Hubbry Logo
XDR DRAMXDR DRAMMain
Open search
XDR DRAM
Community hub
XDR DRAM
logo
7 pages, 0 posts
0 subscribers
Be the first to start a discussion here.
Be the first to start a discussion here.
XDR DRAM
XDR DRAM
XDR DRAM
View on Wikipedia
from Wikipedia
XDR DRAM.

XDR DRAM (extreme data rate dynamic random-access memory) is a high-performance dynamic random-access memory interface. It is based on and succeeds RDRAM. Competing technologies include DDR2 and GDDR4.

Overview

[edit]

XDR was designed to be effective in small, high-bandwidth consumer systems, high-performance memory applications, and high-end GPUs. It eliminates the unusually high latency problems that plagued early forms of RDRAM. Also, XDR DRAM has heavy emphasis on per-pin bandwidth, which can benefit further cost control on PCB production. This is because fewer lanes are needed for the same amount of bandwidth. Rambus owns the rights to the technology. XDR is used by Sony in the PlayStation 3 console.[1]

Technical specifications

[edit]

Performance

[edit]
  • Initial clock rate at 400 MHz.
  • Octal data rate (ODR): Eight bits per clock cycle per lane.
  • Each chip provides 8, 16, or 32 programmable lanes, providing up to 230.4 Gbit/s (28.8 GB/s) at 900 MHz (7.2 GHz effective).[2]

Features

[edit]
  • Bi-directional differential Rambus Signalling Levels (DRSL)
  • Programmable on-chip termination
  • Adaptive impedance matching
  • Eight bank memory architecture
  • Up to four bank-interleaved transactions at full bandwidth
  • Point-to-point data interconnect
  • Chip scale package packaging
  • Dynamic request scheduling
  • Early-read-after-write support for maximum efficiency
  • Zero overhead refresh

Power requirements

[edit]
  • 1.8 V Vdd
  • Programmable ultra-low-voltage DRSL 200 mV swing
  • Low-power PLL/DLL design
  • Power-down self-refresh support
  • Dynamic data width support with dynamic clock gating
  • Per-pin I/O power-down
  • Sub-page activation support

Ease of system design

[edit]
  • Per-bit FlexPhase circuits compensate to a 2.5 ps resolution
  • XDR Interconnect uses minimum pin count

Latency

[edit]
  • 1.25/2.0/2.5/3.33 ns request packets

Protocol

[edit]

An XDR RAM chip's high-speed signals are a differential clock input (clock from master, CFM/CFMN), a 12-bit single-ended request/command bus (RQ11..0), and a bidirectional differential data bus up to 16 bits wide (DQ15..0/DQN15..0). The request bus may be connected to several memory chips in parallel, but the data bus is point to point; only one RAM chip may be connected to it. To support different amounts of memory with a fixed-width memory controller, the chips have a programmable interface width. A 32-bit-wide DRAM controller may support 2 16-bit chips, or be connected to 4 memory chips each of which supplies 8 bits of data, or up to 16 chips configured with 2-bit interfaces.

In addition, each chip has a low-speed serial bus used to determine its capabilities and configure its interface. This consists of three shared inputs: a reset line (RST), a serial command input (CMD) and a serial clock (SCK), and serial data in/out lines (SDI and SDO) that are daisy-chained together and eventually connect to a single pin on the memory controller.

All single-ended lines are active-low; an asserted signal or logical 1 is represented by a low voltage.

The request bus operates at double data rate relative to the clock input. Two consecutive 12-bit transfers (beginning with the falling edge of CFM) make a 24-bit command packet.

The data bus operates at 8x the speed of the clock; a 400 MHz clock generates 3200 MT/s. All data reads and writes operate in 16-transfer bursts lasting 2 clock cycles.

Request packet formats are as follows:

XDR DRAM request packet formats[3]
Clock
edge		 Bit NOP Column read/write Calibrate/power-down Precharge/refresh Row Activate Masked write
Bit Bit Description Bit Description Bit Description Bit Description Bit Description
↓ RQ11 0 0 COL opcode 0 COLX opcode 0 ROWP opcode 0 ROWA opcode 1 COLM opcode
↓ RQ10 0 0 0 0 1 M3 Write mask
low bits
↓ RQ9 0 0 1 1 R9 Row address
high bits		 M2
↓ RQ8 0 1 0 1 R10 M1
↓ RQ7 x WRX Write/Read bit x reserved POP1 Precharge delay (0–3) R11 M0
↓ RQ6 x C8 Column address
high bits		 x POP0 R12 reserved C8 Column address
high bits
↓ RQ5 x C9 x x reserved R13 C9
↓ RQ4 x C10 reserved x x R14 C10 reserved
↓ RQ3 x C11 XOP3 Subopcode x R15 C11
↓ RQ2 x BC2 Bank address XOP2 BP2 Precharge bank BA2 Bank address BC2 Bank address
↓ RQ1 x BC1 XOP1 BP1 BA1 BC1
↓ RQ0 x BC0 XOP0 BP0 BA0 BC0
↑ RQ11 x DELC Command delay (0–1) x reserved POP2 Precharge enable DELA Command delay (0–1) M7 Write mask
high bits
↑ RQ10 x x reserved x ROP2 Refresh command R8 Row address
low bits		 M6
↑ RQ9 x x x ROP1 R7 M5
↑ RQ8 x x x ROP0 R6 M4
↑ RQ7 x C7 Column address
low bits		 x DELR1 Refresh delay (0–3) R5 C7 Column address
low bits
↑ RQ6 x C6 x DELR0 R4 C6
↑ RQ5 x C5 x x reserved R3 C5
↑ RQ4 x C4 x x R2 C4
↑ RQ3 x SC3 Sub-column address x x R1 SC3 Sub-column address
↑ RQ2 x SC2 x BR2 Refresh bank R0 SC2
↑ RQ1 x SC1 x BR1 SR1 Sub-row address SC1
↑ RQ0 x SC0 x BR0 SR0 SC0

