AMD Zen 4 CPUs Rumored On Track For 2021 Release | Hacker News Reader

AMD Zen 4 CPUs Rumored On Track For 2021 Release

hothardware.com

54 pointslettergram30 comments

Very cool. For me this hopefully means higher end Zen 3 CPUs will bomb in price so I can upgrade my 2600X

I'd first like to see what Zen 3 CPUs are like though.

Holding out hope for PCIe5 and DDR5 too...

Ok, then - expect to wait at least 5 years, maybe more.

PCIe 4 was standardized in 2011 and it's still not that widely adopted. So let's say 9 years from standardization to wide availability.

PCIe 5 was standardized in 2017... 2017+9 = 2025. Good luck.

I wouldn't count on it. There aren't even that many pci-e 4 devices on the market.

Intel is planning for PCIe5 in 2021 (skipping on PCIe4), so AMD having gen 5 support in that generation would not be too unexpected. Though with the seperation of IO/compute dies I would not be suprised if PCIe5 would only appear in EPYC/Threadripper components.

I expect that the upcoming Ryzen socket refresh (AM5, which is expected to come with Zen 4/Ryzen 5xxx) to have support for (or even require) DDR5, as by then DDR5 will already be available for a year.

In the server market where the extra bandwidth is actually useful, the transition from PCIe 3 to PCIe 4 is pretty far along. Switches, NICs, SSDs and FPGAs supporting PCIe 4 are all either available or announced, and server-class GPUs should also show up next year. Most of those products have good reason to continue on to PCIe 5 as soon as suitable hosts are on the way.

Server-class GPUs kinda showed up first, or at least at the same time: Radeon Instinct MI50 was the first PCIe 4 GPU ever. (The consumer Radeon VII is the same chip)

Ah, I forgot about that one. Possibly because they didn't enable PCIe 4 functionality on the Radeon VII.

There was less than two years between the release of the two standards, so two years between the release of supporting CPUs wouldn't be surprising.

Plenty of gen 4 SSDs on the market

I'm more interested in HBM memory becoming cheaper for GPUs.

I want HBM for CPUs. It's faster, lower power, and lower latency. The only potential downside is being harder to upgrade, but I think that could also be handled with some work.

HBM is not lower latency than traditional DDR. It is significantly higher bandwidth, but the latency can be higher. On GPUs this is not a problem, because they are quite latency tolerant. CPUs might have more trouble hiding the latency impact.

You are correct about HBM2 (latencies between 12 and 16ns vs 8-12ns latencies in typical DDR4 systems). My understanding was that HBM3 reduces those latencies and HBM4 potentially reduces them even further (in addition to lower latency due to closer proximity).

AMD seems considers HBM on-chip to be an upcoming reality.

https://www.guru3d.com/news-story/amd-working-on-vertical-(3...

That's what HBM 3 is meant to accomplish.

Perhaps 2021 will be the year I upgrade. I've been rocking an i5-4690k stably overclocked at 4.5GHz for the past 5 years without issue, but I'd like to get some more threads

The note that Huawei's Hisilicon is one of the three customers for TSMC 5nm is interesting.

The US trade ban is still in effect for Huawei. If we don't lift it, we will likely be blocking out the fastest Android.

Huawei is accused of state-sponsored spying using their position as a major LTE backend equipment vendor.

I don't think worrying about who has the fastest Android phone is important here.

I wasn't trying to make a political statement. I just wasn't aware Huawei was able to get in as one of the only 3 customers. It seems to position them well versus Samsung, Google, etc.

Accused, but still no proof, right?

It's not actually been accused, has it?

The proposed rules are on the basis of hypothetical future spying capabilities.

(Personally and technically, I don't think the ban is a terrible idea. Non-audited software, from a country without an independent legal system, running on privileged network hardware seems a reasonable concern. Especially deployed by rural ISPs who can't perform security due diligence. Even if the public debate is being framed poorly)

It is at least not hypothetical for the US. https://arstechnica.com/tech-policy/2014/05/photos-of-an-nsa...

The US NSA TAO implants are a bit different ball of wax.

Fundamentally, they illustrate that the NSA (government) can't / won't ask Cisco (vendor) to directly install backdoors for them.

So in many ways a validation that checks and balances are working as intended, whatever you think of TAO separately, if supply chain interdiction is required.

The Huawei bans are predicated on the belief that Huawei is not an independent vendor, separate from the interests of the mainland Chinese government.

Cisco was repeatedly caught hardcoding username+password into devices/software. Repeatedly as in “They never seem to learn from their mistakes”. I simply cannot believe Cisco as large vendor will make the same mistakes over and over again. It’s like they try to sneak backdoors into their products, get called out, but continue, yet no one ever cries that the US is spying. Probably because the other four eyes profit from it, as well as some other western countries that have intelligence-sharing deals with US intelligence agencies. Huawei is only called out because their backdoors (… if any exist at all) are not our backdoors.

Sure. But, of course, most cell phones and network equipment is made in China. Vendor cooperation doesn't seem to be needed.

This is an argument of nuances, but I would make the point that the design and software / firmware stages provide fundamentally different capabilities than manufacturing.

You have entire industries around supply chain assurance, so manufacturing corruption is still an issue (although usually more targeted at lower-cost, replaced components).

But as the post-SuperMicro reporting laid out, it's an order of magnitude more difficult to embed physical components on a board, in an undetected manner, that create covert communication channels. Vs doing so at privileged software layers.

It's likely that Qualcomm, Samsung, and HiSilicon will have essentially the same performance because they will use the same cores on the same or similar processes.

I'm ok with waiting for a faster smartphone if that means a totalitarian regime is not spying on me.

Current Intel is so f#%$ed! Either someone at the company has the power and knowledge to make deep foundational changes or not even their size and “business prowess” will be able to maintain the current structure and culture.

Reply on news.ycombinator.com