation without waiting for
write cycles to actually update the main memory.
A detailed description of the cache operation and parameters is
included in Chapter 2.
The 82396SX Smart Cache has an interface to two electrically isolated
busses. The interface to the i386 SX Microprocessor bus is referred to
as the Local Bus (LB) interface. The interface to the main memory and
other system devices is referred to as the 82396SX Smart Cache System
Bus (SB) interface. The SB interface emulates the i386 SX
Microprocessor. The SB interface, as does the i386TM SX Micro-
processor. operates in pipeline mode.
In addition, it is enhanced by an optional burst mode for Line Fills.
The burst mode provides faster line fills by allowing consecutive read
cycles to be executed at a rate of up to one word per clock
cycle. Several bus masters (or several 82396SX Smart Caches) can share
the same System Bus and the arbitration is done via the SHOLD/SHLDA
mechanism (similar to the i486 SX Microprocessor).
Cache consistency is maintained by the SAHOLD/SEADS# snooping
mechanism, similar to the i486 SX Microprocessor. The 82396SX Smart
Cache is able to run, a zero wait state i386 SX Microprocessor
non-pipelined read cycle if the data exists in the cache. Memory write
cycles can run with zero wait states if the write buffer is not full.
The 82396SX Smart Cache organization provides a higher hit rate than
other standard configurations. The 82396SX Smart Cache, featuring the
new high performance write buffer cache architecture, provides full
concurrency between the electrically isolated Local Bus and System
Bus. This allows the 82396SX Smart Cache to service read hit cycles on
the Local Bus while running line fills or buffered write cycles on the
System Bus.
1.2 Features
1.2.1 823858X-LIKE FEATURES
o The 82396SX Smart Cache maps the entire physical address range of
the i386 SX Microprocessor (16MB) into an 16KB cache. Unified code
and data cache.
o Cache attributes are handled by hardware. Thus the 82396SX Smart
Cache is transparent to application software. This preserves the
integrity of system software and protects the users software
investment.
o Word and Byte writes, Word reads.
o Zero wait states in read hits and in buffered write cycles. All i386
SX Microprocessor cycles are non-pipelined (Note: The i386 SX
Microprocessor must never be pipelined when used with the 82396SX
Smart Cache - NA# must be tied to Vcc).
o A hardware cache FLUSH# option. The 82396SX Smart Cache will
invalidate all the Tag Valid bits in the Cache Directory and clear
the System Bus line buffer when FLUSH# is activated tor a minimum of
four CLK’s.
o The 82396SX Smart Cache supports non-cacheable accesses.
o The 82396SX Smart Cache internally decodes the i387 SX Math
Coprocessor accesses as Local Bus cycles.
o The System Bus interface emulates a i386 SX Microprocessor
interface.
o The 82396SX Smart Cache supports pipelined and non-pipelined system
interface.
o Provides cache consistency (snooping): The 82396SX Smart Cache
monitors the System Bus address via SEADS# and invalidates the cache
address if the System Bus address matches a cached location.
1.2.2 NEW FEATURES
o 16KB on chip cache arranged in four banks, one bank for each way. In
Read hit cycles, one word is read. In a write hit cycle, any byte
within the word can be written. In a cache fill cycle, the whole
line (16B) is written. This large line size increases the hit rate
over smaller line size caches.
o Cache architecture similar to the i486 SX Microprocessor cache: 4
Way set associative with Pseudo LRU replacement algorithm. Line
size is 16B and a full line is retrieved from memory for every cache
miss. A Tag Valid Bit and a Write Protect Bit are associated with
every Line.
o New write buffer architecture with four word deep write buffer
provides zero wait state memory write cycles. I/O, Halt/ Shutdown
and LOCK#ed writes are not buffered.
o Concurrent Line Buffer Cacheing: The 82396SX Smart Cache has a line
buffer that is used as additional memory. Before data gets written
to the cache memory at the completion of a Line Fill it is stored in
this buffer. Cache hit cycles to the line buffer can occur before
the line is written to the cache.
o In i387 SX Math Coprocessor accesses, the 82396SX Smart Cache drives
the READYO# in one wait state if the READYI# was not driven in the
previous clock.
Note that the timing of the 82396SX Smart Cache’s READYO# generation
for i387 SX Math Coprocessor cycles is incompatible with 80287
timing.
o An enhanced System Bus interface:
a) Burst Option is supported in line-fills similar to the i486 SX
Microprocessor. SBRDY# (System Burst READY) is provided in
addition to SRDY#. A burst is always a 16 byte line fill (cache
update) which is equivalent to eight word cycles.
b) System cacheability attribute is provided (SKEN#). SKEN# is used
to determine whether the current cycle is cacheable. It is used
to qualify Line Fill requests.
c) SHOLD/SHLDA system bus arbitration mechanism is supported. A
Multi i386 SX 82396SX Smart Cache cluster can share the same
System Bus via this mechanism.
f) Cache invalidation cycles supported via SEAD$#. This is used to
provide cache coherency.
