Discussion Qualcomm Snapdragon Thread

[DavidC1]

Score Per GHz:
View attachment 110645

Source

Why is David Huang's results so poor?

The 8 Elite P core is getting 2.12 per GHz, while the E core is getting 1.74. So the E core scores are ok but the P core scores are trash.

The Zen 5 and 5c results are weird too. The 5C is slightly faster per clock. Other results such as Geekerwan and C&C showed something like 5% ahead for regular Zen 5. I can't completely trust Huang's results because of these things.

 

[Hesperax]

Why is David Huang's results so poor?

The 8 Elite P core is getting 2.12 per GHz, while the E core is getting 1.74. So the E core scores are ok but the P core scores are trash.

The Zen 5 and 5c results are weird too. The 5C is slightly faster per clock. Other results such as Geekerwan and C&C showed something like 5% ahead for regular Zen 5. I can't completely trust Huang's results because of these things.

It's in David's disclaimer:



We would need to compare the delta in setup between the reviewers to understand why.

 

[Hesperax]

For Reference, Qualcomm claimed (*) a SPECint rate of 2.47 (10.64/4.3) and SPECfp rate of 4.13 (17.77/4.3) for Snapdragon X Elite. I would assume that Snapdragon 8 Elite should be around that mark with optimal conditions.

* I don't see the frequency listed in the Slides but I assume they used their Top-End SKU with the 4.3 GHz Clock.

 

[FlameTail]

A wild bit of speculation is that Qualcomm has an Oryon-S core in the works.

If L = Large and M = Medium, doesn't that suggest that there will be a "Small" core?

Cortex X ≈ Oryon-L
Cortex A7xx ≈ Oryon-M
Cortex A5xx ≈ Oryon-S

Oryon-S would allow Qualcomm to scale the Oryon CPU all the way down to low end Snapdragon 4 series chips and Snapdragon Wearable chips.

If Oryon-S is an Oryon-M but with less cache and lower clocks, which means Oryon-S is also an OoO core, then the more correct analogy would be this:

Cortex X ≈ Oryon-L
Performance Cortex A7xx ≈ Oryon-M
Efficiency Cortex A7xx ≈ Oryon-S

The performance A725 cores would have more cache and clock higher (~3 GHz), whereas the efficiency A725 cores would have less cache and lower clock speeds (~2 GHz).

Historically, the Cortex A7xx line has always been deployed as performance cores, but the Dimensity 9300/9400 have demonstrated that they can be used as efficiency cores too (and excellent ones at that!).

 

[Hesperax]

Based on 8 Elite, we can construct a hypothetical X Elite 2;

Getting the sum of the added component areas of the hypothetical X Elite 2 SoC, and adding it to the 8 Elite, we get 176 mm².

The X Elite 2 will also have some additional I/O, so that would add about 5 mm². But X Elite 2 will be on N3P, which is 4% denser than the N3E used on 8 Elite. So they cancel each other out.

For ~175 mm² N3P, they can make a great laptop SoC with an efficient and powerful GPU.

If we assume the leak from Tigerick is correct:

Then I doubt Mahua would have a die size around ~175 mm² N3P. It would be too expensive for the bulk of the Windows market.

Qualcomm already had to rebate X Elite quite heavily to hit "mainstream" price point for premium laptops (~$1000). And the X Elite was ~170 mm² of N4 (or N4P?).

I think Mahua would need to be around the same die size as X Elite but scaled for N3P (so around ~150 mm² in N3P) to hit similar price points.

 

[FlameTail]

If we assume the leak from Tigerick is correct:

Then I doubt Mahua would have a die size around ~175 mm² N3P. It would be too expensive for the bulk of the Windows market.

175 mm² N3P would be similar in cost to;

Strix Point = 232 mm² N4P (Monolithic)
Lunar Lake = 140 mm² N3B + 40 mm² N3B (Foveros).

Also Tigerick stated that Mahua is a drect successor to Hamoa, which means it's a premium part (belong to the same segment as Strix Point/Lunar Lake, $1000< laptops).

