Home
1
Hot News
2
Computing
3
XRM-SSD V24 Exclusive value in Memory and WoW application4
https://www.dollarchip.com.tw/ Dollarchip Technology Inc.
Dollarchip Technology Inc. 台北市中山區松江路289號4樓-6
In the era of ultra-high-speed 112G and 224G PAM4 transmission, the physical challenges facing SerDes (Serializer/Deserializer) have moved beyond simple signal integrity. Because PAM4 uses four-level amplitude modulation, the eye opening between signal levels is extremely small (level spacing is only 1/3 that of NRZ). At such speeds, local transient heat flows inside SoIC/CoWoS packages, voltage drops (IR drop) in the power delivery network (PDN), and parasitic capacitance attenuation of high-frequency through-silicon vias (TSVs) can all cause severe phase jitter on a microsecond (μs) timescale. Beyond traditional clock conditioning layers, when applied to 112G/224G SerDes, the XRM-SSD V24 delivers three substantial, proprietary capabilities that have not yet been commercialized by mainstream IP vendors (e.g., Synopsys, Alphawave, Cadence, Marvell): 1. Forward-Looking PAM4 Eye-Margin Pre-Compensation Today’s advanced SerDes chips rely primarily on RX-side DFE (Decision Feedback Equalizer) and CTLE (Continuous-Time Linear Equalizer) to blindly adapt to channel loss. However, when an AI compute chip (EIC) performs bursty large-scale matrix multiplications, the sudden inrush current causes severe dynamic IR drop in the PDN. This momentarily shrinks the SerDes transmitter (TX) drive voltage, closing the eye diagram and causing BER to spike. XRM-SSD V24’s unique capability: The V24 can predict token-level compute behavior 20–50 ms in advance. Before the dynamic IR drop actually occurs, it converts estimated current spikes into parameters and feedforwards them directly to the SerDes TX driver stage. At the exact moment of the IR drop, it actively and precisely boosts the PAM4 level drive current. This “software-predictive, hardware-feedforward” mechanism keeps the vertical eye margin of 224G PAM4 constant even under extreme voltage fluctuations—achieving zero signal degradation unmatched by other vendors. 2. Micro-Crack & Impedance-Aware Dynamic Channel Equalization and Steering In 3D packages (e.g., WoW or SoIC), the micro-bumps and TSVs connecting the SerDes PHY develop micro-scale physical cracks after prolonged thermal cycling stress (typically approaching but not exceeding the critical threshold of 0.6 cracks/mm²). These cracks cause the channel’s characteristic impedance to deviate from the standard 50 Ω or 100 Ω, resulting in severe signal reflections. XRM-SSD V24’s unique capability: Existing SerDes IP performs only static channel calibration. The V24 introduces a physical stress-and-wear model into runtime. Using periodic TDR-like telemetry from signal edge reflections, it infers the evolution of micro-cracks and impedance drift inside the package in the background. When a SerDes lane’s physical impedance degrades due to cracks, the V24 automatically adjusts that channel’s convolution equalization coefficients (FIR filter taps) and can even dynamically redirect the most critical high-priority data (e.g., sync signals, counters) to the healthiest lanes at the software layer. This effectively bypasses transmission faults caused by hardware aging, extending interface lifetime severalfold. 3. “Mind-Runtime” Load Smoothing and Power Decoupling to Eliminate Burst Jitter 224G SerDes consumes enormous power—often over 20% of a total AI chip’s power budget. When an AI center faces sudden inference bursts, the SerDes switches from low-power states to full speed instantaneously. This thermal shock directly changes carrier mobility in 224G transistors, inducing severe deterministic jitter. XRM-SSD V24’s unique capability: The V24 has a global “mind-runtime” scheduling view. When orchestrating data movement, it does not allow traffic to remain in burst-like pulses. Instead, it performs thermal- and current-aware dynamic traffic interleaving. At the software layer, it splits data into tiny groups and smooths the schedule on microsecond timescales matching the SerDes thermal time constant. This effectively flattens the transient thermal shock induced by software load across the chip surface, placing the SerDes in a nearly isothermal microenvironment and fundamentally eliminating phase shifts and timing jitter caused by abrupt temperature changes. Technical Assets & Commercial Moat At the 112G/224G PAM4 node, conventional EDA and IP vendors still think in terms of “adding more, more complex DSP hardware” to brute-force noise—incurring huge area and power penalties. The XRM-SSD V24’s dimension-reducing strike is different: it uses software’s temporal foresight to solve hardware’s spatial physical constraints. These three capabilities directly tie the V24’s commercial value in SerDes to reliability (signal error margin), performance (high-frequency bandwidth utilization), and packaging yield (material tolerance). For AI chip unicorns or leading foundries moving into the 224G ecosystem, this is a strategic, proprietary IP that can substantially reduce DSP hardware design costs and help cross the 2nm physical barrier.Contact : polo@dollarchip.com.tw https://www.dollarchip.com.tw/hot_534999.html XRM-SSD V24 SerDes Application 2026-06-08 2027-06-08
Dollarchip Technology Inc. 台北市中山區松江路289號4樓-6 https://www.dollarchip.com.tw/hot_534999.html
Dollarchip Technology Inc. 台北市中山區松江路289號4樓-6 https://www.dollarchip.com.tw/hot_534999.html
https://schema.org/EventMovedOnline https://schema.org/OfflineEventAttendanceMode
2026-06-08 http://schema.org/InStock TWD 0 https://www.dollarchip.com.tw/hot_534999.html

