Hynix Demos First 16Gb DDR5-5200 DIMMs

November 25, 2018

SK Hynix announced on Thursday that it has successfully tested both conventional DDR5 RAM and

the first registered DDR5 DIMMs as well.

The new chips have a capacity of 16Gb (that’s per-IC, with typically 8-16 ICs per physical DIMM,

for 8-16GB of total memory capacity) and are clocked at DDR5-5200,

which provides up to 41.2GB/s of memory bandwidth per channel.

A conventional desktop outfitted in a standard dual-channel configuration would offer 82.4GB/s of overall memory bandwidth.

That’s a significant improvement over current desktop DDR4. While you can buy DDR4 at a variety of clocks,

the commonly available DDR4-3200 provides 51.2GB/s of bandwidth in a dual-channel configuration.

DDR5 should be more power-efficient than DDR4,

thanks to at least some voltage improvements.

Hynix is advertising a 9 percent decrease in core voltage on their 1Ynm process and claims that their chip meets all JEDEC specifications.

Since the DDR5 spec hasn’t actually been published yet we don’t know if this voltage decrease is typical for the class, but either way,

it’ll yield at least some improvements in power efficiency.

The chips aren’t set to be commercialized until 2020,

and we probably won’t see instant adoption from Intel or AMD — AMD has promised to support current AM4 sockets through Zen 3,

which means either they make a clean break with the old platform with Zen 4 in 2021 or support two different memory standards on the same silicon.

AMD has gone both of these routes in the past, so we’ll have to wait and see what the company opts to do this time.

Intel has not commented on when it will introduce DDR5,

but it’s not unusual for it to take several years between early memory standard demonstration and full commercialization.

As is typical,

we’d expect DDR5 to start off at a cost premium relative to DDR4 before crossing over with the older standard and eventually becoming less expensive.

One thing to always keep in mind is that latency doesn’t improve nearly as quickly as bandwidth. In fact,

DRAM latencies don’t necessarily come down at all unless you overclock or pay for top-quality RAM.

Even as clock speed increases, latency timing keeps pace,

which is why you don’t see DDR4-3200 being sold at the 2-2-2-5 timings that typified really good DDR-400. RAM timings tend to impact the performance of AMD systems a bit more than Intel rigs,

but in both cases,

the adoption of integrated memory controllers and large on-die caches has ameliorated the impact of using slower RAM.

Relative RAM latency

Relative RAM latency

The big winner in all this will likely be integrated graphics solutions.

Dual-channel DDR5 would represent a 1.63x improvement in memory channel bandwidth,

and solutions like AMD’s integrated GPUs continue to be highly sensitive to increases in bandwidth.

APUs like the Ryzen 2400G scale quite well when you open the proverbial throttle and with 1080p’s enduring popularity,

we suspect we’d see these gains translate into higher frame rates as opposed to an immediate push to bump resolution (AMD, of course, would likely advertise both,

but driving higher frame rates in 1080p would be a higher priority we think).

While there’s always been a push and pull between the level of quality offered by integrated graphics and the increased demands of newer titles,

if you look at the situation over the long term,

integrated graphics solutions have improved faster than GPU minimum requirements.

We’ve gotten much better at driving games from iGPUs than we were in 2010 and memory bandwidth increases + on-die GPUs tightly integrated with their CPU counterparts are the two principal reasons why.

There was a time when it looked like HBM might cut into the desktop and laptop market,

even if only a little, but that moment seems to have passed courtesy of HBM’s enduring high prices. DDR5 looks set to have the desktop market to itself,

barring some unusual moves from AMD or Intel, neither of which has shown much inclination to pick up the banner of alternative memory technologies.