News Big movements afoot in ARM server land

[soresu]

Haven't seen a new thread discussing it yet and loathe to use necromancy on an older thread, so here talking about the Neoverse V1 and future N2 cores just announced.

V1 (codename Zeus on previous roadmap) is available for licensing now and seems to be basically the Cortex X1/Hera core with 2x256b SVE1 units, projected as max 96C per socket and +50% IPC over N1.

N2 (codename Perseus) is coming next year and supposedly based on the Cortex Axx core to succeed A78/Hercules, so possibly Matterhorn - it has 2x 128b SVE units though ARM are tight lipped on whether they are SVE2 or not, projected as max 128C per socket and +40% IPC over N1.

Link here to the Anandtech article with lotsa depth.

 

[Gideon]

Considering what Graviton pulls off now with N1, the Zeus platform looks mighty impressive. As Andrei mentioned:

If we take the conservatively clocked Graviton2 with its 2.5GHz N1 cores as a baseline, a theoretical 3GHz V1 chip would represent an 80% uplift in per-core single-threaded performance. Not only would such a performance uptick vastly exceed any current x86 competition in the server space in terms of per-core performance, it would be enough to match the current best high-performance desktop chips from AMD and Intel today (Though we have to remember it’ll compete against next-gen Zen3 Milan and Willow Cove Sapphire Rapids products).

It'll be fun to see how Sapphire Rapid and Genoa compare (though it looks that new Cortex-A and Cortex-X chips arrive a bit sooner).

Regardless, looking at Gravitons current perfomance though, they will almost certainly lose in per-core perf.

If the ARM's sale to Nvidia goes through, the top-of-the-line gaming rig might have a Nvidia CPU inside in 5 years 0.o (though backwards compatiblity will be nonexistant).
It's also looking more and more likely that either Microsoft or Sony will go ARM on their next-gen consoles.

 

[itsmydamnation]

It's also looking more and more likely that either Microsoft or Sony will go ARM on their next-gen consoles.

No it doesn't you just invented that

 

[tamz_msc]

No it doesn't you just invented that

Sooner rather than later the performance equation involving x86 and ARM will change in favour of ARM. When that happens it would be stupid to ignore the higher performance design for a given use case.

 

[Gideon]

No it doesn't you just invented that

Well obviously, as this is a hypothesis not fact. Next-gen console design hasn't probably even started yet, so vendors are still scoping it for a while.
But putting fanboy glasses away, forever using the same AMD CPU + GPU (for every generation till the end of time) will not let them differentiate much between themselves. One or the other will at least try something different at one point.

Obviously it's far from certain. They'll lose backwards compatiblity yet again among other things, But this has not stopped them in the past.

Sooner rather than later the performance equation involving x86 and ARM will change in favour of ARM. When that happens it would be stupid to ignore the higher performance design for a given use case.

Exactly. If ARM Cortex-X cores will significantly outperform x86 cores, I'd be really sad if they do not find themselves in consumer hardware (and me being an AMD fanboy at heart will not change that).

I'm still mad at the Nvidia deal though, as it will 100% turn the industry for the worse and it looks almost certain that it will go through.

 

[A///]

Sooner rather than later the performance equation involving x86 and ARM will change in favour of ARM. When that happens it would be stupid to ignore the higher performance design for a given use case.

If it happens, I can see NVidia winning the contract and not AMD like before. There is a weird clause in NVidia's sales contract that Microsoft is a part of. I caught a portion of a tweet posted on an OC forum elsewhere but someone had dug into NVidia's fiscal reports and found that at some point MS had invested in Nvidia with an acquisition clause. Really interesting stuff. It hadn't bounced on my radar before a couple weeks ago. Last time it got reported on was 2011 on various e-news sites. I don't know how I missed it then.

 

[mikegg]

Sooner rather than later the performance equation involving x86 and ARM will change in favour of ARM. When that happens it would be stupid to ignore the higher performance design for a given use case.

I think ARM has a huge opportunity in the server and laptop space.

But I don't see ARM getting into the PS6/Xbox Next because of backwards compatibility.

 

[A///]

I think ARM has a huge opportunity in the server and laptop space.

But I don't see ARM getting into the PS6/Xbox Next because of backwards compatibility.

Depends if they can pull off an x86-ARM emulation like Apple plans to. Though there would be carry over. Didn't the PS5 drop PS1, PS2 and PS3 support? And Xbox Next? Is Next an official name? Did you just unmask yourself as an Xbox team engineer? Dun dun dun! Just kidding.

