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

 

[FlameTail]

https://d1io3yog0oux5.cloudfront.net/_33a7aedec0c124030df2d09df76fbc53/qualcomm/db/703/7060/file/+Investor_Day_2021_JThompson_PDF.pdf

Stumbled across this old document from Qualcomm. Some very interesting information. Notably they said they can scale the Adreno GPU upto 15 TFLOPS for Laptop chips. The X Elite's 4.6 TFLOP is a far cry from that.

 

[FlameTail]

AnandTech Forums: Technology, Hardware, Software, and Deals

Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

www.anandtech.com

Revisiting Nuvia's claim for the Phoenix core back then. Has the Oryon core matched/exceeded that claim?

 

[FlameTail]

As noted earlier, both Wi-Fi and the modem are discrete for this product. The chip is intended to be paired with Qualcomm’s FastConnect 7800 silicon in the form of an M.2 card. The 7800 is their latest-generation Wi-Fi 7 solution, with support for 4 spatial streams as well as Bluetooth 5.4. The modem pairing is the Snapdragon X65, a high-performance 5G modem which was also available for the 8CX Gen 3.

The fact that neither wireless system is integrated into the SoC is unusual for Qualcomm, but perhaps not too surprising since they want to bring the Elite to market ASAP. Integrating these modules would take further time, and as a laptop SoC, Qualcomm doesn’t need to be as space efficient. In any case, the official line from Qualcomm is that the discrete modem is for OEM flexibility – to give OEMs the option to either include a modem or not – though Qualcomm of course will be strongly encouraging OEMs to include one as a major feature differentiator of the platform.

-Anandtech article on the SD X Elite.

That's quite interesting. Is it beneficial to have these modules integrated or discrete? Nodes are getting more expensive while there are diminishing gains in analog density.


Perhaps integrating the WiFi+Bluetooth module while leaving the 5G modem discrete is the best way. Every PC needs WiFi and BT, whereas 5G remains optional.

 

[FlameTail]

I am guessing this is the integrated Fastconnect (WiFi + Bluetooth) module in the Snapdragon 8 gen 2.

If so, it isn't taking up a lot of area, is it?

 

[soresu]

View attachment 88747
I am guessing this is the integrated Fastconnect (WiFi + Bluetooth) module in the Snapdragon 8 gen 2.

If so, it isn't taking up a lot of area, is it?

Wut - the VPU at the top middle is enormous!?!?

That can't be just encoder/decoder ASICs surely....

 

[FlameTail]

Wut - the VPU at the top middle is enormous!?!?

That can't be just encoder/decoder ASICs surely....

Hmm. Perhaps they mislabelled it?
Also isn't the NPU suspiciously too small?

 

[soresu]

Hmm. Perhaps they mislabelled it?
Also isn't the NPU suspiciously too small?

As the 'NPU' is the Hexagon DSP block which is one of their major silicon lineages all the way back to the pre ARMv7 days I'd say yes, it looks much too small to be correct.

 

[mikegg]

Hey guys.

Some thoughts...

View attachment 88675

Oryon: 3227@4.3 GHz
M2 Max: 2841@3.7 GHz

They say Oryon can match M2 Max performance at 30% less power. That's significant.

So for Oryon to match M2 Max... it's clockspeed would be:

3.78 GHz.

Which is roughly equal to the base clock speed of 3.8 GHz.

_______

Oryon at 3.8 GHz matches M2 Max ST performance at 30% less power.

That is significant. The performance per watt of this core looks great. I guess this is what Amon and Gerard were referring to when they talked about how the real silicon turned out to be much better than the simulations. It's not the raw performance, but the efficiency.

But that isn't my point.

My point is... why did they then limit the base clock speed to 3.8 GHz?

Why the need for Dual-core boost?

Considering how good the performance per watt is, they could run all cores at 4.3 GHz. Or atleast 4 GHz like Apple's doing in the M3 series.

Unless... as the rumours have us suspecting- there are deficiencies in the uncore.

That would explain why the all-core speed is limited to 3.8 GHz, and also why the X Elite- a CPU with 12 P cores is underperforming in MT tests.

How much power did it use to get that 3,227 score though? Did they significantly overclock it well beyond its sweet spot to get that score?

On my M1 Pro, running GB6 ST uses 0.3w - 5w of power. Likely similar for M3 Pro.

