Question Qualcomm's first Nuvia based SoC - Hamoa

[poke01]

Qualcomm's working on a 2024 PC chip codename "Hamoa" with up to 12 (8P+4E).

Said to the same cache layout as M1. large private L1$, per-cluster L2$ (cluster = 4 cores, 12MB for every cluster) and a lot of LLC.

source:

https://twitter.com/x/status/1589405172979339264

 

[adroc_thurston]

I disagree completely. I could right now replace my work laptop with it and continue working like nothing happened. Node.js, VSCode, WSL2... That's pretty much all I need. Things are improving and more and more native software is available.

Webdevs can use basically anything, they're the least sensitive segment anywhere.
PC s/w ecosystem is helluva lot wider than that.

 

[naukkis]

Duh.
Chongus shared L2 is a really-really bad fit for server which is why everyone not Apple moved to private L2 + shared L3.

Everyone already did.
It's 32/64K L1's 1/2/3/whatever megs of private L2 and then a pool of shared victim L3.
Zen, coves, Cortex-A/X, Neoverse N/V, pretty much all R-V designs, weird chinesium; all do that.

<redacted> L1 size has nothing to with L2 being shared or private - Apple does that scheme because it can. And sharing L2 isn't problem at all - lets look about IBM latest and greatest Z16 - 256KB L1 with 32MB private L2 per core which can be shared between cores, thus eliminating need of additional shared cache level. L1 size has most impact of designs energy-efficiency - it's a tradeoff between cache traffic energy vs. L1 energy consumption and at this point it's pretty obvious that larger L1 caches are clearly more efficient. And eliminating unnecessary L2-L3 traffic too like IBM have done will improve efficiency too - even in that Apple's shared L2 approach.

Please refrain from using profanity in the tech section. -AT Moderator Shmee

 

[moinmoin]

Perhaps on Windows, I don't know much about their quality in that arena, but in the Linux shop they have 2 open source drivers, Freedreno for OGL and Turnip for Vulkan - both of which only have basic support from Qualcomm themselves and are largely community driven efforts.

I'm not sure about their level of optimisation, but they are easily the most feature complete ARM SoC gfx drivers for Linux and support everything up to OGL 4.6 and VK 1.3 which blows everything else in the ecosystem out of the water including thf V3D drivers for Raspberry Pi 4 and 5.

That's more like an indictment of the overall absolutely sorry state of driver support all ARM SoC manufactures deliver right now. Which is a complete embarrassment considering their products are being used with Android for like more than a decade now, and the driver layer there still is Linux. The fact that Freedreno and Turnip are largely community driven efforts should have QC die in terrible shame.

 

[Tup3x]

Webdevs can use basically anything, they're the least sensitive segment anywhere.
PC s/w ecosystem is helluva lot wider than that.

Sure but the point is that for quite a few the software support is a non issue and things are improving. Also X Elite is pretty ideal CPU for my use case. So, at least they got are covered now.

Big things need to happen before ARM chips have any place in gamer's rigs and powerful workstations (CAD etc.) though. This hardware doesn't really cover that use case anyway.

If we see NVIDIA/AMD GPU in ARM PC then things could change. Also Qualcomm did say that they are working with game developers so I guess they are trying to do at least something.

 

[FlameTail]

I am interested to see what kind of SKU segmentation the Snapdragon X Elite will have.

The 12-core CPU with the 4.6 TFLOP GPU can't be the only SKU.

 

[NTMBK]

<redacted> L1 size has nothing to with L2 being shared or private - Apple does that scheme because it can. And sharing L2 isn't problem at all - lets look about IBM latest and greatest Z16 - 256KB L1 with 32MB private L2 per core which can be shared between cores, thus eliminating need of additional shared cache level. L1 size has most impact of designs energy-efficiency - it's a tradeoff between cache traffic energy vs. L1 energy consumption and at this point it's pretty obvious that larger L1 caches are clearly more efficient. And eliminating unnecessary L2-L3 traffic too like IBM have done will improve efficiency too - even in that Apple's shared L2 approach.

Please refrain from using profanity in the tech section. -AT Moderator Shmee

If bigger L1 caches without blowing up latency was easy, everyone would do it. Having more cache close to the CPU is a very obvious win, but it's extremely hard to increase the size without making it slower.