There are a large number of timing constraints giving minimum times that must elapse between various commands (see Dynamic random-access memory § Memory timing); the DRAM controller sending them must ensure they are all met.

Some commands contain delay fields; these delay the effect of that command by the given number of clock cycles. This permits multiple commands (to different banks) to take effect on the same clock cycle.

Row activate command

[edit]

This operates equivalently to standard SDRAM's activate command, specifying a row address to be loaded into the bank's sense amplifier array. To save power, a chip may be configured to only activate a portion of the sense amplifier array. In this case, the SR1..0 bits specify the half or quarter of the row to activate, and following read/write commands' column addresses are required to be limited to that portion. (Refresh operations always use the full row.)

Read/write commands

[edit]

These operate analogously to a standard SDRAM's read or write commands, specifying a column address. Data is provided to the chip a few cycles after a write command (typically 3), and is output by the chip several cycles after a read command (typically 6). Just as with other forms of SDRAM, the DRAM controller is responsible for ensuring that the data bus is not scheduled for use in both directions at the same time. Data is always transferred in 16-transfer bursts, lasting 2 clock cycles. Thus, for a ×16 device, 256 bits (32 bytes) are transferred per burst.

If the chip is using a data bus less than 16 bits wide, one or more of the sub-column address bits are used to select the portion of the column to be presented on the data bus. If the data bus is 8 bits wide, SC3 is used to identify which half of the read data to access; if the data bus is 4 bits wide, SC3 and SC2 are used, etc.

Unlike conventional SDRAM, there is no provision for choosing the order in which the data is supplied within a burst. Thus, it is not possible to perform critical-word-first reads.

Masked write command

[edit]

The masked write command is similar to a normal write, but no command delay is permitted and a mask byte is supplied. This permits controlling which 8-bit fields are written. This is not a bitmap indicating which bytes are to be written; it would not be large enough for the 32 bytes in a write burst. Rather, it is a bit pattern which the DRAM controller fills unwritten bytes with. The DRAM controller is responsible for finding a pattern which does not appear in the other bytes that are to be written. Because there are 256 possible patterns and only 32 bytes in the burst, it is straightforward to find one. Even when multiple devices are connected in parallel, a mask byte can always be found when the bus is at most 128 bits wide. (This would produce 256 bytes per burst, but a masked write command is only used if at least one of them is not to be written.)

Each byte is the 8 consecutive bits transferred across one data line during a particular clock cycle. M0 is matched to the first data bit transferred during a clock cycle, and M7 is matched to the last bit.

This convention also interferes with performing critical-word-first reads; any word must include bits from at least the first 8 bits transferred.

Precharge/refresh command

[edit]

This command is similar to a combination of a conventional SDRAM's precharge and refresh commands. The POPx and BPx bits specify a precharge operation, while the ROPx, DELRx, and BRx bits specify a refresh operation. Each may be separately enabled. If enabled, each may have a different command delay and must be addressed to a different bank.

Precharge commands may only be sent to one bank at a time; unlike a conventional SDRAM, there is no "precharge all banks" command.

Refresh commands are also different from a conventional SDRAM. There is no "refresh all banks" command, and the refresh operation is divided into separate activate and precharge operations so the timing is determined by the memory controller. The refresh counter is also programmable by the controller. Operations are:

  • 000: NOPR Perform no refresh operation
  • 001: REFP Refresh precharge; end the refresh operation on the selected bank.
  • 010: REFA Refresh activate; activate the row selected by the REFH/M/L register and selected bank for refresh.
  • 011: REFI Refresh & increment; as for REFA, but also increment the REFH/M/L register.
  • 100: LRR0 Load refresh register low; copy RQ7–0 to the low 8 bits of the refresh counter REFL. No command delay.
  • 101: LRR1 Load refresh register middle; copy RQ7–0 to the middle 8 bits of the refresh counter REFM. No command delay.
  • 110: LRR2 Load refresh register high; copy RQ7–0 to the high 8 bits of the refresh counter REFH (if implemented). No command delay.
  • 111 reserved

Calibrate/powerdown command

[edit]

This command performs a number of miscellaneous functions, as determined by the XOPx field. Although there are 16 possibilities, only 4 are actually used. Three subcommands start and stop output driver calibration (which must be performed periodically, every 100 ms).

