Discussion Qualcomm Snapdragon Thread

[Magio]

Article about the state of that Tuxedo/Linaro Linux laptop based on X Elite here.

Seems it's finally nearing daily drivable status, tho still prototype stage. Both nice that progress has been ongoing and a woeful indictment of Qualcomm's Linux support that we're just nearing full feature enablement a year after launch...

It sucks because I'm actually excited about X Elite Gen 2 from a hardware POV, but I'm so disillusioned by QC's (lack of) efforts regarding Linux that I can't picture myself buying into the platform even if it delivers on its wildest performance expectations.

 

[mvprod123]

Blender for Windows on ARM with Vulkan support is officially released.

Blender for Windows on Arm — Blender Developers Blog

Past, present and future work to make Blender available for Windows on Arm.

code.blender.org

Some benchmarks for X1-85 GPU

 

[soresu]

Blender for Windows on ARM with Vulkan support is officially released.

Blender for Windows on Arm — Blender Developers Blog

Past, present and future work to make Blender available for Windows on Arm.

code.blender.org

Some benchmarks for X1-85 GPU

I wonder if this is just a modification of the Intel OneAPI/SYCL backend.

 

[CouncilorIrissa]

https://twitter.com/x/status/1955039921888170010

 

[poke01]

samsung SM-S947U - Geekbench

Benchmark results for a samsung SM-S947U with a Qualcomm ARMv8 processor.

browser.geekbench.com

Here is the link

 

[Magio]

https://twitter.com/x/status/1955039921888170010

Oh wow, and seemingly it's capped at 4GHz for this score.

4000 ST might actually be achieved once it gets to its 4.7GHz target.

Edit: Actually, wait. This says ARMv8, wasn't 8 Elite 2 speculated to be ARMv9?

 

[mvprod123]

Oh wow, and seemingly it's capped at 4GHz for this score.

4000 ST might actually be achieved once it gets to its 4.7GHz target.

Edit: Actually, wait. This says ARMv8, wasn't 8 Elite 2 speculated to be ARMv9?

It's ARMv9

 

[mvprod123]

https://twitter.com/x/status/1955231604764447087

 

[Hitman928]

https://twitter.com/x/status/1955231604764447087

It’s almost nothing though if you take out the AI/ML stuff

 

[Magio]

View attachment 128648

It's ARMv9

But then that's how it's listed:


I mean, early leaked benchmarks having errors can happen, but it could also be indicative of a faked bench.

It’s almost nothing though if you take out the AI/ML stuff

It's still +11% IPC without SME, that's not nothing. Apple's P-core hasn't had a 10+% YoY bump to IPC since 2019.

Most importantly, this gives 8 Elite Gen 2 roughly the same IPC as A18 Pro, while still (seemingly) targeting high clocks. If it's still as area efficient as the Oryon in 8 Elite Gen 1, that's a monster core in the making.

 

[mvprod123]

But then that's how it's listed:
View attachment 128650

I mean, early leaked benchmarks having errors can happen, but it could also be indicative of a faked bench.



It's still +11% IPC without SME, that's not nothing. Apple's P-core hasn't had a 10+% YoY bump to IPC since 2019.

Most importantly, this gives 8 Elite Gen 2 roughly the same IPC as A18 Pro, while still (seemingly) targeting high clocks. If it's still as area efficient as the Oryon in 8 Elite Gen 1, that's a monster core in the making.

GB still identifies all SoCs as V8. This may be a bug in Geekbench itself.

 

[DZero]

But then that's how it's listed:
View attachment 128650

I mean, early leaked benchmarks having errors can happen, but it could also be indicative of a faked bench.



It's still +11% IPC without SME, that's not nothing. Apple's P-core hasn't had a 10+% YoY bump to IPC since 2019.

Most importantly, this gives 8 Elite Gen 2 roughly the same IPC as A18 Pro, while still (seemingly) targeting high clocks. If it's still as area efficient as the Oryon in 8 Elite Gen 1, that's a monster core in the making.