Apple has the best team in the industry, and that's why they can make such big L1 caches perform well.

 

[darkswordsman17]

Its probably been mentioned by now but I guess Intel went out of their way to be like "eh, no one cares", which means Intel likely does care and is a good sign this could be solid.

Personally, even if the chip lives up to the potential, there are almost guaranteed to be many many other issues that makes it not especially great other than for basic tasks (think Chromebook level) that most any current SoC including ARM can probably do well enough), that I won't be buying such a device.

The main thing I'm hoping is this makes AMD and Intel finally make the GPU heavy consumer APU that would be what I'd want. Basically double Phoenix with faster soldered RAM or CAMM, maybe with an extra memory channel or two, or whatever will offer bandwidth to make the extra GPU worthwhile (I said years back how I wish from the get go with Zen that we'd have gotten APUs with sizable GPU and then use HBM as system memory). Basically I wanted an Apple like design from AMD and Intel, which they have dragged their feet on leaving the door open for another ARM design to offer similar and get similar enthusiasm about.

Sure but the point is that for quite a few the software support is a non issue and things are improving. Also X Elite is pretty ideal CPU for my use case. So, at least they got are covered now.

Big things need to happen before ARM chips have any place in gamer's rigs and powerful workstations (CAD etc.) though. This hardware doesn't really cover that use case anyway.

If we see NVIDIA/AMD GPU in ARM PC then things could change. Also Qualcomm did say that they are working with game developers so I guess they are trying to do at least something.

Haven't they been saying that for years in the mobile space? I could swear they even touted updatable drivers at one point (although maybe that was someone else).

That's the biggest issue, Qualcomm doesn't really deliver on the software side of things. They tout a lot of features but many of them don't end up getting implemented directly (often times companies develop their own alternative even). And they leverage their cellular modem to gain the hardware advantage. Which, that's actually I think the most significant part of this and why this might finally gain traction, namely with businesses (who would love to keep even better tabs on work equipment. Personal use, we'll see (I don't believe cellular companies were receptive to these types of devices as they tend to be power users that use a lot of bandwidth, but then they can just force people to a new higher tier so they probably won't complain if these become more popular).

And there's a lot of Windows software that needs updated.

I am interested to see what kind of SKU segmentation the Snapdragon X Elite will have.

The 12-core CPU with the 4.6 TFLOP GPU can't be the only SKU.

I actually hope it is, as I think that will be to their benefit. It enforces a certain level of performance and a single chip focuses development. Then in the future they can make higher tier versions, like Apple did with the M series). Which maybe they could offer a slightly lower version with like 10 cores or somewhat scaled back GPU. But I wouldn't want something that's like half. And 2/3 might be ok but chances are there's not going to be significant difference in price.

Maybe they could offer a lower clocked efficiency focused version. I wouldn't mind seeing that. But then I'd rather that just be an optional switch (maybe work with device manufacturers to make a physical button or hardware switch, or have it have to be with different modes like docked vs undocked, etc).

 

[soresu]

That's more like an indictment of the overall absolutely sorry state of driver support all ARM SoC manufactures deliver right now. Which is a complete embarrassment considering their products are being used with Android for like more than a decade now, and the driver layer there still is Linux. The fact that Freedreno and Turnip are largely community driven efforts should have QC die in terrible shame.

ARM Ltd is no different.

They seem to have had a change of heart lately* with supporting OSS efforts under Panfrost after Alyssa Rosenberg stood it up, but their past support was non existent.

*presumably due to the popularity of RK3588 based products and the increasing dearth of ARM gfx customers in the hi end.

I agree wholeheartedly, all ARM gfx vendors should die of shame from what I have heard on the likes of the FEX-emu discord, and the Dolphin emu (Android port) dev team in the past.

 

[naukkis]

If bigger L1 caches without blowing up latency was easy, everyone would do it. Having more cache close to the CPU is a very obvious win, but it's extremely hard to increase the size without making it slower.

Apple has the best team in the industry, and that's why they can make such big L1 caches perform well.

Cache sizing and optimizing is trivial thing to do. But it's a big step to go towards cache arrangement that mostly benefits efficiency when designing high-performing part. Apple doesn't care about piss contest of being fastest so they could optimize design for efficiency. Being closo to be fastest seems to be just coincidence.