If ARM can match x86 in all use cases and use less power under sustained load, say 2-3 minutes minimum at lower TDP and corresponding heat, then x86 is toast. There's a lot of headroom lost on x86 due to terribly designed software. Improving your processor is a better way of overcoming performance obstacles.

 

[soresu]

It's also looking more and more likely that either Microsoft or Sony will go ARM on their next-gen consoles.

RDNA was literally designed for backwards compatibility with the GCN uArch's used in the PS4 and XB1 console generation.

Given that fact the likelihood of a switch away from x86 is non existent in the next gen as it was in this gen.

There's already plenty CPU perf coming in PS5 and XSX, and likely there will be at least 50% IPC uplift by the time PS6/XSX2 are in major design mode given AMD is likely to be on Zen 6+ by then.

I doubt that they will need more than that, the super optimisation of console titles from the limited platform variability usually means that they are more constrained by graphics power anyway (Jaguar cores being the weal sauce exception).

Given that Renoir can manage PS4 APU graphics performance and much greater CPU performance with only 35W on 7nm, you can imagine what they could do at say 3nm = probably a PS4 mini console at less than 10W.

 

[A///]

RDNA was literally designed for backwards compatibility with the GCN uArch's used in the PS4 and XB1 console generation.

Given that fact the likelihood of a switch away from x86 is non existent in the next gen as it was in this gen.

There's already plenty CPU perf coming in PS5 and XSX, and likely there will be at least 50% IPC uplift by the time PS6/XSX2 are in major design mode given AMD is likely to be on Zen 6+ by then.

I doubt that they will need more than that, the super optimisation of console titles from the limited platform variability usually means that they are more constrained by graphics power anyway (Jaguar cores being the weal sauce exception).

Given that Renoir can manage PS4 APU graphics performance and much greater CPU performance with only 35W on 7nm, you can imagine what they could do at say 3nm = probably a PS4 mini console at less than 10W.

Maybe a lot earlier. Don't mid-cycle refreshes on consoles bring increased SoC speeds. Didn't the PS4 get some resolution and color depth improvements that pointed toward a better variant of Jaguar? You may see RDNA3 and Zen3 on the PS5 Pro or whatever they name the refresh.

 

[soresu]

If it happens, I can see NVidia winning the contract and not AMD like before.

You may be unaware of this, but neither Sony nor Microsoft have any desire to work with nVidia on their consoles again.

Prior experience has taught them that nVidia are not especially good team players.

Being top dog means they have a lot of money to throw about, but it also means that they tend to think that all business interactions should favour them, even if they are essentially acting as a contractor in the arrangement.

Given how well the current console agreements have worked out for Sony and MS I highly doubt that they would seek alternative suppliers as anything but a bargaining chip at the table with AMD contract negotiators.

 

[A///]

You may be unaware of this, but neither Sony nor Microsoft have any desire to work with nVidia on their consoles again.

Lol oh I know. I tried telling someone that once and they flamed me out on another board. NVidia has a habit of screwing over their partners. Which is why I found the rumor video from last week about NVidia developing RTX for Nintendo jaw droppingly stupid. Last more confident analysis pointed toward an AMD and RDNA low power SoC for the next big Nintendo product. Probably have part of that wrong.

Won't be surprised if 5 years from now they lock AIB partners out from touching their products. Someone briefly mentioned this... I want to say Steve Burke from GN. He's got a point, regardless of how delicate it is.

 

[soresu]

Don't mid-cycle refreshes on consoles bring increased SoC speeds.

I'd expect GPU improvements but little on the CPU side for a mid crucle refresh this time.

Jaguar was so weak that game devs often ended up using the GPU compute to do things that would otherwise have been done on the CPU, this is why they increased Jaguar clocks on PS4 Pro and XB1X.

On the other hand PS5 and XSX CPU power is a huge step up over the previous gen, not to mention they have a boatload of ASIC offloading hardware to take the strain off the CPU in many situations anyway.

They might increase the clocks a bit, but I would predict a higher concentration on decreasing power so that they can squeeze in more GPU and/or decrease cooling infrastructure which increases costs and volume of the case (which also increases costs).

It would be nice to see them do a PS2 and make a real slim for once - neither PS3 nor PS4 were that impressive in their 'slim' variants.

 

[soresu]

Last more confident analysis pointed toward an AMD and RDNA low power SoC for the next big Nintendo product. Probably have part of that wrong.

It would certainly make sense for them.

Being the only ARM based console means that they are the odd ones out and probably have to make more of an effort to bring developers to their p;platform regardless of how good the tools are.