The fourth subcommand places the chip in power-down mode. In this mode, it performs internal refresh and ignores the high-speed data lines. It must be woken up using the low-speed serial bus.

Low-speed serial bus

[edit]

XDR DRAMs are probed and configured using a low-speed serial bus. The RST, SCK, and CMD signals are driven by the controller to every chip in parallel. The SDI and SDO lines are daisy-chained together, with the last SDO output connected to the controller, and the first SDI input tied high (logic 0).

On reset, each chip drives its SDO pin low (1). When reset is released, a series of SCK pulses are sent to the chips. Each chip drives its SDO output high (0) one cycle after seeing its SDI input high (0). Further, it counts the number of cycles that elapse between releasing reset and seeing its SDI input high, and copies that count to an internal chip ID register. Commands sent by the controller over the CMD line include an address which must match the chip ID field.

General structure of commands

[edit]

Each command either reads or writes a single 8-bit register, using an 8-bit address. This allows up to 256 registers, but only the range 1–31 is currently assigned.

Normally, the CMD line is left high (logic 0) and SCK pulses have no effect. To send a command, a sequence of 32 bits is clocked out over the CMD lines:

  • 4 bits of 1100, a command start signal.
  • A read/write bit. If 0, this is a read, if 1 this is a write.
  • A single/broadcast bit. If 0, only the device with the matching ID is selected. If 1, all devices execute the command.
  • 6 bits of serial device ID. Device IDs are automatically assigned, starting with 0, on device reset.
  • 8 bits of register address
  • A single bit of "0". This provides time to process read requests, and enable the SDO output in case of a read,
  • 8 bits of data. If this is a read command, the bits provided must be 0, and the register's value is produced on the SDO pin of the selected chip. All non-selected chips connect their SDI inputs to their SDO outputs, so the controller will see the value.
  • A single bit of "0". This ends the command and provides time to disable the SDO output.

See also

[edit]

References

[edit]
[edit]
Revisions and contributorsEdit on WikipediaRead on Wikipedia

XDR DRAM

View on Grokipedia
from Grokipedia
XDR DRAM, or eXtreme Data Rate , is a high-performance (DRAM) interface architecture developed by Inc. that enhances standard DRAM cores with a high-speed serial interface to achieve superior bandwidth efficiency using fewer pins and lower power compared to contemporary DDR technologies. Announced in July 2003 as a successor to Rambus's earlier , XDR DRAM employs octal data rate (ODR) signaling, transmitting eight bits of data per clock cycle per pin, with an initial of 400 MHz enabling per-pin data rates of 3.2 Gbps. This differential signaling approach, using Rambus's FlexPhase timing alignment, allows scalable operation up to 7.2 Gbps or higher in later implementations, providing peak bandwidths such as 9.6 GB/s from a single 2-byte-wide device at 4.8 Gbps. Key features include programmable lane widths (up to 32 lanes per chip), support for 8 internal banks, dynamic width control (x1 to x16), and low-latency access times around 2-3.33 ns per request packet, all powered at 1.8 V with options for power-down and self-refresh modes to optimize energy use. The technology was first commercialized by in December 2003 with 512-Mbit devices, followed by and Elpida (now part of Micron), targeting applications in , graphics processing, and networking where high bandwidth in compact, cost-sensitive systems is critical. By 2009, over 100 million XDR DRAM units had shipped, with notable adoption in Sony's console, which utilized 256 MB of XDR DRAM at 3.2 GHz to deliver 25.6 GB/s of system shared between CPU and GPU. Despite its performance advantages, XDR DRAM saw limited mainstream PC uptake due to proprietary licensing requirements and competition from open-standard variants, though evolved the architecture into XDR2 (announced 2005) and later XDRn variants for mobile and high-end computing.