o Full Local Bus/System Bus concurrency is attained by:
a) Servicing cache read hit cycles on the Local Bus while completing
a Line Fill on the System Bus. The data requested by the i386 SX
Microprocessor is provided over the local bus as the first word
of the Line Fill.
b) Servicing cache read hit cycles on the Local Bus while executing
buffered write cycles on the system bus.
c) Servicing cache read hit cycles on the Local Bus while another
bus master is running (DMA, other i386 SX Microprocessor, 82396SX
Smart Cache, i486 SX Microprocessor, etc...) on the System Bus.
d) Buffering write cycles on the Local Bus while the system bus is
executing other cycles. Write protected areas are supported by
the SWP# input. This enables caching of ROM space or shadowed ROM
space.
o No Post Input (NPI#) provided for disabling of write buffers per
cycle. This option supports memory mapped l/O designs.
o Byte Enable Mask (BEM) is provided to mask the processor byte
enables during a memory read cycle.
o A2oM# input provided for emulation of 8086 address wrap-around.
o SRAM test mode, in which the TAGRAM and the cache RAM are treated as
standard SRAM, is provided. A Tristate Output test mode is also pro-
vided for system debugging. In this mode the 82396SX Smart Cache is
isolated from the other devices in the board by floating all its
outputs.
o Single chip, 132 lead PQFP package, 1 micron CHMOS-IV technology.
***Versions:...
***Features:...
**82485 Turbo Cache (and 485Turbocache) c90...
**82489DX Advanced Programmable Interrupt Controller 10/12/92...
**82495DX/490DX DX CPU-Cache Chip Set <Sep91...
**82495XP/490XP Cache Controller / Cache RAM (for i860) 06/05/91...
**82496/491 Cache Controller / Cache RAM (for P5 Pentium) 03/22/93...
**82497/492 Cache Controller / Cache RAM (for P54 Pentium) <Nov94...
**82498/493 Cache Controller / Cache RAM (for P54 Pentium) <Nov94...
**
**Later chipsets (basic spec):
**440 series:...
**450NX (?) 06/29/98:...
**????? (Profusion) c:99...
**800 series
***810 (Whitney) 04/26/99...
***810L (Whitney) 04/26/99...
***810-DC100 (Whitney) 04/26/99...
***810e (Whitney) 09/27/99...
***810e2 (Whitney) 01/03/01...
***815 (Solano) 06/19/00...
***815e (Solano-2) 06/19/00...
***815em (Solano-?) 10/23/00...
***815ep (Solano-3) c:Nov'00...
***815p (Solano-3) c:Mar'01...
***815g (Solano-3) c:Sep'01...
***815eg (Solano-3) c:Sep'01...
***820 (Camino) 11/15/99...
***820e (Camino-2) 06/05/00...
***830M (Almador) 07/30/01...
***830MP (Almador) 07/30/01...
***830MG (Almador) 07/30/01...
***840 (Carmel) 10/25/99...
***845 (Brookdale) 09/10/01...
***845MP (Brookdale-M) 03/04/02...
***845MZ (Brookdale-M) 03/04/02...
***845E (Brookdale-E) 05/20/02...
***845G (Brookdale-G) 05/20/02...
***845GL (Brookdale-GL) 05/20/02...
***845GE (Brookdale-GE) 10/07/02...
***845PE (Brookdale-PE) 10/07/02...
***845GV (Brookdale-GV) 10/07/02...
***848P (Breeds Hill) c:Aug'03...
***850 (Tehama) 11/20/00...
***850E (Tehama-E) 05/06/02...
***852GM (Montara-GM) 01/14/03...
***852GMV (Montara-GM) ???...
***852PM (Montara-GM) 06/11/03...
***852GME (Montara-GM) 06/11/03...
***854 (?) 04/11/05...
***855GM (Montara-GM) 03/12/03...
***855GME (Montara-GM) 03/12/03...
***855PM (Odem) 03/12/03...
***860 (Colusa) 05/21/01...
***865G (Springdale) 05/21/03...
***865PE (Springdale-PE) 05/21/03...
***865P (Springdale-P) 05/21/03...
***865GV (Springdale-GV) c:Sep'03...
***875P (Canterwood) 04/14/03...
*Headland/G2...
*HMC (Hulon Microelectronics)...
*Logicstar...
*Motorola...
*OPTi...
*PC CHIPS/Amptron/Atrend/ECS/Elpina/etc...
*SIS...
**5571 (Trinity) Pentium PCI/ISA Chipset (75MHz) <12/09/96
***Info:
[no general section in datasheet]
3. Functional Description
3.1 DRAM Controller
3.1.1 DRAM Type
The SiS5571 can support up to 384MBytes (3 banks) of DRAMs and each
bank could be single or double sided 64 bits FP (Fast Page mode) DRAM,
EDO (Extended Data Output) DRAM, and SDRAM (Synchronous DRAM)
DRAM. Half populated bank(32-bit) is also supported.
The installed DRAM type can be 256K, 512k, 1M, 2M, 4M or 16M bit deep
by n bit wide DRAMs, and both symmetrical and asymmetrical type DRAM
are supported. It is also permissible to mix the DRAMs (FP/EDO/SDRAM)
bank by bank and the corresponding DRAM timing will be switched
automatically according to register settings.