 

[eek2121]

Why do you believe it won't cost as much? Have they announced ANY information about pricing? What about Qualcomm's overall margins and general greediness makes you believe they will price the systems attractively? If they will have better power/performance ratios than Intel alternatives why won't they be priced at a premium to the competition?

Maybe they won't be priced as high as Apple, we'll see, but they are competing much more with Intel/AMD than with Apple so that price comparison is more relevant.

It’s qualcomm. You think Apple is expensive? Just wait.

Qualcomm’s strategy is this: You will buy our high margin SoC + modem, license our obscene patents (including modem, even if you don’t need one) if you buy from someone else, or buy nothing. If you don’t license from us? We litigate.

Qualcomm is both the champion and villain of the ARM world.

If history is any indicator: this chip will be in $1,200+ (or higher) laptops, won’t be competitive with then-current products, and have plenty of issues at launch.

Remember, this chip will actually not be competing with anything you can use right now. It will be competing with Zen 5 and possibly Zen 6, Arrow Lake, M4, etc.

That is one of the issues with hyping your product 6-12 months before launch. The competitive landscape will be very different in H2 2024.

 

[SpudLobby]

How much power did it use to get that 3,227 score though? Did they significantly overclock it well beyond its sweet spot to get that score?

On my M1 Pro, running GB6 ST uses 0.3w - 5w of power. Likely similar for M3 Pro.

First, that’s not overclocking in any classical use of that term, they’re also only shipping one die.

But also, yes, it very likely is that way with the 4.3GHz RE: absurd power. However on total ST power draw they come out ahead of the M2 Max @ 3.7GHz by 30% iso-performance (so at a similar frequency, not 4.3GHz.) I don’t know what that total for the M2 Max is, but given the frequency and power boost over the M1 it’s probably like 10-15W, with QC coming in 30% lower.

Second: dude you are using PowerMetrics and PowerMetrics is not accurate. You need a proper sampling rate over the duration of the test anyways in order to determine the average power consumption.

I’d also like to point out even if it were accurate, that’s not showing you the platform power and in particular DRAM anymore (they removed that from PowerMetrics recently, by the way) or the package power delivery losses etc.


And you want to see all of those things minus idle/statics — that is what Andrei or Geekerwan’s tests at the USB-C or PMIC fuel gauges get vastly closer to effectively measuring. (And before you balk, minimizing DRAM power consumption is also part of how chips in any functional constraint are made more efficient, because additional cache can minimize data movement, including DRAM access which is the most energy expensive — so this is a dependent variable of an SoC that we want to see).

So you don’t have any good information from the PowerMetrics utility on account of inaccuracy — and then insufficient sampling rate for a quick test — and lastly no DRAM and/or active platform power.

You guys have got to stop shooting from the hip, for Christ’s sake. It’s like a new lesson every day. And I’d wager power went up for the M3 Pro over the M1 Pro anyways.

 

[SpudLobby]

Another thing: the Qualcomm chip matches Intel’s 1360P (4+8) 48-50W GB6 MT performance with about 16-18W of *platform* power draw per the graphics QC gave.


So the X Elite runs about a GB6 9-10K at 16-18W.

It’s a pretty good result under the measurement conditions IMO.


https://images.anandtech.com/doci/21105/Snapdragon X Elite Pre-Briefing Deck 10.jpeg

 

[FlameTail]

Let's do a thought experiment.

Qualcomm claimed Oryon core could match the M2 Max ST performance at 30% less power.

Let's take M2 Max (Avalanche @3.7 GHz)'s power consumption as the baseline (100).

Avalanche @3.7GHz : 100

Oryon performance matches the M2 Max performance at 3.8 GHz.

Oryon @3.8GHZ : 70

But the Oryon core can go all the way to 4.3 GHz.
3.8 → 4.3 is a 13% gain.

So,

70 × 113% = 79.1

So,

Oryon @4.3 GHz : 79.1

But... of course this is incorrect! Power consumption does not scale linearly with performance!

So,

Avalanche @3.7 : 100
Oryon @3.8 : 70
Oryon @4.3 : ??

Will Oryon @4.3 exceed the power consumption of Avalanche @3.7 ?

70 → 100 is a 42% gain in power consumption.

So will Oryon's power consumption increase by 42% while performance only increases by 13% ?

We have no other data, so we are in the dark regarding this.

But I am bullish on the power efficiency of Oryon. At the worst, I think @4.3 GHz it will match or exceed Avalanche @3.7 GHz, but not by much.