History and Development

Origins and Announcement

Inc. developed as a successor to its earlier technology, aiming to deliver significantly higher bandwidth for demanding applications. The technology, formerly code-named Yellowstone, was officially announced on July 10, 2003, in collaboration with Elpida Memory Inc. and Corp., who committed to manufacturing the devices. This partnership marked a strategic effort to position XDR as a scalable, high-performance solution. The initial specifications targeted a data rate of 3.2 Gbps using octal data rate signaling, with a roadmap extending to 6.4 Gbps and beyond, enabling system bandwidths up to 100 GB/s—eight times that of contemporary PC memory. Toshiba began sampling 512 Mbit XDR DRAM devices in December 2003 at 3.2 Gbps per pin. Samples were slated to ship in 2004, with volume production ramping up in 2005 across densities from 256 Mbit to 8 Gbit and device widths from x1 to x32. The motivation stemmed from the need to overcome bandwidth limitations in consumer electronics, graphics, and networking systems, providing a cost-effective alternative to specialty DRAMs while supporting emerging broadband architectures like Sony's Cell processor. This came after RDRAM's commercial challenges, including high costs and compatibility issues that limited its mainstream adoption despite its speed advantages. Key partnerships expanded beyond the initial collaborators, with beginning mass production of 256 Mbit XDR DRAM devices in early 2005, followed by 512 Mbit versions later that year. These Samsung devices, operating at up to 3.2 Gbps per pin, were claimed to be the world's fastest DRAM at the time, offering up to 9.6 GB/s bandwidth in applications and underscoring XDR's early production milestones.

Evolution and Variants

Following its initial introduction, XDR DRAM underwent iterative improvements in data rates to meet demands for higher bandwidth in high-performance computing applications. The technology began with a pin data rate of 3.2 Gbps in 2003, enabling peak bandwidths of approximately 6.4 GB/s per 16-bit device. By 2006, advancements allowed operation at 4.0 Gbps, increasing peak bandwidth to 8.0 GB/s per device through refinements in signaling and timing circuits. Further escalation occurred in 2008 with the achievement of 4.8 Gbps, delivering up to 9.6 GB/s per device and supporting sustained transfers in the 8-9.6 GB/s range for optimized architectures. Capacity enhancements paralleled these speed increases, with early 512 Mbit devices from 2003 giving way to broader adoption including 256 Mbit and higher densities by 2005, which better accommodated denser memory configurations while maintaining the high-speed interface. This growth in density, combined with the core architecture's octal prefetch mechanism, enabled reliable high-throughput operations without proportional increases in power consumption. In July 2005, proposed XDR2 as a successor variant, announced on July 7 with an initial target data rate of 8 Gbps to achieve even greater bandwidth, incorporating features like micro-threading for parallel access and . Intended for licensing and potential shipping by 2007, particularly in applications, XDR2 was never commercialized, remaining a conceptual extension of the XDR family. Market adoption reflected these refinements, with over 50 million XDR DRAM units shipped worldwide by March 2008, driven by integration in . Shipments surpassed 100 million by June 2009, underscoring the technology's niche scaling in specialized high-bandwidth systems despite competition from DDR variants.

Technical Architecture

Core Design Principles

XDR DRAM utilizes a hybrid architecture that integrates a conventional DRAM core with Rambus's proprietary high-speed signaling interface to deliver enhanced performance while maintaining compatibility with standard memory fabrication processes. This design leverages the reliability and cost-effectiveness of traditional DRAM arrays for and retrieval, augmented by specialized circuitry for rapid I/O operations. The core operates on established principles of , including capacitor-based cells refreshed periodically to retain data, but the interface innovations enable significantly higher transfer rates without altering the fundamental storage mechanism. Central to the architecture is the use of differential signaling for and clock signals using Differential Rambus Signaling Level (DRSL), while address and control signals use single-ended Signaling Level (RSL). This approach transmits signals over paired true and complementary lines for DRSL, providing improved immunity and enabling bi-directional communication at multi-GHz speeds without dedicated ground pins, thus optimizing pin efficiency compared to single-ended methods. DRSL supports data rate (ODR) encoding, where eight bits are transferred per clock cycle on each lane, allowing a 400 MHz clock to achieve effective data rates up to 3.2 Gbps initially, with scalability to higher frequencies. A key design principle emphasizes minimizing the number of high-speed pins to maximize per-pin bandwidth and simplify board routing, contrasting with the wider parallel buses in synchronous DRAMs. For instance, configurations typically employ 32 data pins—comprising 16 differential pairs (DQ and DQN)—to handle narrow but ultra-fast channels, reducing crosstalk and power dissipation while supporting programmable widths such as x8, x16, or x32 for flexibility in system integration. This serialized, point-to-point topology facilitates higher aggregate throughput in bandwidth-intensive applications. Signal integrity in multi-device environments is ensured through on-die termination (ODT), a programmable feature that matches channel impedance directly at the receiver to minimize reflections and stubs. ODT resistors, typically valued at 40–60 Ω, are integrated into the device and calibrated for varying loads, enabling robust operation in daisy-chain or multi-drop topologies without external components. This innovation, rooted in Rambus's earlier developments, addresses challenges of high-frequency signaling over longer traces. The inherently supports multi-channel configurations to scale bandwidth, with devices organized into up to eight internal banks for interleaved access, allowing systems to aggregate multiple independent channels—such as the dual-channel setup in —for overall system throughput exceeding 25 GB/s in practical deployments.