The issue is that over 4 Ghz, the heat generated would be insane. Let alone 4.5 Ghz and more. That explains why phones are starting to use fans like the Red Magics.

 

[Hitman928]

But then that's how it's listed:
View attachment 128650

I mean, early leaked benchmarks having errors can happen, but it could also be indicative of a faked bench.



It's still +11% IPC without SME, that's not nothing. Apple's P-core hasn't had a 10+% YoY bump to IPC since 2019.

Most importantly, this gives 8 Elite Gen 2 roughly the same IPC as A18 Pro, while still (seemingly) targeting high clocks. If it's still as area efficient as the Oryon in 8 Elite Gen 1, that's a monster core in the making.

I get 1.8% IPC improvement for INT. . .

FP can show a bit more improvement than INT if you want to include the reduced precision tests or not (not typically used when discussing IPC)

 

[Magio]

The issue is that over 4 Ghz, the heat generated would be insane. Let alone 4.5 Ghz and more. That explains why phones are starting to use fans like the Red Magics.

8 Elite Gen 1 drew 7.6W in single thread at its 4.3/4.4GHz clocks. That's 1W more than the A18 Pro P-core (6.6W at 4GHz), but it's not a completely insane draw. This will be on a newer, better process node and refreshed core so one would assume the 4.74GHz target is at similar draw. If so, that's not a problem. (MT is always going to be thermally limited to whatever the phone's cooling can sustain, on the other hand.) Source

I get 1.8% IPC improvement for INT. . .

FP can show a bit more improvement than INT if you want to include the reduced precision tests or not (not typically used when discussing IPC)

1.8% is the difference in INT (excluding object detection) for pure performance, with 8E2 clocked at 4.05GHz and 8E1 at 4.44GHz.

On this bench, 8E2 averages 3210 in those 8 tests, 8E1 averages 3194. But for PPC/IPC you need to adjust for clocks, and if you do that 8E2 is at 792/GHz, while 8E1 is at 719/GHz (again, excluding object detection/SME).

That's a 10+% improvement in INT IPC.

 

[Hitman928]

8 Elite Gen 1 drew 7.6W in single thread at its 4.3/4.4GHz clocks. That's 1W more than the A18 Pro P-core (6.6W at 4GHz), but it's not a completely insane draw. This will be on a newer, better process node and refreshed core so one would assume the 4.74GHz target is at similar draw. If so, that's not a problem. (MT is always going to be thermally limited to whatever the phone's cooling can sustain, on the other hand.) Source



1.8% is the difference in INT (excluding object detection) for pure performance, with 8E2 clocked at 4.05GHz and 8E1 at 4.44GHz.

On this bench, 8E2 averages 3210 in those 8 tests, 8E1 averages 3194. But for PPC/IPC you need to adjust for clocks, and if you do that 8E2 is at 792/GHz, while 8E1 is at 719/GHz (again, excluding object detection/SME).

That's a 10+% improvement in INT IPC.

I missed that the results in the table aren’t ISO frequency as the prior result. With that adjustment I get 10.3% IPC improvement for INT (I don’t agree with taking total score as then you are weighting some tests more than others). It’s a nice improvement, but I don’t know if it’ll be enough to really change anything in their market positioning. It will be more interesting (to me at least) to see if they can actually yield their top end SKU with consistent 1t performance in laptops this time around.

 

[mvprod123]

I missed that the results in the table aren’t ISO frequency as the prior result. With that adjustment I get 10.3% IPC improvement for INT (I don’t agree with taking total score as then you are weighting some tests more than others). It’s a nice improvement, but I don’t know if it’ll be enough to really change anything in their market positioning. It will be more interesting (to me at least) to see if they can actually yield their top end SKU with consistent 1t performance in laptops this time around.

I think some people have greatly overestimated the Nuvia team. There is no magic, and everything comes down to physics.

