Nuvia: don’t hold your breath

[dada_dave]

Qualcomm wins a legal battle over Arm chip licensing

The jury delivered a win for Qualcomm with a split verdict.

www.theverge.com

Qualcomm won. But jury couldn’t reach a verdict on Nuvia. So I dunno.

I suppose the verdict isn't hugely surprising, especially the part where the jury couldn't decide. The behavior of both companies seemed to be at odds with common sense and at times even their own interests. Some of ARM's demands and positions seemed unreasonable, but then so did Qualcomm's. As a layman, I just didn't understand ARM's trial strategy here and as lawyer you didn't seem to either. With such a verdict will there likely be retrials or appeals?

As a sidebar: I felt the whole argument over how much silicon/RTL was reliant on ARM technologies to be an odd issue to fight over in the first place. It's an ARM CPU core. Many of those other functional blocks in the 99% are great for performance, but you can still build a CPU without them technically. So even if one agreed that ARM's tech was restricted to just the decoders and that you could swap those ARM decoders out for any other decoder based on any other ISA - like say x86 (ha!) - that the CPU would continue to function exactly as it did before, the CPU cannot function without decoders. And as you said, to get access to those that means an ARM license. The relevant question to me was therefore just whether Nuvia breached its agreement and whether Qualcomm had the right to develop ARM cores under its license - which for the latter yeah it clearly does, so really I guess it was only the first question that was relevant and finding in favor of Qualcomm on the latter seems more than reasonable. That the jury couldn't decide on the former also seems reasonable given from what I could tell given the seemingly confusing Nuvia ALA but that had little to do with what percentage of the silicon/RTL was based on the ISA or ARM ARM. Because it isn't 0%. So fighting over that is just a distraction. Reading back over my own paragraph, I guess that's a long winded version of your snooty waffle argument.

As a sidebar to the sidebar: I was amused that the Qualcomm CEO said they'd be happy to go back to paying for ARM's off-the-shelf designs if they were competitive.

Cristiano explained that Arm’s lagging behind Apple in performance was the main reason for buying Nuvia, and he would be happy to get TLA from Arm if it provided high-performance cores even now.

Wish granted!

Okay sure the Oryon-L is slightly better, but the ARM X925 core is definitely competitive! Think he's going back to paying the TLA for ARM cores? Nah me neither ...

 

[dada_dave]

ARM going for a retrial rather than appeal:

Arm will seek a new trial against Qualcomm

Only one of the three original charges can be relitigated.

www.tomshardware.com

 

[Cmaier]

ARM going for a retrial rather than appeal:

Arm will seek a new trial against Qualcomm

Only one of the three original charges can be relitigated.

www.tomshardware.com

for sure. there was no final judgment on half the case, so I’m not sure what you could even appeal

 

[mr_roboto]

for sure. there was no final judgment on half the case, so I’m not sure what you could even appeal

Huh. How often does that happen? I naively thought the job of the court was to produce a final judgment, not kick the can down the road.

 

[Cmaier]

Huh. How often does that happen? I naively thought the job of the court was to produce a final judgment, not kick the can down the road.

well, if the jury can’t reach a verdict, then the jury can’t reach a verdict. You urge them to go back and try again, but, in the end, you can’t force them to reach a verdict. It’s fairly rare in civil cases, though.

 

[dada_dave]

for sure. there was no final judgment on half the case, so I’m not sure what you could even appeal

I guess my question was what the rules were wrt appeals vs retrial on the part the jury did reach a verdict on. Is it that if a jury fails to reach a verdict on any part you can retry all of it? I would have naively thought that you would have to appeal the decision that was reached while retrying the original question that wasn’t, but maybe that would be too complicated in practice?

 

[Cmaier]

I guess my question was what the rules were wrt appeals vs retrial on the part the jury did reach a verdict on. Is it that if a jury fails to reach a verdict on any part you can retry all of it? I would have naively thought that you would have to appeal the decision that was reached while retrying the original question that wasn’t, but maybe that would be too complicated in practice?