Interface Specifications

The XDR DRAM interface utilizes a compact 144-ball fine-pitch ball grid array (FBGA) package to accommodate high-density pin assignments while minimizing footprint for integration in space-constrained systems. This package supports dedicated pins for key signals, including a differential clock pair (CFM and CFMN) for precise timing synchronization, 16 differential data pin pairs (DQ[15:0] and corresponding DQN[15:0]) for bidirectional transfers, and 12 multiplexed pins (RQ[11:0]) that handle address, command, and control information in a serialized format. The request bus allows parallel (multi-drop) connection to multiple memory devices, enabling shared access to commands and addresses. Additional pins manage termination voltage (VTERM), reference voltage (VREF), and a low-speed serial interface (SDI/SDO with CMD and SCK) for device configuration and initialization, enabling robust operation without external configuration hardware. Signaling in the XDR DRAM interface employs proprietary standards optimized for multi-gigabit speeds, with Differential Signaling Level (DRSL) used for the lines to provide noise immunity and high bandwidth through low-voltage differential pairs, akin to LVDS but tailored for protocols. Signaling Level (RSL), a single-ended low-voltage complementary metal-oxide-semiconductor (LVCMOS-like) scheme, drives the request and control pins for simpler, lower-power transmission of commands and addresses. The interface operates at rates, transferring 8 bits per clock cycle per pin via a combination of (transfers on both clock edges) and internal , achieving up to 4 Gbps per pin at a 500 MHz clock while maintaining signal integrity through on-die termination (ODT) features. The channel architecture of XDR DRAM is fundamentally point-to-point for the high-speed data paths, ensuring minimal reflections and optimal signal quality by directly connecting each device to the without shared buses for data. This design supports dynamic bus width adjustment from x1 to x16 via programmable registers, allowing flexibility for varying system bandwidth needs, and interleaves transactions across eight internal banks for sustained throughput. The low-speed serial configuration bus, however, employs a daisy-chain connecting up to multiple devices in series (RST, SCK, and CMD driven in parallel to all chips, with SDI/SDO chained), facilitating initialization and mode setting without impacting the primary data channel. Electrical specifications emphasize low-voltage operation to reduce power and , with a core supply voltage (VDD) of 1.8 V ±0.09 V for internal logic and array operations, and I/O signaling at 1.2 V ±0.06 V for both DRSL termination (VTERM,DRSL) and reference levels. This dual-voltage approach separates core and interface domains, enabling efficient power delivery while supporting the high-speed differential clock with cycle times as low as 2.00 ns for maximum performance. The interface incorporates fly-by elements in the clock and command distribution to minimize stubs and timing skew in multi-device configurations, though primary remains point-to-point to preserve bandwidth.

Performance Characteristics

Bandwidth and Throughput

XDR DRAM achieves peak bandwidth of up to 9.6 GB/s per device when operating at 4.8 Gbps per pin in x16 mode, leveraging its 16-bit data interface with differential signaling across 32 pins (16 DQ and 16 DQN pairs). This configuration enables high per-pin data rates through octal data rate (ODR) signaling, where data is transferred on multiple clock phases to maximize throughput efficiency. The theoretical maximum can be derived as follows: with 4.8 Gbps per differential pair across 16 pairs (32 pins total), the aggregate is 4.8 Gbps × 16 = 76.8 Gbps, or 9.6 GB/s when divided by 8 bits per byte. Sustained throughput typically ranges from 6.4 GB/s to 8.0 GB/s in multi-burst operations, depending on the clock speed (3.2 Gbps to 4.0 Gbps per pin) and system configuration, achieving over 95% bus utilization in optimized scenarios. This performance is supported by burst length of 16 words, allowing sequential data transfers without bank conflicts in ideal conditions, which minimizes idle time on the bus. Channel aggregation further scales bandwidth; for instance, a 3-channel setup can reach 28.8 GB/s by interleaving accesses across multiple independent channels. The architecture achieves high effective data rates of 3.2 to 4.8 Gbps per pin through octal data rate (ODR) signaling, transmitting eight bits per clock cycle, building on multi-data-rate techniques from prior architectures.

Latency and Timing

XDR DRAM's latency profile is defined by key timing parameters that balance high-speed data transfer with reliable access times. The column address strobe (, denoted as tCAC, is programmable with absolute values typically ranging from 2.0 to 3.33 ns, depending on the speed bin and operating , allowing the memory to deliver the first data word after the column command while accommodating the signaling overhead of the octal data rate interface. Row-related command timings further shape access patterns in XDR DRAM. The row-to-column delay (tRCD) is generally around 15 ns for read operations, representing the minimum interval between a row (ACT) command and a subsequent read (RD) or write (WR) command, with values spanning 5 to 7 cycles depending on the speed bin. Similarly, the row precharge time (tRP), which specifies the delay before a new row can be activated in the same bank, is approximately 10 ns or 6 to 7 cycles. These timings support bank-level parallelism across the device's 8 internal banks. Refresh operations occur to maintain , with the full array requiring refresh within a 64 ms retention window, distributed across multiple refresh commands to minimize disruption. To handle the demands of high-frequency operations, XDR DRAM incorporates deep pipelining, enabling up to 4 outstanding transactions to overlap across banks and reduce effective wait states. However, this results in higher cycle-count latencies compared to SDRAM technologies like DDR, as the faster clock and differential signaling require additional cycles for and synchronization. The interface's periodic ZQ , performed via dedicated commands to adjust output driver impedances, introduces negligible overhead—typically on the order of a few cycles per event—but ensures precise timing margins under varying voltage and temperature conditions.

