Advanced System Memory Infrastructure (ASMI) ============================================== Rationale ========= The lagging of the DRAM semiconductor processes behind the logic processes has led the industry into a subtle way of ever increasing memory performance. Modern devices feature a DRAM core running at a fraction of the logic frequency, whose wide data bus is serialized and deserialized to and from the faster clock domain. Further, the presence of more banks increases page hit rate and provides opportunities for parallel execution of commands to different banks. A first-generation SDR-133 SDRAM chip runs both DRAM, I/O and logic at 133MHz and features 4 banks. A 16-bit chip has a 16-bit DRAM core. A newer DDR3-1066 chip still runs the DRAM core at 133MHz, but the logic at 533MHz (4 times the DRAM frequency) and the I/O at 1066Mt/s (8 times the DRAM frequency). A 16-bit chip has a 128-bit internal DRAM core. Such a device features 8 banks. Note that the serialization also introduces multiplied delays (e.g. CAS latency) when measured in number of cycles of the logic clock. To take full advantage of these new architectures, the memory controller should be able to peek ahead at the incoming requests and service several of them in parallel, while respecting the various timing specifications of each DRAM bank and avoiding conflicts for the shared data lines. Going further in this direction, a controller able to complete transfers out of order can provide even more performance by: (1) grouping requests by DRAM row, in order to minimize time spent on precharging and activating banks. (2) grouping requests by direction (read or write) in order to minimize delays introduced by bus turnaround and write recovery times. (3) being able to complete a request that hits a page earlier than a concurrent one which requires the cycling of another bank. The first two techniques are explained with more details in [1]. To enable the efficient implementation of these mechanisms, a new communication protocol with the memory controller must be devised. Migen and Milkymist SoC (-NG) implement their own bus, called ASMIbus, based on the split-transaction principle. ASMIbus - the big picture ========================= The ASMIbus consists of two parts: the control signals, and the data signals. The control signals are used to issue requests. * Master->Slave: - ADR communicates the memory address to be accessed. The unit is the word width of the particular implementation of ASMIbus. - WE is the write enable signal. - STB qualifies the transaction request, and should be asserted until ACK goes high. * Slave->Master - TAG_ISSUE is an integer representing the transaction ("tag") attributed by the memory controller. The width of this signal is determined by the maximum number of in-flight transactions that the memory controller can handle. - ACK is asserted at least one cycle after STB when TAG_ISSUE is valid and the transaction has been accepted by the memory controller. The data signals are used to complete requests. * Slave->Master - TAG_CALL is used to identify the transaction for which the data is "called". It takes the tag value that has been previously attributed by the controller to that transaction during the issue phase. - CALL qualifies TAG_CALL. - DATA_R returns data from the DRAM in the case of a read transaction. It is valid for one cycle after CALL has been asserted and TAG_CALL has identified the transaction. The value of this signal is undefined for the cycle after a write transaction data have been called. * Master->Slave - DATA_W must supply data to the controller from the appropriate write transaction, on the cycle after they have been called using CALL and TAG_CALL. - DATA_WM are the byte-granular write data masks. They are used in combination with DATA_W to identify the bytes that should be modified in the memory. The DATA_WM bit should be high for its corresponding DATA_W byte to be written. DATA_W and DATA_WM must always be driven low by a master, except during the data call for a write transaction that it has requested. Tags represent in-flight transactions. The memory controller can reissue a tag as soon as the cycle when it appears on TAG_CALL. Performance considerations ========================== Note that the payload of a transaction lasts for only one cycle (i.e. there are no bursts). Therefore, to be able to achieve 100% bandwidth utilization, the issuance of a tag should also take no more than a cycle. For this purpose, the control signals are pipelined. When ACK is asserted, STB can qualify a new request in the same cycle. This puts a constraint on the arbiter, which must be able to switch combinatorially to the next transaction on the assertion of ACK and still meet timing. The controller is not allowed to generate ACK combinatorially from STB. However, the master can generate STB combinatorially from ACK in order to maximize bus bandwidth. STB <0><1><0><0><1><1><0><1><1><0> ACK <0><0><1><0><0><0><1><0><1><1> TAG ------------------ SDRAM burst length and clock ratios =================================== A system using ASMI must set the SDRAM burst length B, the ASMIbus word width W and the ratio between the ASMIbus clock frequency Fa and the SDRAM I/O frequency Fi so that all data transfers last for exactly one ASMIbus cycle. More explicitly, these relations must be verified: B = Fi/Fa W = B*[number of SDRAM I/O pins] For DDR memories, the I/O frequency is twice the logic frequency. Environment =========== The ASMI consists of a memory controller (e.g. ASMIcon) and optionally an arbiter/switch (e.g. ASMIswitch) thanks to which multiple masters can access the shared system memory. Links between them are using the same ASMIbus protocol described above. In order to avoid duplicating the tag matching and tracking logic, the master->slave data signals must be driven low when they are not in use, so that they can be simply ORed together at the arbiter. This way, only masters have to track (their own) transactions. It is suggested that memory controllers use an interface to a PHY compatible with DFI [2]. The DFI clock can be the same as the ASMIbus clock, with optional serialization and deserialization happening across the PHY, as specified in the DFI standard. +-------+ +----------+ |Master1|<==>| | +----------+ +-------+ | | +-------+ +-------+ | Off-chip | |ASMIswitch|<==>|ASMIcon|<-->|DDR PHY|<<====>>| SDRAM | +-------+ | | +-------+ +-------+ |device(s) | |Master2|<==>| | +----------+ +-------+ +----------+ <====> ASMIbus links <----> DFI (or similar) links <<==>> PCB traces to external SDRAM chips Example transactions ==================== Basic transaction: CTL ------------------------ ISSUE------< T1 >------------------ CALL ------------------< T1 >------ DAT_R------------------------ DAT_W------------------------------ Two simple transactions: CTL ------------------------------------ ISSUE------< T1 >------< T2 >------------------------ CALL ------------------------< T1 >------< T2 >------ DAT_R------------------------------------ DAT_W------------------------------------------------ Interleaved transactions: CTL ------------------------------------ ISSUE------< T1 >------< T1 >< T2 >< T1 >< T1 >------------------------ CALL ------------------< T1 >------< T1 >< T1 >------< T1 >< T2 >------ DAT_R------------------------------------------------ DAT_W------------------------------------------------------ Read address x Write address x < Tn > Tag n Data to/from address x [1] http://www.xilinx.com/txpatches/pub/documentation/misc/ improving%20ddr%20sdram%20efficiency.pdf [2] http://www.ddr-phy.org/