3.1.2 DRAM Configuration
The SiS5571 can support single sided or double sided DRAM modules for
each bank. The basic configurations are shown as the following:
3.1.3 Double-sided DRAM [omitted see datasheet]
3.1.4 Single-sided DRAM [omitted see datasheet]
3.1.5 DRAM Scramble Table [omitted see datasheet]
3.1.6 64-bit mapping table [omitted see datasheet]
3.2 DRAM Performance [omitted see datasheet]
3.3 CPU to DRAM Posted Write FIFOs
There is a built-in CPU to Memory posted write buffer with 8 QWord
deep ( CTMFF). All the write access to DRAM will be buffered. For the
CPU read miss / Line fill cycles, the write- back data from the second
level cache will be buffered first, and right after the data had been
posted write into the FIFO, CPU can performs the read operation by the
memory controller starting to read data from DRAMs. The buffered data
are then written to DRAM whenever no any other read DRAM request
comes. With this concurrent write back policy, many wait states are
eliminated. If there comes a bunch of continuous DRAM write cycles,
some ones will be pending if the CTMFF is full.
3.4 32-bit (Half-Populated) DRAM Access
For the read access, there will be either single or burst read cycle
to access the DRAM which depends on the cacheability of the cycle. If
the current DRAM configuration is half-populated bank, then the
SiS5571 will assert 8 consecutive cycles to access DRAM for the burst
cycle. For the single cycle that only accesses DRAM within a DWord,
the SiS5571 will only issue one cycle to access DRAM. For the single
cycle that accesses one Qword or cross DWord boundary, the SiS5571
will issue two consecutive cycles to access DRAM.
3.5 Arbiter
The arbiter is the interface between the DRAM controller and the host
which can access DRAMs. In addition to pass or translate the
information from outside to DRAM controller, arbiter is also
responsible for which master has higher priority to access DRAMs. The
arbiter treats different DRAM access request as DRAM master, and that
makes there be 5 masters which are trying to access DRAMs by sending
their request to the arbiter. After one of them get the grant from the
arbiter, it owns DRAM bus and begins to do memory data transaction.
The masters are: CPU read request, PCI master, Posted write FIFO write
request, and Refresh request. The order of these masters shown above
also stands for their priority to access memory.
3.6 Refresh cycle
The refresh cycle will occur every 15.6 us. It is timed by a counter
of 14Mhz input. The CAS[7:0]# will be asserted at the same time, and
the RAS[5:0]# are asserted sequentially.
3.7 PCI bridge
SiS5571 is able to operate at both asynchronous and synchronous PCI
clocks. Synchronous mode is provided for those synchronous system to
improve the overall system performance. While in the PCI master write
cycles, post-write is always performed. And function of Write Merge
with CPU-to-DRAM post-write buffer is incorporated to eliminate the
penalty of snooping write-back. On the other hand, prefetch is enabled
for master read cycles by default, and such function could be disabled
optionally. And, Direct-Read from CPU-to-DRAM post-write buffer is
implemented to eliminate the overhead of snooping write-back also. In
addition to Write-Merge and Direct-Read, Snoop-Ahead also hides the
overhead of inquiry cycles for master to main memory cycles. These key
functions, Write-Merge, Direct-Read and Snoop-Ahead, achieve the
purpose of zero wait for PCI burst transfer. The post-write and
prefetch buffers are both 16 Double-Word deep FIFOs.
3.8 Snooping Control [omitted see datasheet]
3.9 AHOLD/BOFF# Process and Arbiter Interface [omitted see datasheet]
3.10 Target Initiated Termination [omitted see datasheet]
3.11 DATA Flow [omitted see datasheet]
3.12 PCI Master Read/Write DRAM Cycle [omitted see datasheet]
***Configurations:...
***Features:...
**5581/5582 (Jessie) Pentium PCI/ISA Chipset (75MHz) <04/15/97...
**5591/5592/5595 (David) Pentium PCI A.G.P. Chipset <01/09/98...
**5596/5513 (Genesis) Pentium PCI Chipset <03/26/96...
**5597/5598 (Jedi) Pentium PCI/ISA Chipset <04/15/97...
**530/5595 (Sinbad) Host, PCI, 3D Graphics & Mem. Ctrl.<11/10/98...
**540 (Spartan) Super7 2D/3D Ultra-AGP Single C.S.<11/30/99...
**55x SoC (System-on-chip) <03/14/02...
**
**Support chips:
**85C206 Integrated Peripheral Controller [no datasheet] ?...
**5595 Pentium PCI System I/O <12/24/97...
**950 LPC I/O <07/16/99...
**Other:...
**PII/III/Pro...
**Athlon etc...
*Symphony...
*TI (Texas Instruments)...
*UMC...
*Unresearched:...
*VIA...
*VLSI...
*Western Digital...
*Winbond...
*ZyMOS...
*General Sources:...
(c) Copyright mR_Slugs Warehouse - All rights Reserved