 

[soresu]

If we see NVIDIA/AMD GPU in ARM PC then things could change

There's no reason hypothetically that this can't happen already with the right SoC and board configuration having a PCIe x16 slot.

The only thing stopping this from happening is purely software bound (beyond lack of PCIe x16 IO from the SoC 😅) as Linaro people have already had a discrete gfx card working on an Ampere type system.

 

[igor_kavinski]

Intel doesn’t think that Arm CPUs will make a dent in the laptop market

"They've been relegated to pretty insignificant roles in the PC business."

arstechnica.com

The same Intel that turned down Apple for the iPhone SoC and then failed to make a decent mobile CPU for the Android phone market?

This is not gonna go the way they think it will

 

[soresu]

The same Intel that turned down Apple for the iPhone SoC and then failed to make a decent mobile CPU for the Android phone market?

The preexisting native ARM software library on Google Play also played a role in that.

If everything ran on Java things might have turned out differently for Intel.

Alas they got almost the entire Chrome OS market as a consolation prize - ironically that is slowly going down the tubes too as the Android tablet market picks up steam again with Google making large screen UI improvements to the OS.

 

[FlameTail]

Question: A lot of apps already have ARM versions in MacOS right. So why are the developers dragging their feet in bringing ARM versions to Windows?

The already existing ARM version in MacOS can simply be ported into Windows.

If that's not so simple and easy, atleast the said devs have experience with x86 -> ARM porting, when Macs transitioned to Apple Silicon.

 

[Thibsie]

Apple has the best team in the industry, and that's why they can make such big L1 caches perform well.

Had ?

 

[soresu]

This is not gonna go the way they think it will

So far I haven't been given to the impression that QC are very sensible about pricing in the higher end.

They are going to need to be a lot more competitive on price in order to take a big bite out of the Windows market or it won't matter how good X Elite or Oryon is, because at the end of the day many people will go on price.

 

[Thibsie]

Question: A lot of apps already have ARM versions in MacOS right. So why are the developers dragging their feet in bringing ARM versions to Windows?

The already existing ARM version in MacOS can simply be ported into Windows.

If that's not so simple and easy, atleast the said devs have experience with x86 -> ARM porting, when Macs transitioned to Apple Silicon.

Because Apple has a clue how to do it, they're used to it.
M$ has no clue.

 

[soresu]

The already existing ARM version in MacOS can simply be ported into Windows.

Backend code should be very portable (barring stuff like native Metal GPU code), but the MacOS frontend isn't getting ported, that's a totally different beast to Windows with different UI APIs and the BSD/Unix based OS it's running on.

It would be far more likely that an existing Wintel based UI would be ported to use WoA UI APIs instead like with foobar2000.

I'd be extremely surprised if the whole of Adobe CC isn't well into the process of porting if it's not done already.

As for the rest like Autodesk, SideFX, Epic/Unreal, Unity etc etc that will happen when it happens - some of them are still in the process of porting their apps to run on MacOS which has a far larger marketshare than WoA devices at the moment.

If it doesn't already have a Wintel version I wouldn't expect it to get a WoA port for quite some time like we saw with the Affinity Photo app years back taking its sweet time to release on Windows after MacOS.

 

[DrMrLordX]

The same Intel that turned down Apple for the iPhone SoC and then failed to make a decent mobile CPU for the Android phone market?

Shh Gelsinger is still watching AMD in his rearview mirror, he hasn't got time for all that nonsense.

 

[FlameTail]

Shh Gelsinger is still watching AMD in his rearview mirror, he hasn't got time for all that nonsense.

He'd better check his windscreen soon or he's car is gonna crash into a huge truck called "ARM". /s

 

[beginner99]

maybe they could offer a slightly lower version with like 10 cores or somewhat scaled back GPU

Don't they have to? nothing yields perfectly and I bet they will have enough salvage parts for a cut. 8-cores and smaller gpu makes the most sense.

They are going to need to be a lot more competitive on price in order to take a big bite out of the Windows market or it won't matter how good X Elite or Oryon is, because at the end of the day many people will go on price.

Fully agree. I see a niche for this as mentioned earlier (traveling business manger that needs battery life, 5g modem but only few common apps). But for more mainstream adaption, price will matter a lot especially since most persons don't want to pay a premium for having to deal with software compatibility issues.