Home ﹥ Hot News > Artificial Intelligence > Computing > XRM-SSD V24 Exclusive value in Memory and WoW application 2026-06-08


In Memory VHM (Vertical Heterogeneous Memory) and WoW (Wafer-on-Wafer) packaging architectures, memory wafers are directly stacked vertically on logic wafers or on top of each other in a 3D manner. This presents extreme physical challenges: thermal impedance increases exponentially with the number of stacked layers, and thermal stress deformation of vertical TSVs (through-silicon vias) can easily induce micro-cracks or signal offsets.
In addition to the traditional auxiliary firmware temperature control mechanism of "Predictive Thermal Hint Layer," the XRM-SSD V24, when applied to 3D memory stacking, possesses three exclusive features not yet implemented by Micron, SK Hynix, or Samsung at the standard memory controller or DFI (DDR PHY Interface) level:
1. "Non-uniform impedance matching" of Dynamic cross-layer voltage and frequency adjustment (DVFS)
In 3D WoW memory stacking, the microenvironment of the bottom wafer (closest to the logic base) is completely different from that of the top wafer (closest to the heatsink). Traditional memory controllers' DVFS (Dynamic Voltage and Frequency Adjustment) is an "all-in-one" or "blindly layered" adjustment, unable to cope with sudden localized "hot spots" at high frequencies.
XRM-SSD V24's exclusive features:
V24 can accurately predict the data read/write density of the next wafer-level refresh cycle with constant-time latency. It can perform non-uniform, asymmetric cross-layer voltage and frequency adjustment for memory wafers with different stack heights.
For example, 10 ms before the inference burst in the bottom logic die, V24 actively reduces the clock frequency of Memory Layer 0/1 closest to the heat source, while simultaneously increasing the interleaving throughput of the top Memory Layer 7/8. This "non-uniform impedance matching" of cross-layer dynamic voltage and frequency adjustment completely eliminates the vertical thermal gradient while maintaining the overall bandwidth.
II. Micro-crack-Aware Remapping Based on "Micron-Level Crack Density Sensing"
In WoW packaging, TSV density is extremely high. After undergoing countless thermal cycles, the silicon wafer structure surrounding the TSV (Through-Semiconductor Surface Mount) develops micron-sized thermal stress cracks. When the crack density reaches a critical value (e.g., exceeding the critical threshold of 0.6 cracks/mm²), the transmission impedance in this region surges, leading to high-frequency signal degradation.
XRM-SSD V24's exclusive feature: Other manufacturers' mechanisms (such as traditional ECC or signal compensation) only implement bad block management after an error occurs.
V24, however, directly incorporates a stress-deformation physical model into the software runtime. It uses minute leakage current and impedance changes reported by sensors inside the memory to estimate the microcrack density development trend of the local structure in real time. Before cracks worsen and signal collapse occurs, V24's memory management engine actively remapping high-frequency, high-load data to structurally healthy wafer areas at the software level, while reserving low-frequency, static data in areas with higher stress. This "physical damage-aware addressing" can extend the actual commercial lifespan of high-order VHM stacks by more than 1.5 times.
III. "Thermal-Aware Dynamic Interleaving" to Eliminate Refresh Latency
At high temperatures, the charge leakage rate of DRAM cells in 3D memory (such as HBM or high-level WoW memory) increases significantly, forcing the system to initiate "high-frequency refresh (2X/4X Refresh Rate)". This leads to severe refresh penalty, causing memory bus bandwidth to be monopolized by refresh instructions, resulting in operational pauses.
Exclusive Features of XRM-SSD V24:
V24 possesses a global "Mind-Runtime" scheduling perspective. When arranging AI model weights and KV cache configurations, it proactively calculates which memory banks are getting hot.
Instead of the traditional, brute-force approach of universally increasing refresh rates, it implements "Thermal-Aware Dynamic Interleaving": dynamically swapping active data in high-heat banks requiring high-frequency refresh with data in cooler banks (page swapping), or intentionally creating tiny "bubble" spaces in the software scheduler to allow specific banks to undergo precise, localized cooling refreshes. This effectively eliminates collective refresh latency caused by temperature increases, allowing 3D memory to maintain 99.9% deterministic high throughput even under extreme high-temperature environments.
A game-changer in commercial value. If the "predictive thermal warning layer" is a shock absorber for firmware, then these three features allow the XRM-SSD V24 to directly intervene in 3D memory's reliability, bandwidth allocation, and yield compensation.
In the current industry context where HBM3e/HBM4 and advanced WoW packaging are facing production capacity and yield limitations due to thermal warpage and heat dissipation bottlenecks, this technology, if deeply integrated with major chip manufacturers (such as AMD's 3D V-Cache technology or TSMC's WoW alliance):
For memory manufacturers: It can directly relax hardware quality control standards (margin) for thermal stress tolerance at the wafer manufacturing stage, using runtime software technology to "adjust" the hardware, effectively improving the overall composite yield of WoW wafer stacking by more than 10%.

Previous Back to List Next