Qualcomm's solution to the mainstream laptop market ($750-$1000) in this generation is Purwa (belong to the same segment as Kraken Point/ARL-U). Tigerick didn't state anything about a Purwa successor, but I assume one is in development.

X1 -> X2
Purwa -> ???
Hamoa -> Mahua
N/A -> Glymur

Qualcomm already had to rebate X Elite quite heavily to hit "mainstream" price point for premium laptops (~$1000). And the X Elite was ~170 mm² of N4 (or N4P?).

I don't think manufacturing cost is the problem for X Elite. 170 mm² N4P is cheap compared to Intel/AMD parts I mentioned above.

The problem is Qualcomm (and OEMs) seem to have overestimated the X Elite's value. The GPU is subpar compared to the competition, and the Windows-on-ARM ecosystem still has the pains of growing software compatibility.

 

[MS_AT]

Why is David Huang's results so poor?

The 8 Elite P core is getting 2.12 per GHz, while the E core is getting 1.74. So the E core scores are ok but the P core scores are trash.

The Zen 5 and 5c results are weird too. The 5C is slightly faster per clock. Other results such as Geekerwan and C&C showed something like 5% ahead for regular Zen 5. I can't completely trust Huang's results because of these things.

Are those even Huang scores? He is not posting FP results most of the time? and the screenshots are signed with watermark under a different name.

Anyway, as was already said above in this thread, the setup differences matter for SPEC as every person is compiling the benchmark on their own. So with Geekbench you usually care about HW differences [for the same OS] as people are using a precompiled binary, for SPEC you need to be mindful of compiler options as they will alter the results.

 

[Nothingness]

Also the IPC is lower than the X4 interestingly. A bit of a surprise there.

And yet people think Arm Ltd cores are seriously underperforming the competition.

Well, good luck getting X4 to run over 4 GHz. IPC is just part of the equation.

Yeah definitely. But the score of X925 isn't that far behind in SPECint.

 

[DZero]

I suspect that Qualcomm had in mind an S core. Out of Order of course.

 

[Meteor Late]

Perf/watt of Phoenix M doesn't seem so great vs A720, which is already a year old core.

Much different performance points, though, one would have to compare power at the same performance or performance at the same power to get an idea.

 

[Meteor Late]

Qualcomm is just pushing the M core way past its sweetspot, because its whole point is not more power efficiency but merely more MT performance in an area efficient manner.
Apple E core on the other hand, and watching its architecture and area size, there is no way that it's so much slower than Oryon M, it's most likely extremely frequency capped to be very efficient.

 

[Meteor Late]

Don't see much point in an Oryon-S core, keep in mind that ARM is serving many more markets, they need a super tiny core for using that in cheap TVs, Routers and what not. But it was a really bad core even in terms of efficiency, Oryon M is small enough, I see no point going lower.
If they want to do cheaper, just do a half L2 cluster for Oryon M, so for example a CPU with 4 Oryon M would get 4MB L2.

 

[FlameTail]

Much different performance points, though, one would have to compare power at the same performance or performance at the same power to get an idea.

Yup, We have curves from Geekerwan;

Don't see much point in an Oryon-S core, keep in mind that ARM is serving many more markets, they need a super tiny core for using that in cheap TVs, Routers and what not. But it was a really bad core even in terms of efficiency, Oryon M is small enough, I see no point going lower.
If they want to do cheaper, just do a half L2 cluster for Oryon M, so for example a CPU with 4 Oryon M would get 4MB L2.

My speculation for Oryon-S is that it would be same core as Oryon-M, but clocked significantly lower (~2 GHz) and with less L2 cache.

 

[Meteor Late]

Yup, We have curves from Geekerwan;
View attachment 110689View attachment 110690

My speculation for Oryon-S is that it would be same core as Oryon-M, but clocked significantly lower (~2 GHz) and with less L2 cache.

Yeah that would just be a renaming but the overall architecture remains the same.