 

[soresu]

Qualcomm is both the champion and villain of the ARM world.

Their villainy is really only through their modem market monopoly, which isn't necessarily coupled to ARM.

 

[SpudLobby]

It’s qualcomm. You think Apple is expensive? Just wait.

Qualcomm’s strategy is this: You will buy our high margin SoC + modem, license our obscene patents (including modem, even if you don’t need one) if you buy from someone else, or buy nothing. If you don’t license from us? We litigate.

Qualcomm is both the champion and villain of the ARM world.

If history is any indicator: this chip will be in $1,200+ (or higher) laptops, won’t be competitive with then-current products, and have plenty of issues at launch.

Remember, this chip will actually not be competing with anything you can use right now. It will be competing with Zen 5 and possibly Zen 6, Arrow Lake, M4, etc.

That is one of the issues with hyping your product 6-12 months before launch. The competitive landscape will be very different in H2 2024.

Don’t think it will be competing with Zen 6 out of the gate.

And you can barely find Phoenix — which was a let down and gets crushed on battery life or ST efficiency — in the style of laptop this is targeted to 10 months after the announcement.


Zen 5 mobile will be an H2 2024 product for announcement which in practice with AMD means another year if they ever hit the kind of laptops this is targeted to at scale. In Europe and Asia they just narrowly.


Arrow Lake IDK but same deal with paper launches — though Intel when they do launch will hit the market hard and especially for the kind of ultraportable with LPDDR that Qualcomm is after (even in the 45W range there are plenty of those). The problem is doing that on time.

 

[SpudLobby]

There will be probably be a serious Envy/Spectre and XPS equivalent Qualcomm laptop before an AMD one.

AMD kinda has had Envy laptops but not at the volume or seriousness of Intel and there is no 7040 series on those. Just Pavilions

 

[Doug S]

Avalanche @3.7 : 100
Oryon @3.8 : 70
Oryon @4.3 : ??

Will Oryon @4.3 exceed the power consumption of Avalanche @3.7 ?

70 → 100 is a 42% gain in power consumption.

So will Oryon's power consumption increase by 42% while performance only increases by 13% ?

We have no other data, so we are in the dark regarding this.

But I am bullish on the power efficiency of Oryon. At the worst, I think @4.3 GHz it will match or exceed Avalanche @3.7 GHz, but not by much.

Well we've got some data from the graph Qualcomm provided. It "matches competitor performance at 68% less power" i.e. vs i7-1360p performance. Assuming the y axis is linear (it is unmarked so you can't know for sure) it appears that at max power it is about 30% faster than i7-1360p. More than triple the power is required to gain 30% performance.

We can also observe that as expected the curve is flattening beyond that iso performance mark, culminating in the 30% better performance it gets at 3.8 GHz (because it won't run any cores at 4.3 GHz in a multithreaded test) and we know the the curve will continue to further flatten from 3.8 GHz to 4.3 GHz.

If you require more than triple the power to gain 30% in performance, I'd guess that in an even flatter part of the curve (to the right of what is shown) you're gonna be in the ballpark of doubling power to get that extra 13% at 4.3 GHz.

 

[FlameTail]

Can you post the relevant slide?

 

[Doug S]

Can you post the relevant slide?

I was referring to the one SpudLobby posted earlier today. That is MT but the slope of MT power increase is going to be "better" than the slope of ST power increase (because of shared caches, memory controllers, etc.) so that's a 'best case' slope for what the ST slope would be.

 

[SpudLobby]

I was referring to the one SpudLobby posted earlier today. That is MT but the slope of MT power increase is going to be "better" than the slope of ST power increase (because of shared caches, memory controllers, etc.) so that's a 'best case' slope for what the ST slope would be.

Yeah.

 

[SpudLobby]

I definitely don’t think the 4.3GHz ST will look great. And we know the MT clocks are at best 3.8GHz fwiw. Tops out at 80W platform power there so I mean in practice for ST I could see the power draw at 9-12W or something at 3.8GHz. Which isn’t too bad, the perf on GB6 is around 2770 or so at that point? I think AMD still draws more for similar.

Apple M2 Review

The new Apple M2 is being touted as a much faster SoC than its competitors, but we'll have to see that for ourselves as we test it...

www.techspot.com

Here’s a review that took power from the wall minus idle (so display mostly and statics removed). M2 ST is about 8W.

