Computer Architecture System Interface Units Iolanthe II approaches Coromandel Harbour

System Interface Unit (also BusIU) Positioned between cache and system bus (external to die)

System Interface Unit (also BusIU) Cache  main memory bus Responsible for Matching cache line length to memory bus eg PowerPC byte cache line 64 bit (8 byte) bus Memory transactions are “bursts” of 4 double words Giving priority to reads Read requests stall CPU pipeline Writes are assumed “complete” when they exit the pipeline Detecting reads on data in the write buffers Cache coherence - later!

System Interface Unit Read queue 2-4 entries Read will fetch a cache line (PowerPC 8 words) Need to fetch requested word first May be word 0-7 in the cache line Bus transactions may be “out of order” Requested word first Bus clock << processor clock MHz vs MHz ð2 words / bus cycle or 2 words / CPU cycles! ðMany CPU cycles if requested word is last read! 2003 data Add a factor of 3 or so!

System Interface Unit Write buffer 2-3 entries Lower priority than read buffer Additional (R10000) Incoming buffer External Agent (EA) supplying data writes to it Processor doesn’t control transfer EA signals completion and data forwarded to cache Uncached buffer Pages may be marked “not cached” ðFaster I/O transactions

System Interface Unit Overlap  greater bus utilisation Separate Address and Data buses Multiplexing two types of information on the same bus Saves pins Costs time - unable to assert both at once Expensive for writes! Required since …. whenever Separate Address and Data bus tenure Address and data phases split Issue address 1, wait for data 1 Issue address 2 while waiting for data 1

PowerPC organisation PowerPC 601 ~1993 Boundary of the Si die New - Look in the “Example Processors” section of the Web notes

PowerPC organisation PowerPC 601 ~1993 Boundary of the Si die New - Look in the “Example Processors” section of the Web notes 3-way SuperScalar Integer Branch Floating Point

PowerPC organisation PowerPC 601 ~1993 Boundary of the Si die New - Look in the “Example Processors” section of the Web notes MMU Unified TLB (Data and I-misses) Instruction TLB

PowerPC organisation PowerPC 601 ~1993 Boundary of the Si die New - Look in the “Example Processors” section of the Web notes Cache Unified (Data and Inst)

PowerPC organisation PowerPC 601 ~1993 Boundary of the Si die New - Look in the “Example Processors” section of the Web notes SIU Read Q (2 entries) Write Q (3 entries)

Bus Transactions Primary task of SIU Implement bus protocol Sequence of signals and responses 1.Requesting access Several devices can be bus masters 2.Granting access 3.Indicating type of transaction Read, write, read-modify-write, cache invalidate, … 4.Transaction tag Allows multiple transactions ‘in flight’ Slow devices don’t block system 5.Address 6.Data

Bus Protocols Depend on type of bus Serial Single (or small number) of data wires Parallel Smallest number of wires Multiplexed address and data Separate address and data Most complex Requiring separate grants to address and data buses Highest throughput Greatest tolerance for slow devices See following diagram

System Interface Unit Separate Address and Data bus tenure Address and data phases split Issue address 1, wait for data 1 Issue address 2 while waiting for data 1

System Interface Unit Separate Address and Data bus tenure Address and data phases split Issue address 1, wait for data 1 Issue address 2 while waiting for data 1

Separate Address and Data buses Overlap  greater bus utilisation Separate Address and Data buses Separate Address and Data bus tenure Address and data phases split Issue address 1, wait for data 1 Issue address 2 while waiting Increases utilisation of both buses Device latency can be long SIU doesn’t idle while device responds to access request

System Interface Unit Next generation ( eg PowerPC 620) Multiple transactions active at any time Each transaction identified by a transaction ID (3 bits on bus) Allows multiple processors to interleave transactions on the bus Further tolerance for long, variable latencies Memory and devices may have different latencies Multiple levels of memory Devices: discs, networks, graphics, etc Very long latency operation doesn’t block a (just) long latency one

Parallelism Modern systems derive performance from ability to perform many tasks in parallel Datapath Pipelined – one instruction / stage Instruction Issue Unit Instruction Q Superscalar 3+ functional units execute multiple instructions SIU I/O transaction Qs System bus Several transactions active at any time Peripherals Several servicing requests at the same time