Going AMD would give them both x86 console parity and a fully modern and RT acceleration capable GPU.

Clearly Samsung have tasted the RDNA sauce and want it for their own low power stuff, so it wouldn't surprise me if Nintendo thought similarly.

 

[soresu]

Didn't the PS5 drop PS1, PS2 and PS3 support?

Functionally yes, but the Zen2 CPU cores could handle PS1 and PS2 easily, and likely with platform specific optimisation PS3 too given how well the RPCS3 team have been doing.

We already know that Sony are familiar with the concept of emulators from the PS mini which uses an emulator, so it's not much of a stretch at all.

Damn, you all derailed me from ARM server talk - you sly dogs got me monologuing!

 

[A///]

I'd expect GPU improvements but little on the CPU side for a mid crucle refresh this time.

Jaguar was so weak that game devs often ended up using the GPU compute to do things that would otherwise have been done on the CPU, this is why they increased Jaguar clocks on PS4 Pro and XB1X.

On the other hand PS5 and XSX CPU power is a huge step up over the previous gen, not to mention they have a boatload of ASIC offloading hardware to take the strain off the CPU in many situations anyway.

They might increase the clocks a bit, but I would predict a higher concentration on decreasing power so that they can squeeze in more GPU and/or decrease cooling infrastructure which increases costs and volume of the case (which also increases costs).

It would be nice to see them do a PS2 and make a real slim for once - neither PS3 nor PS4 were that impressive in their 'slim' variants.

I'll take your word for it.

I was never intrinsically aware of how bad Jaguar was or rather is. I've caught discussion on company chat channels about it. We have off topic and gaming discussion channels and people were talking about how the PS4 was plagued with games that would stutter or slow down and then go back to speed. I had the original PS1, PS2 and the slim, and the PS3. I remember it being an issue with the PS1 with certain games where the game would stick and then go back to fluid gameplay. Or I had bad discs. This was like 20-23 years ago. The slim was pretty remarkable. Around that time I was really into consoles and all those blog sites thought future playstations would be smaller and thinner, or a small cube. I was lucky enough to get my PS3 the week after it came out. It was a nice upgrade, but oh my god that thing was ugly as sin. I never did the class action rebate. It didn't feel right to me because back then I didn't care and still don't care.

Going off topic, the only reason I don't care about backwards compatibility is while I sometimes crave to play an old game, how it looked in my head is usually 300x better compared to actually playing it. There's some original PS1 games that look fine today, because their modern versions are more refined. Then there's games that look awful. There was this one game I remember playing non stop. The MC was a skeleton in armor with a scepter as weapon, IIRC. Some of the film or show based games look nasty now if you look up gameplay videos.

 

[Gideon]

Given that fact the likelihood of a switch away from x86 is non existent in the next gen as it was in this gen.

The likelihood of that IMO highly depends on two things:

1. When will next-gen appear? Is it in 5 years or closer to 7-8 years as the Xbox 360 generation, but this time with incremental upgrades in between. To me the latter makes more sense as current gen is bigger investment up-front than last gen was (production costs of PS5/Xbox Series X are certainly higher than PS4/One were at launch). It's more similar to PS3/Xbox 360 generation in that sense.

2. Will ARM have gained any traction on desktop by that time (next gen might be a bit early for that, especially if it's a 5-year generation).

If Arm gains significant traction in PC gaming during this time (still far from certain) then unified instruction-set across current gen might be more important than backwards-compatibility. I do agree that 5 years is probably too short time-frame for that, even if Arm were to decimate x86 (which is even farther from certain).

EDIT:
Sorry for derailing the thread, perhaps mods could move the console discussion to another thread called "Consoles on ARM or sth"

 

[itsmydamnation]

So i cant find this (might be the long work hours...lol) , when can we actually buy chips?

 

[Gideon]

So i cant find this (might be the long work hours...lol) , when can we actually buy chips?

This is more of a preview as Andrei said in the Anandtech article:

Today’s announcement has been more of a teaser or unveiling, with the company planning to go into more details about the architecture and microarchitectures of the designs at a later date. Arm's DevSummit is scheduled for October 6-8th - and might be where we'll hear a bit more about the new architecture.

Looking at the timeframe on the chart below and when N1 was actually somewhat available (March - May 2020 in Amazon Graviton 2 preview) I wouldn't expect V1 chips before Q1 2021, N2 probably late 2021 or early 2022

 

[piokos]

I can't believe ARM server chip topic turned into another discussion about consoles and Zen. Seriously, come on people...