 

[Magio]

I missed that the results in the table aren’t ISO frequency as the prior result. With that adjustment I get 10.3% IPC improvement for INT (I don’t agree with taking total score as then you are weighting some tests more than others). It’s a nice improvement, but I don’t know if it’ll be enough to really change anything in their market positioning. It will be more interesting (to me at least) to see if they can actually yield their top end SKU with consistent 1t performance in laptops this time around.

This time around X Elite 2 should be the same cores on the same node as 8 Elite 2. That alone should make the laptop chips a hell of a better proposition.

 

[hemedans]

8 Elite Gen 1 drew 7.6W in single thread at its 4.3/4.4GHz clocks. That's 1W more than the A18 Pro P-core (6.6W at 4GHz), but it's not a completely insane draw. This will be on a newer, better process node and refreshed core so one would assume the 4.74GHz target is at similar draw. If so, that's not a problem. (MT is always going to be thermally limited to whatever the phone's cooling can sustain, on the other hand.) Source

Phone can sustain around 5W, 7.6W is good for burst perfomance only, in real life like gaming those core usual run below 2ghz.

 

[Magio]

Phone can sustain around 5W, 7.6W is good for burst perfomance only, in real life like gaming those core usual run below 2ghz.

Gaming is a multi-threaded workload + heavy on the GPU, I'm talking about ST CPU performance.

 

[CouncilorIrissa]

On this bench, 8E2 averages 3210 in those 8 tests, 8E1 averages 3194. But for PPC/IPC you need to adjust for clocks, and if you do that 8E2 is at 792/GHz, while 8E1 is at 719/GHz (again, excluding object detection/SME).

FWIW SME usage isn't limited to Object Detection. Photo Library and Background Blur also leverage SME.

 

[Magio]

FWIW SME usage isn't limited to Object Detection. Photo Library and Background Blur also leverage SME.

Sure but the point is it throws off Object Detection results so much that it unbalances the entire benchmark result (to the point where Geekbench should really rebalance their benchmark already).

If not for Object Detection there wouldn't be a point in excluding SME in the first place, since it does help performance in a variety of scenarios and benchmarks exist to quantify that. But its impact on that on specific test is too disproportionate in the current Geekbench version.

 

[CouncilorIrissa]

Sure but the point is it throws off Object Detection results so much that it unbalances the entire benchmark result (to the point where Geekbench should really rebalance their benchmark already).

If not for Object Detection there wouldn't be a point in excluding SME in the first place, since it does help performance in a variety of scenarios and benchmarks exist to quantify that. But its impact on that on specific test is too disproportionate in the current Geekbench version.

True, just pointing that out since the post only focuses on OD for some reason. Photo Library probably would be in the negatives as well without SME with a clock disadvantage. Still, that would be a 8-9% INT PPC increase which is decent YoY gain.

This time around X Elite 2 should be the same cores on the same node as 8 Elite 2. That alone should make the laptop chips a hell of a better proposition.

I personally won't be surprised if laptops get Oryon-L on N3E, except juiced even higher.

 

[Magio]

I personally won't be surprised if laptops get Oryon-L on N3E, except juiced even higher.

I hope that's not the case and it gets all the 8 Elite 2 arch improvements + N3P. Would be pretty disappointing if the laptop platform continued to lag behind the phone chips.

 

[Doug S]

Sure but the point is it throws off Object Detection results so much that it unbalances the entire benchmark result (to the point where Geekbench should really rebalance their benchmark already).

If not for Object Detection there wouldn't be a point in excluding SME in the first place, since it does help performance in a variety of scenarios and benchmarks exist to quantify that. But its impact on that on specific test is too disproportionate in the current Geekbench version.

It is for exactly this reason that SPEC uses a geometric mean rather than arithmetic mean. In SPEC's case because the vendors were writing the compilers to "break" their benchmarks, had they not used geometric mean we'd have seen overall results go up by WAY more from when Sun & Intel broke benchmarks than from object detection's SME gains.