Operational Features

Power Management

XDR DRAM employs a core operating voltage of 1.8 V ± 0.09 V for both the memory core and interface logic, paired with a 1.2 V ± 0.06 V termination voltage for its differential signaling levels (DRSL) I/O interface. This dual-voltage architecture balances high-speed performance with reduced power dissipation in the signaling domain. Under active operation, a typical XDR DRAM device consumes approximately 2-3 W at 4 Gbps per pin data rates, with read currents around 1.2 A and write currents near 1.12 A at 1.8 V, translating to roughly 2.16 W and 2.02 W respectively for a 256 Mb x16 device. At maximum speeds up to 4.8 Gbps, consumption scales to about 4 W per device due to increased switching activity. Standby power remains low at around 0.61 W (340 mA), dropping further to approximately 17 mW in power-down self-refresh mode, where internal refresh maintains without external clocking. Key power management features include dedicated power-down modes, entered via a PDN command in the column address packet (with XOP[3:0] = 1100), which deactivates the high-speed interface while enabling self-refresh through the Refresh Bank Control Register. This mode, analogous to traditional DRAM's CKE-low power-down, requires a minimum of 16 clock cycles post-command for entry and up to 4096 cycles for subsequent command issuance upon exit, allowing significant energy savings during idle periods. In terms of efficiency, XDR DRAM delivers superior bandwidth per watt compared to its predecessor , achieving 2-3 GB/s per watt at peak operation—for instance, an 8 GB/s bandwidth device at ~3 W yields over 2.6 GB/s per watt—thanks to optimized signaling and low-power PLL/DLL designs that enhance overall energy proportionality. The high-speed interface from the core design principles supports this by enabling sustained transfers with reduced overhead, contributing to up to 40% lower power than comparable graphics memory systems at equivalent bandwidths.

System Integration Aspects

XDR DRAM's channel architecture supports flexible configurations, allowing up to four devices per channel through a daisy-chain for serial interfaces like SDO and SDI lines, where the output of one connects to the input of the next, and the final output links back to the controller. This daisy-chaining reduces the number of (PCB) traces required, simplifying board layout and lowering manufacturing costs compared to parallel configurations that demand dedicated lines for each device. On-chip features significantly ease system design by minimizing the need for external components. Programmable on-die termination (ODT) with adaptive adjusts automatically, providing internal termination resistance for data pins typically between 40 Ω and 60 Ω to mitigate reflections without additional discrete resistors. Similarly, integrated voltage regulation support, including dedicated VTERM pins for differential signaling level (DRSL) termination at 1.2 V ± 0.06 V, reduces reliance on board-level regulators and enhances stability across process, voltage, and temperature variations. The of XDR DRAM devices, such as the 144-ball package in typical implementations, facilitates better thermal dissipation by allowing more efficient heat spreading across the package and board. However, operating at high data rates up to 4 Gbps per pin necessitates careful PCB routing to manage and prevent thermal hotspots from concentrated power delivery. Junction temperatures are specified to remain below 100°C under normal operation to ensure reliability. Compatibility with standard controllers is achieved through Rambus-provided (IP) blocks, including the XDR PHY (XIO) and (XCG), which integrate seamlessly into system-on-chip (SoC) designs for and graphics applications. This IP enables pin-count reduction and supports high-bandwidth interfaces without major redesigns, as demonstrated in mobile XDR variants that deliver over 17 GB/s from a single device while aligning with existing SoC manufacturing processes.

Protocol and Commands

Data Transfer Commands

The XDR DRAM protocol employs a set of core commands for managing data transfers, utilizing a 12-bit command bus (RQ[11:0]) that multiplexes addresses over two clock cycles per command to enable efficient high-speed operations. This structure supports the transmission of 24-bit request packets, including opcodes and address information, across the request queue (RQ) signals during complementary clock phases (CFM/CFMN). The row activate command, denoted by the ACT opcode in the ROWA packet, selects and opens a specific row within a designated bank, preparing it for subsequent column accesses. It includes the bank address (BA) and row address (R) fields, multiplexed over the two-cycle packet, and establishes the row buffer for data availability. Following issuance, the minimum row-to-column delay (tRCD) must elapse before a read or write can target that bank, 5-7 clock cycles for reads and 1-3 for writes depending on the speed grade, ensuring internal array stabilization. Read commands utilize the RD opcode within the packet, specifying the column and initiating from the activated row. The command supports a burst length of 16 transfers, with prefetch mechanisms allowing sequential column data to be queued for output on the differential data strobe (DQS) lines, optimizing throughput in multi-bank interleaving scenarios. Column addresses (C) and sub-column bits (SC) are provided in the packet, enabling fine-grained access to the row buffer contents. Write commands employ the WR in the COL packet for unmasked transfers or the WRM variant in the packet for byte-level masking, directing data input to the specified columns. Masked writes use an 8-bit mask in the command to selectively enable or disable individual bytes within each burst transfer, preventing overwrites on non-targeted data lanes and supporting partial updates. Like reads, writes operate with a burst length of 16, adhering to a write-to-read delay (tWTR) after completion to maintain protocol integrity.

