DDR5
The fifth-generation Double Data Rate synchronous dynamic random-access memory standard. Compared with DDR4, DDR5 offers increased memory bandwidth and improved energy efficiency, with an operating voltage reduced to 1.1 V. It is widely deployed in modern server and consumer computing platforms.
Core Facts
Timeline (last 90 days)
The bandwidth of graphics card VRAM (such as GDDR7) is usually several times that of system memory DDR5.
When DDR5 and the RTX 40 series become the standard, people still using DDR4 memory, 8GB of VRAM, and 7B quantized models are at a disadvantage in the hardware community.
Offloading inactive MoE experts to host DDR5 RAM via PCIe Gen5 degrades token generation latency by 8 to 14 times.
Mainstream DDR4 products have been in mass production on the 1y nm (approximately 14–16 nm) and 1z nm (approximately 13 nm) nodes for many years, while DDR5 mass production is concentrated on the 1α nm (approximately 12 nm) node and even more advanced nodes.
Against the current backdrop of memory shortages, manufacturers that had originally planned to discontinue or reduce DDR4 production are reassessing capacity allocation and restarting or maintaining DDR4 production.
Ordinary DDR5 system memory bandwidth is only approximately 100 GB/s.
This custom PC comes with 64GB of DDR5 high-frequency memory.
A screenshot showing 512GB of DDR5 memory with a power consumption of only 16 watts sparked heated discussion on Reddit.
Enterprise-grade memory in the 512GB configuration often carries a price tag of four figures in USD or higher, far exceeding mainstream PC build budgets.
The latest generation of HBM3E can include 8 to 12 DRAM die per stack, providing over 1 TB/s of bandwidth, which is more than 20 times the bandwidth of conventional DDR5 memory (50–60 GB/s).
7 more timeline events
All Facts (16)
The typical bandwidth of DDR5 memory ranges from 50 to 80 GB/s
80%Verified128GB DDR5 memory can fully load all weights of a 70B-class model, but at a speed of approximately 2–5 tokens/s.
65%UnverifiedArrow Lake配套芯片组为Intel 800系列（如Z890、B860等），原生支持DDR5内存和PCIe 5.0通道
90%UnverifiedWith 500 Mbps upload bandwidth spread across 21 concurrent streams, each stream averages only about 24 Mbps, whereas watching 4K content typically requires 25–40 Mbps
60%UnverifiedThe bandwidth of graphics card VRAM (such as GDDR7) is usually several times that of system memory DDR5.
50%UnverifiedWhen DDR5 and the RTX 40 series become the standard, people still using DDR4 memory, 8GB of VRAM, and 7B quantized models are at a disadvantage in the hardware community.
50%UnverifiedOffloading inactive MoE experts to host DDR5 RAM via PCIe Gen5 degrades token generation latency by 8 to 14 times.
50%UnverifiedMainstream DDR4 products have been in mass production on the 1y nm (approximately 14–16 nm) and 1z nm (approximately 13 nm) nodes for many years, while DDR5 mass production is concentrated on the 1α nm (approximately 12 nm) node and even more advanced nodes.
50%UnverifiedOrdinary DDR5 system memory bandwidth is only approximately 100 GB/s.
50%UnverifiedThis custom PC comes with 64GB of DDR5 high-frequency memory.
50%UnverifiedEnterprise-grade memory in the 512GB configuration often carries a price tag of four figures in USD or higher, far exceeding mainstream PC build budgets.
50%UnverifiedThe latest generation of HBM3E can include 8 to 12 DRAM die per stack, providing over 1 TB/s of bandwidth, which is more than 20 times the bandwidth of conventional DDR5 memory (50–60 GB/s).
50%UnverifiedSince the second half of 2024, DRAM prices have entered a new upward cycle, driven by reasons including a surge in demand for HBM from AI servers causing memory manufacturers to shift production capacity toward HBM, thereby squeezing supply of conventional DDR5.
50%UnverifiedGPU video memory (VRAM) bandwidth is typically on the order of 500–1,000 GB/s, whereas DDR5 system memory bandwidth is only approximately 50–80 GB/s, a difference of more than 10x.
50%UnverifiedMemory training finds the optimal signal timing window by automatically adjusting dozens of parameters such as DQS delay, drive strength, and termination resistance at boot time, with training results typically cached in SPI Flash or CMOS; if a new version of AGESA changes the training algorithm or parameter range, it may cause certain memory modules to fail training and subsequently trigger boot failures.
50%Unverified现代存储金字塔中GPU显存(HBM)带宽可达数TB/s但消费级容量通常8-24GB，CPU内存(DDR5)带宽约50-100GB/s，NVMe固态硬盘带宽约5-14GB/s
50%