New DDR5 SDRAM standard supports double the bandwidth of DDR4
JEDEC, the controlling body of the RAM standard, has published details about the DDR5 SDRAM spec to address the demand for ever-faster RAM.
The new DDR5 spec is designed to enable scaling memory performance without degrading channel efficiency at higher speeds. This was done by doubling the burst-length to BL16 and bank-count to 32 from 16. DDR5 DIMM also has two 40-bit independent sub-channels, increasing efficiency, and reliability.
A new feature deemed Decision Feedback Equalization (DFE) enables IO speed scalability for higher bandwidth and performance improvement. DDR5 supports double the bandwidth of its predecessor, DDR4 with 4.8 gigabits per second possible -- but not shipping at launch.
DDR5 also supports on-die ECC scaling features, which enable manufacturing on advanced process nodes.
"The DDR5 standard offers the industry a critical advancement in main memory performance to enable the next-generation of computing required to turn data into insight across cloud, enterprise, networking, high-performance computing and artificial intelligence applications," said Frank Ross, JEDEC Board of Directors member and senior member of Technical Staff at Micron.
Power requirements were dropped from 1.2V to 1.1V compared to DDR4. A voltage regulator on DIMM design reduces power consumption and offers better voltage tolerance for better DRAM yields. DDR5 uses the MIPI Alliance I3C Basic specification for system management bus.
"DDR5 is prepared to improve the Computing Performance by applying various features to overcome future technology scaling challenge and improve performance compared to DDR4. On this basis, DDR5 will lead the evolution of the data-centric era, and will play a pivotal role in the 4th Industrial Revolution," said Uksong Kang, Head of DRAM Product Planning at SK hynix, one of the JEDEC members. "SK hynix is opening up a new sector in the market through the development of the industry's first DDR5 that meets JEDEC standards. We have been working with many partners to verify DDR5 Ecosystem through development of test chips, and modules since 2018, and doing our best to secure mass-production levels in the second half of this year."
The product is expected to arrive in data centers and other cloud computing applications first, and migrate to consumers shortly thereafter. JEDEC has made the spec available for purchase.
The new DDR5 spec is designed to enable scaling memory performance without degrading channel efficiency at higher speeds. This was done by doubling the burst-length to BL16 and bank-count to 32 from 16. DDR5 DIMM also has two 40-bit independent sub-channels, increasing efficiency, and reliability.
A new feature deemed Decision Feedback Equalization (DFE) enables IO speed scalability for higher bandwidth and performance improvement. DDR5 supports double the bandwidth of its predecessor, DDR4 with 4.8 gigabits per second possible -- but not shipping at launch.
DDR5 also supports on-die ECC scaling features, which enable manufacturing on advanced process nodes.
"The DDR5 standard offers the industry a critical advancement in main memory performance to enable the next-generation of computing required to turn data into insight across cloud, enterprise, networking, high-performance computing and artificial intelligence applications," said Frank Ross, JEDEC Board of Directors member and senior member of Technical Staff at Micron.
Power requirements were dropped from 1.2V to 1.1V compared to DDR4. A voltage regulator on DIMM design reduces power consumption and offers better voltage tolerance for better DRAM yields. DDR5 uses the MIPI Alliance I3C Basic specification for system management bus.
"DDR5 is prepared to improve the Computing Performance by applying various features to overcome future technology scaling challenge and improve performance compared to DDR4. On this basis, DDR5 will lead the evolution of the data-centric era, and will play a pivotal role in the 4th Industrial Revolution," said Uksong Kang, Head of DRAM Product Planning at SK hynix, one of the JEDEC members. "SK hynix is opening up a new sector in the market through the development of the industry's first DDR5 that meets JEDEC standards. We have been working with many partners to verify DDR5 Ecosystem through development of test chips, and modules since 2018, and doing our best to secure mass-production levels in the second half of this year."
The product is expected to arrive in data centers and other cloud computing applications first, and migrate to consumers shortly thereafter. JEDEC has made the spec available for purchase.
Comments
Is DDR5 mainly for the CPU and how does it compare to GDDR6 and HBM2?
2 channels of DDR4: 50 GB/s
2 channels of DDR5: 100 GB/s
2 channels of GDDR6: 250 GB/s
2 channels of HBM2: 1000 GB/s
DDR5 is for system or main memory for PCs, servers and such, but it usually comes down to cost. If HBM was cheap, there would be systems using it for main memory, but it is two expensive to be used as system memory for a regular PC. It will be interesting if Apple uses it as main memory for high end Apple Silicon Macs though. I'm almost half expecting it.
There are latency differences that can sway usage of one type of RAM over the other depending on primary application as well.
They sell the Radeon Pro Vega 2 w/32 GB of HBM for $2400. Just multiple by 4, and then add some more. Or look at the 2 Radeon Pro Vega II Duo MPX module upgrade price. Total of 128 GB of HBM with 4 GPUs: $10,800.
Apple is seemingly saying Apple Silicon Macs will have main memory in the SoC package. That has some big ramifications that people still need to digest. It's going to be some very interesting packaging if they put 128 GB of LPDDR5 in-package, let alone 128 GB of HBM. If they make 8 GB/s SSD standard with smaller pools of high speed system memory, workflows that require 1 TB of memory might work out fine or better.
DDR4 and DDR5 are specs for system memory on desktop computers.
LPDDR is for notebook computers, LP standing for Low Power. This type of memory is typically more expensive than the standard desktop computing RAM.
The 'G' in GDDR is graphics memory for discrete GPUs. You cannot stick GDDR into a desktop computer's system memory slots. They are not compatible.
HBM is a newer mobile graphics memory type. Rather than hang separately off the graphics bus, it is usually more highly integrated with the GPU itself, often in the same package. This technology can provide better mobile graphics performance but with higher cost.