Control and Maintenance Commands

XDR DRAM employs specific control commands to manage operations and ensure through precharge and refresh mechanisms. The precharge command, denoted as PRE, closes an active specified by the bits BA within a ROWP packet, initiating the precharge phase with a row precharge delay of t_RP cycles, 6-7 clock cycles depending on the speed grade. This command is essential for deactivating open rows to prepare for subsequent activations in the same . For refresh operations, the command, encoded in the ROWP packet using the ROP field (such as REFA for all-bank refresh or REFI for incremental refresh), performs auto-refresh across all banks, maintaining with a refresh interval of 64 ms and a per-bank refresh time of t_RFC, which aligns with parameters detailed in the latency specifications. These commands prevent in the volatile DRAM cells by periodically restoring charge levels. Calibration and power management commands in XDR DRAM facilitate signal integrity and energy efficiency. The ZQCL command, issued via a COLX packet with the XOP field set to CALZ or similar encodings, performs impedance calibration by adjusting on-die termination (ODT) resistors to match external conditions, executed periodically every 100 ms with a calibration duration of approximately 12 t_CYCLE to ensure optimal output driver strength and input matching. Power-down entry and exit are controlled through the PD command in the COLX packet (XOP=1100), transitioning the device to a low-power state while preserving data, with entry latency of 16 t_CYCLE and exit managed via the Clock Enable (CKE) signal or Power Management (PM) register settings; CKE low initiates power-down, and high resumes normal operation. These features allow XDR DRAM to reduce power consumption during idle periods without compromising accessibility. Mode register sets () configure key operational parameters in XDR DRAM during initialization and runtime adjustments. The command, transmitted over the command bus, programs registers such as the Configuration (CFG) register for burst length (fixed at 16 transfers) and the () enable bit to synchronize internal clocks with the external clock, reducing skew for high-speed operations. is set via the DLY register, specifying additive latency values like 6 t_CYCLE for read-to-output timing, ensuring precise data timing aligned with . These settings are loaded via serial or parallel interfaces post-reset, enabling flexible adaptation to different system bandwidth needs. The low-speed serial bus in XDR DRAM provides a dedicated interface for device initialization, mode register programming, and tasks like reporting, operating independently of the high-speed data paths. This bus uses a multi-wire configuration including reset (RST), serial clock (SCK), command (CMD), serial data in (SDI), and serial data out (SDO) signals in a daisy-chain , allowing broadcast or targeted access to multiple devices with a of around 50 MHz. Commands follow a structured format with a 4-bit (e.g., SBW for serial broadcast write, SDR for serial device read), followed by address, payload data, and a (CRC) for detection, typically spanning 32 SCK cycles per transaction to configure registers or report status without interrupting main memory operations. This bus ensures reliable setup and diagnostics, particularly during power-up sequences and periodic .

Applications and Legacy

Commercial Adoption

The primary commercial application of XDR DRAM was in the console, launched in 2006, which utilized 256 MB XDR DRAM modules clocked at 3.2 GHz to achieve a system bandwidth of 25.6 GB/s. This implementation leveraged XDR's high-speed differential signaling to support the console's demanding graphics and processing requirements. Beyond gaming consoles, XDR DRAM found use in networking equipment and graphics accelerators, where its superior bandwidth addressed high-throughput needs in specialized systems. However, adoption in personal computers remained limited due to competition from more cost-effective standards. Major manufacturers including , Elpida Memory, and produced XDR DRAM devices, with total global shipments surpassing 100 million units by 2009, driven largely by console demand. By the , XDR DRAM was phased out as DDR3 and GDDR5 technologies dominated mainstream consumer, graphics, and computing markets, offering better scalability and lower costs without proprietary licensing requirements.