You also notice MT power — and we saw this in the Mac Mini Andrei review too — is a bit higher than Apple’s claims probably because they do roughly Intel/AMD on their power figures when they say the M2 is 15W at top (or M3, 17W) — they are only referring to the CPU or SoC.

Also see the ST is a bit higher from even the AMD and Intel stuff. I’m not 100% sure what Phoenix looks like but I’d wager it peaks as badly as Rembrandt just with more performance.

 

[SpudLobby]

Fwiw: AMD’s Zen 4 7840U even at best on GB6 ST with 5.1GHz is running about a 2600 +-40 GB6.

That’s still about 150-200 points behind QC @ 3.8GHz from the live demo, and yes result on Windows.

I suspect the Snapdragon’s 3.8GHz perf/watt is closer to Apple (see the base M2 around 8 watts) than it is AMD at 5GHz there who is likely running total power into 20W+ give or take some, and not matching QC.

 

[FlameTail]

So Oryon needs to improve it's IPC in future generations without cranking up the frequency even more.

Mr. Williams and Mr. Bruno if you are reading this,

WE WANT MOAR IPC!!

 

[eek2121]

Don’t think it will be competing with Zen 6 out of the gate.

And you can barely find Phoenix — which was a let down and gets crushed on battery life or ST efficiency — in the style of laptop this is targeted to 10 months after the announcement.


Zen 5 mobile will be an H2 2024 product for announcement which in practice with AMD means another year if they ever hit the kind of laptops this is targeted to at scale. In Europe and Asia they just narrowly.


Arrow Lake IDK but same deal with paper launches — though Intel when they do launch will hit the market hard and especially for the kind of ultraportable with LPDDR that Qualcomm is after (even in the 45W range there are plenty of those). The problem is doing that on time.

Zen 5 is coming earlier than you think, and these next gen mobile chips for laptops are coming later than you think.

Even meteor lake, while not setting speed records, is poised to have competitive perf/watt. Arrow Lake is a step above that.

Mobile Zen 5 is going to be something else…

Look, I am not bashing on ARM/Qualcomm/anyone else, but there is a reason Intel and AMD dominate the desktop, server, and mobile segments, and a margin loving company like Qualcomm isn’t going to change that any time soon.

…they are STILL including the modem in their laptop parts, do you think they will give it away for free? lol no.

Fwiw: AMD’s Zen 4 7840U even at best on GB6 ST with 5.1GHz is running about a 2600 +-40 GB6.

That’s still about 150-200 points behind QC @ 3.8GHz from the live demo, and yes result on Windows.

I suspect the Snapdragon’s 3.8GHz perf/watt is closer to Apple (see the base M2 around 8 watts) than it is AMD at 5GHz there who is likely running total power into 20W+ give or take some, and not matching QC.

Have you ever heard people say megahertz doesn’t matter? This is why. It is possible to build a 1ghz chip that runs just as fast. AMD and Intel justdesign their chips around a very specific performance/power/cost/marketing profile.

Intel has always chased high clock frequencies to the point of absurdity, and AMD has followed just to compete. A consumer sees 6ghz and automatically assumes it is the faster chip.

Back in the Athlon days, Athlon chips had significantly higher IPC than Intel chips. They punched well above their weight. A 1.8-2.2ghz Athlon could beat a 3-4ghz Pentium 4. AMD can (lol: is going to soon*) easily produce a high PPC, low clocked, low power part, but very few on the market actually care for or want that, and NOBODY that does any serious work with their PCs wants WoA in it’s current state.

What will it take to change the PC market? An approach that is something other than “me too at twice the price” or “me too at the same price”.

* without the low clocks.

 

[FlameTail]

…they are STILL including the modem in their laptop parts, do you think they will give it away for free? lol no.

The modem is not integrated into the Snapdragon X Elite SoC. It's a discrete part. Qualcomm mentions that OEMs have the flexibility to include it or not.

 

[SpudLobby]

…they are STILL including the modem in their laptop parts, do you think they will give it away for free? lol no.

No they are not. It’s not on the die and it’s optional. You’re in for a ride.

Have you ever heard people say megahertz doesn’t matter?

Buddy if you think opposing MHZ wars helps you I have awful news for you. We’re talking frequencies because we care about performance and AMD and Intel need higher clocks for the same performance. Asking where the frequency/power points on a core’s curve lie is just standard.