الأربعاء، 7 أكتوبر 2020

DDR5-5600 Is Ready, but Intel, AMD Are Sticking With DDR4 for Now

SK Hynix launched the first DDR5-5600 RAM, even though you can’t currently purchase a computer that supports it. SK Hynix is well aware of this, of course, but launching the product is still an important step. It signals to anyone working on proofing a DDR5 solution or designing a DDR5-using chip that they can now buy actual hardware for test evaluation. The first customers, in a situation like this, aren’t end-users or even server deployments — they’re the manufacturers who will build solutions that use these chips. Also, Hynix is starting now so it can have a robust product supply built up by the time the RAM does go mainstream.

DDR5 will arrive with a minimum speed of at 4800Mbit/s, which works out to 76.8GB/s of bandwidth in a dual-channel configuration. DDR5-5600 would support 89.6GB/s, while DDR5-8400 — a still-theoretical configuration that Hynix has pledged to hit — would deliver 134.4GB/s of bandwidth. Once that happens, a dual-channel board would have a bit less bandwidth than an Nvidia GTX 660, a midrange GPU from 2012. A quad-channel board would hit 268.8GB/s, which is only about 20GB/s off the GTX 780, circa 2013.

But, hey, since AMD is prepping an octa-channel board for ultra-high-end workstations, let’s do that comparison, too. An eight-channel DDR5-8400 board, in the event you could load all eight channels at that clock speed, would offer 537.6GB/s of memory bandwidth. At that memory bandwidth level, we get to leave 2013 behind. 537.6GB/s of bandwidth matches the Nvidia Titan Xp from 2017 and beats an RTX 2080 Super from 2019. Now, obviously having the bandwidth of a GPU doesn’t mean your CPU can act like a GPU, but the benchmark nerd in me would love to see how onboard GPU performance scales when you can offer the chip a fire hydrant as opposed to a relative trickle of bandwidth. The vast majority of consumer applications today are latency-sensitive more than bandwidth-sensitive, but there are also interesting implications for AI workloads when you’ve got that kind of bandwidth to play with. GPUs or dedicated hardware would always be faster for AI, but the implications of ~500GB/s of memory bandwidth on a workstation chip are interesting, to say the least.

When Will AMD and Intel Support It?

Zen 3 and Rocket Lake are both DDR4 products, so we won’t see support in 2020 or early 2021. Alder Lake is rumored to support DDR5, but we haven’t heard anything precise about Zen 4 and Zen 5 yet.

It isn’t unusual for both AMD and Intel to introduce supporting consumer chips well after a new memory standard has begun to ship, and the two companies don’t always transfer to a new standard at the same time. Sometimes they delay introductions based on price, to make certain OEMs will be able to build profitable systems using the new RAM.

It would not surprise me if Intel and AMD both waited until 2022 to introduce DDR5 on consumer products. The server launch would happen next year, clearing the way for early product ramps, with consumer following after as yields improve and per-IC pricing falls. If the companies do launch consumer DDR5 support next year, it’ll almost certainly happen at the end of the year, not at the beginning.

One important question in all of this is the degree to which AMD or Intel CPUs are currently memory bound. Zen 3 is rumored to improve AMD’s memory subsystem performance due to the improved L3 cache layout and 8-core unified CCX, but both the Ryzen 9 3950X and the Threadripper 3990X showed strong scaling, even up to 64 cores. There are absolutely memory bandwidth-limited applications that run into the 3990X’s bandwidth limits, but there aren’t many. AMD may not currently feel any particular pressure to move to DDR5. It wouldn’t surprise me if AMD launched Zen at up to eight cores in 2017, doubled consumer core counts with Zen 2 in 2019, then paused for 2-3 years to give developers time to catch up. By 2022, eight-core and higher systems will have been on sale for multiple years.

Even if AMD doesn’t choose to increase core counts at that point, faster GPUs and features like AI will demand DDR5. Memory bandwidth is critical to fast AI processing, and while dedicated chips make use of technologies like GDDR6 or HBM, smartphones and likely some PC processors will probably rely on the good ol’ fashioned DRAM bus. AI is clearly the next frontier in computing and Intel is already integrating CPU-based capabilities like AVX-512.

While we don’t know when AMD will follow suit or how they’ll do it, it’s practically inevitable that they will. That’ll increase the need for memory bandwidth, and we’d expect both AMD and Intel to jump on the DDR5 bandwagon by the time they start integrating dedicated silicon (or additional dedicated silicon) to this sort of processing, if not before. Call it late 2021 at the earliest and probably no later than mid-2023, unless something goes catastrophically wrong with production.

Now Read:



sourse ExtremeTechExtremeTech https://ift.tt/34w4Dct

ليست هناك تعليقات:

إرسال تعليق