Comparison to Competing Technologies

XDR DRAM represents an evolution from its predecessor, , primarily through enhanced bandwidth and improved power efficiency that mitigates the thermal challenges inherent in RDRAM designs. While RDRAM systems, such as those in the , delivered peak bandwidths of 3.2 GB/s across a 32 MB configuration, XDR DRAM scaled to 25.6 GB/s in the PlayStation 3's 256 MB setup, enabling sustained high-throughput operations without the excessive heat generation that plagued RDRAM due to its higher operating voltages and less efficient signaling. This improvement stems from XDR's adoption of differential signaling and octal data rate techniques, which reduce power dissipation per bit transferred compared to RDRAM's earlier architecture. In comparison to , XDR DRAM excels in per-pin bandwidth, achieving 4.8 Gbps versus DDR2-800's 0.8 Gbps, allowing a single XDR device to match the aggregate output of six DDR2-800 x16 devices for equivalent 9.6 GB/s throughput. However, this bandwidth advantage comes at the expense of higher access latency and elevated costs, as XDR's specialized interface demands custom controllers and licensing, making it less suitable for general-purpose computing where DDR2's lower latency (typically 4-6 cycles) and standardization support broader, more affordable integration. XDR thus found favor in bandwidth-intensive applications like gaming consoles, where its superior peak performance justified the trade-offs. Against GDDR4, XDR DRAM offered comparable high-speed capabilities, with data rates up to 4.8 Gbps per pin, but its serial-like differential interface simplified multi-device configurations on narrow buses, potentially easing integration in compact systems. GDDR4, however, gained prevalence in graphics processing units due to its alignment with standards, which facilitated widespread manufacturer support and cost reductions, ultimately leading to its adoption in AMD's HD 2000-4000 series before being supplanted by GDDR5. Overall, XDR DRAM achieved niche success in console hardware, such as the , where its high bandwidth supported demanding real-time rendering, but its proprietary architecture—lacking JEDEC compliance—restricted scalability and ecosystem development, contrasting with DDR technologies' open standards that enabled ubiquitous adoption across PCs and servers.

References

  1. https://www.eetimes.com/rambus-discloses-details-of-new-xdr-dram-interface/
  2. https://www.bdtic.com/datasheet/SAMSUNG/K4Y54164UF_K4Y54084UF_K4Y54044UF.pdf
  3. https://www.e-channelnews.com/toshiba-first-to-commercialize-512-megabit-xdr-dram/
  4. https://www.rambus.com/xdr-dram-surpasses-100-million-units-shipped/
  5. https://www.rambus.com/rambus-technology-is-adopted-in-playstation3-computer-entertainment-system/
  6. https://www.eetimes.com/rambus-to-bring-xdr-memory-to-mainstream-in-06/
  7. https://www.eetimes.com/rambus-reveals-details-of-new-xdr-dram-interface/
  8. https://www.design-reuse.com/news/202506254-rambus-toshiba-and-elpida-announce-xdr-dram-the-world-s-fastest-memory/
  9. https://www.global.toshiba/ww/news/corporate/2003/12/pr2501.html
  10. https://www.cnet.com/tech/tech-industry/rambus-marks-finish-line-for-new-memory/
  11. https://phys.org/news/2005-01-samsung-world-fastest-xdr-dram.html
  12. https://phys.org/news/2005-05-samsung-90nm-512mb-dram-memory.html
  13. https://www.rambus.com/industry-ships-25-million-xdr-dram-memory-devices/
  14. https://www.rambus.com/xdr-dram-passes-50-million-units-shipped-milestone/
  15. https://www.realworldtech.com/cell/11/
  16. https://www.theregister.com/2005/07/07/rambus_xdr_2/
  17. https://investor.rambus.com/press-releases/press-release-details/2009/XDRTM-DRAM-Surpasses-100-Million-Units-Shipped/default.aspx
  18. https://hothardware.com/news/rambus-ships-over-100-million-xdr-dram-modules
  19. https://www.rambus.com/on-die-termination-calibration/
  20. https://www.realworldtech.com/cell-revisited/2/
  21. https://www.rambus.com/elpida-introduces-the-worlds-fastest-dram-based-on-the-rambus-xdr-memory-architecture/
  22. http://application-notes.digchip.com/249/249-48534.pdf
  23. https://www.rambus.com/rambus-to-discuss-trends-and-developments-in-high-speed-interface-technologies-at-denali-memcon-san-jose-2006/
  24. http://utaharch.blogspot.com/2013/11/a-dram-refresh-tutorial.html
  25. https://www.rambus.com/rambus-demonstrates-superior-power-efficiency-of-worlds-fastest-memory/
  26. https://www.rambus.com/rambus-unveils-mobile-xdr-memory-for-next-generation-mobile-products/
  27. https://www.cnet.com/tech/gaming/playstation-3-the-full-specs-list/
  28. https://www.eetimes.com/toshiba-first-to-market-with-512-mbit-xdr-dram/
  29. https://www.infoworld.com/article/2231575/rambus-unveils-xdr-dram-with-elpida-toshiba-2.html
  30. https://www.psdevwiki.com/ps2/Rambus_DRAM
  31. https://ec.europa.eu/competition/antitrust/cases/dec_docs/38636/38636_1192_5.pdf
  32. https://www.rambus.com/qimonda-started-volume-production-of-rambus-xdr-dram-for-playstation3-computer-entertainment-system/
Add your contribution
Related Hubs
User Avatar
No comments yet.