The fact the keynote mentioned nothing about x86->ARM in terms of software and hardware like in Apples M2 tells me it simply doesn't exist.

 

[roger_k]

Looking at QC marketing slides again I can’t shake the impression how similar Oryon is to Firestorm/Avalanche. Same basic cluster organization, very similar feature set (no SVE), even the IPC seems pretty much identical. Oryon is built on a smaller node than M1 or M2, and it has a turbo mode for a nice ST boost, but at the same time I’d expect the 12-core config to have better multi core performance.

A lot has been argued about Apple losing their crack CPU design team with Williams and colleagues, and how their CPU design is stagnating, but it doesn’t really look to me like the crack team (in its new environment) has managed to make any notable progress either. In fact, it seems like what they did was to recreate their work at Apple. Don’t get me wrong, these are obviously very strong cores, but I have expected a bit more after all the bold claims. And I think it’s worrisome that QC didn’t mention anything like x86 memory order emulation, that would be really helpful to improve the performance and stability of ARM Windows.

 

[coercitiv]

Don’t get me wrong, these are obviously very strong cores, but I have expected a bit more after all the bold claims.

It makes sense though, they were iterating on a budget. First you bring a viable product to light, get your funding in order, and only then you get to play with big ideas. I heard that even leaving Apple was accompanied by a colossal waste of energy, or at least the lawsuit was. My impression is the team set a target for performance and efficiency and then worked on delivering exactly that in as little time as possible. No distractions or else...

I was a Nuvia skeptic when they were still "shopping" for a buyer, now after the QC acquisition and the second round of claims I'm willing to give them the benefit of the doubt and wait for their next design: let's see how much of a "crack CPU design team" they really are.

 

[FlameTail]

You know I am glad Nuvia was acquired by Qualcomm. Is there a better company they could have been acquired by?

Some rumours said that Google was mulling acquiring Nuvia before Qualcomm swooped in. If Google had done so, Nuvia would have been relegated to designing cores for a smartphone chip. Not worth for all their talent and ambition.

By being acquired by Qualcomm, the excellent Nuvia/Oryon CPUs can be paired with other strong IP Qualcomm already has- the Adreno GPU, Hexagon NPU, Spectra ISP, Sensing Hub etc... and move onto ambitious projects. We are seeing this awesomeness now with the Snapdragon X Elite breaking new ground in the PC world.

 

[DrMrLordX]

You know I am glad Nuvia was acquired by Qualcomm. Is there a better company they could have been acquired by?

Apple. Hah.

Some rumours said that Google was mulling acquiring Nuvia before Qualcomm swooped in. If Google had done so, Nuvia would have been relegated to designing cores for a smartphone chip. Not worth for all their talent and ambition.

Google has ambitions along the same lines as Amazon: to develop their own in-house server chip/platform. Not sure what came of that, but the Nuvia team would have been a good fit. Their ideas would never have left Google's internal cloud structure, though.

By being acquired by Qualcomm, the excellent Nuvia/Oryon CPUs can be paired with other strong IP Qualcomm already has- the Adreno GPU, Hexagon NPU, Spectra ISP, Sensing Hub etc... and move onto ambitious projects. We are seeing this awesomeness now with the Snapdragon X Elite breaking new ground in the PC world.

At least this way we'll see their product in action on the consumer side. But I doubt they'll ever get to explore their server ambitions. At least not anytime soon.

 

[roger_k]

It makes sense though, they were iterating on a budget. First you bring a viable product to light, get your funding in order, and only then you get to play with big ideas. I heard that even leaving Apple was accompanied by a colossal waste of energy, or at least the lawsuit was. My impression is the team set a target for performance and efficiency and then worked on delivering exactly that in as little time as possible. No distractions or else...

I was a Nuvia skeptic when they were still "shopping" for a buyer, now after the QC acquisition and the second round of claims I'm willing to give them the benefit of the doubt and wait for their next design: let's see how much of a "crack CPU design team" they really are.

Yours is a fair take, and it sounds reasonable. It's just you know, I've been thinking about those charts published by Nuvia in 2020, with that big blue Phoenix rising above everything else, and yet it's three years later and they are only within the lower border of that (admittedly ambiguous) prediction. I suppose we will have to wait and see what they will come up with in the future.