 

[NTMBK]

RDNA was literally designed for backwards compatibility with the GCN uArch's used in the PS4 and XB1 console generation.

Given that fact the likelihood of a switch away from x86 is non existent in the next gen as it was in this gen.

Backwards compatibility chains can and do break. Nintendo kept the same lineage of backwards compatible PowerPC hardware all the way from Gamecube to the Wii U, but then ditched it entirely and moved over to ARM for the Switch.

 

[LightningZ71]

I believe that the biggest battle, going forward, in the server space, is going to be rack density and power. As more and more applications move to a thin client model where, essentially, your entire application is hosted in a web browser, it matters less and less what architecture is living on the other end of that internet connection. It could be x86, it could be ARM, it could be RISCV, it could be POWER. It doesn't mater at all to the end user in almost all situations where a rack-dense server might exist to host them. What's going to matter to them? The cost will! How much is that company going to have to pay, per worker, to provide the part or service that they are in the business of creating?

If ARM can prove that they can host more virtual servers, or virtual dock hosting instances, or whatever other metric that is relevant in that time frame, than x86 or any of the other platforms can, and at a usable performance level, with a lower power draw per unit of performance, then they will get business. The better they are at that, the more business they will get. It's really that simple.

In most cases, its nearly trivial to compile a given program for a different architecture on the back end where that program isn't explicitly asking for very specific features (such as AVX-512). Yes, there are some highly specialized programs that need very particular things, but, those are far more rare in use than many might believe, and most can be rearchitected to use ANY CPU platform that their author might decide to, just with differences in performance that may or may not be important. Not surprisingly, most of those applications have moved on to actually using add in cards, like compute ASICS, customized GPUS, etc. Again, with appropriate drivers, those could be installed in most ANY CPU architecture type servers.

The only thing holding back a wholesale shift to ARM platforms is purchase cycles and the investment it would take in providing software to monitor, manage, and deploy the bare metal, as well as any software work will need to be done for the operating systems themselves that will run on it. So, assuming that the NeoVerse V1 or N2 is the most amazing thing to ever grace a server rack, you're still looking at a year or more before production will hit enough volume to even begin to make a dent in the installed setups in server farms around the world, and, we're assuming that neither of the two leading x86 vendors does nothing to at the very least keep the performance gap from getting large enough to make performing such a migrations anywhere near cost effective.

 

[Hitman928]

This is more of a preview as Andrei said in the Anandtech article:



Looking at the timeframe on the chart below and when N1 was actually somewhat available (March - May 2020 in Amazon Graviton 2 preview) I wouldn't expect V1 chips before Q1 2021, N2 probably late 2021 or early 2022

N1 was originally announced in Feb. 2019 with actual chips becoming available as you said late Q1 - Q2 of 2020. If ARM is only now sharing the V1 architecture with partners then I wouldn't expect actual chips to be available until Q4 2021 and then N2 to be available some time in 2022.

 

[piokos]

In most cases, its nearly trivial to compile a given program for a different architecture on the back end where that program isn't explicitly asking for very specific features (such as AVX-512).

The real question is not how difficult it is to compile for a different architecture. It's about whether you can rebuild at all, as a user. And whether the person who can has time and money to do it properly.

Even assuming ARM will ultimately offer better performance and will replace x86, the total cost of transition will be enormous. In short time horizon, it will be way beyond what companies can take. So, instead of going for ARM, they will just... buy more x86 hardware.

But I see a flaw in the logic of your post.
Yes, we're migrating most of our computing to servers (cloud or not).
Yes, it means the actual architecture running the backend doesn't matter for the end user.
But it absolutely doesn't mean that we should now look for another market-wide standard. It's the opposite: we can think about a more heterogenous reality - with different architectures optimized for different types of loads.
ARM is not best at everything. And it never will be. It will always be possible to make something that runs a particular (but important) code better - databases, rendering or whatever.
It's a bit like sorting N elements. There is no universally best algorithm for any N. Some work best for "large" N, some for "small" N, some for sets that are totally random, some for sets that have just few elements out of order. And for any given N you can write a nested "if" that will beat everything else.

Even if everyone moves to ARM, it will just fork into specialized branches. It's already happening and there's no way to stop it.

 

[LightningZ71]

It doesn't have to be "best at everything", it has to be "Best in the most common use case" while also not being "unusable or cost prohibitivein the less common, corner cases". I think that, over the long run, ARM will prove capable of meeting both of those goals. Again, there's nothing that makes ARM inherently bad as an ISA. What has kept it where it is is the focus of the vast majority of the products that have used it up till now.