I’ll wait and see the papers, but it seems to me the two questions are pretty interwoven. The court may have been willing to override the jury if the jury said Nuvia was liable but Qualcomm wasn’t, for example.

 

[The Flame]

ARM going for a retrial rather than appeal:

Arm will seek a new trial against Qualcomm

Only one of the three original charges can be relitigated.

www.tomshardware.com

Apparently, the Judge isn't keen on having a retrial. She wants mediation first.

Qualcomm vs. Arm trial, Day 5 – Qualcomm wins! – Tantra Analyst

www.tantraanalyst.com

 

[Cmaier]

Apparently, the Judge isn't keen on having a retrial. She wants mediation first.

Qualcomm vs. Arm trial, Day 5 – Qualcomm wins! – Tantra Analyst

www.tantraanalyst.com

yeah, any federal judge will use every opportunity to push ADR. It’s plausible that it settles at mediation, since Arm now has seen they have an uphill battle. But I think it more likely that it doesn’t settle. Note that a mediation takes one day, and will likely occur 2 or 3 months from now (based on typical high-end mediator availability). Any point after that they can rev the case back up and retry it. But the trial likely wouldn’t occur for quite awhile, since free trial dates are scarce in Delaware.

 

[dada_dave]

Wish granted!

View attachment 33210

Okay sure the Oryon-L is slightly better, but the ARM X925 core is definitely competitive! Think he's going back to paying the TLA for ARM cores? Nah me neither ...

Interestingly this seems to be a situation where GB 6.3 and SPEC disagree. Above the X925 and Oryon-L have very similar numbers but in GB, the Oryon-L substantially outperforms the X925.

Dimensity 9500 specifications leak with new CPU architecture, notable frequency uplift, and promise of big performance gain

Specifications of the MediaTek Dimensity 9500 have leaked online. The SoC reportedly features a new "Travis + Gelas" CPU architecture with eight cores. The leak also alleges a big clock speed boost for the SoC with the Dimensity 9500's two big cores alleged to reach up to 4 GHz.

www.notebookcheck.net

(Ignoring predictions of the upcoming Dimensity chip and focusing on current gen it’s 14% slower in GB - of course not sure at what power level the Dimensity was tested at)

 

[The Flame]

Unreal Engine 5 Nanite running on Snapdragon 8 Elite (official demo):



Considering that Nanite is pretty demanding on the hardware, it seems Adreno 8 is a decent GPU arch.

Even Intel's first gen Arc discrete GPU (Alchemist) couldn't run Nanite because it lacked INT64 atomics and Execute Indirect.

 

[The Flame]

An interesting talk about Ray Tracing on Adreno GPUs:



PDF of presentation slides:

https://indico.freedesktop.org/event/6/contributions/302/attachments/225/305/Ray%20Tracing%20for%20Adreno%20GPUs%20on%20Turnip%20%28XDC%202024%29.pdf

8 Gen 2 (Adreno 740) and X Elite (Adreno 741) have 1st generation Adreno RT.

8 Gen 3 (Adreno 750) has 2nd generation Adreno RT. It has a hardware block called AQE that does invocation repacking, which sounds similar to Nvidia's Shader Execution Re-ordering (SER) or Intel's Thread Sorting Unit (TSU).

8 Elite has 3rd generation Adreno RT I guess, but the details are unknown. Looking at the benchmarks though, it seems this is also a minor improvement over 2nd gen Adreno RT.

8 Elite G2 (Adreno 840) will have 4th generation Adreno RT, and it might implement BVH traversal acceleration in hardware.

(1) Apple already introduced this in A17 Pro last year
(2) With Qualcomm's renewed push for PCs, Adreno is being deployed in Snapdragon X SoCs, and advanced RT hardware is going to be important there.

 

[The Flame]

Interestingly this seems to be a situation where GB 6.3 and SPEC disagree. Above the X925 and Oryon-L have very similar numbers but in GB, the Oryon-L substantially outperforms the X925.

Dimensity 9500 specifications leak with new CPU architecture, notable frequency uplift, and promise of big performance gain

Specifications of the MediaTek Dimensity 9500 have leaked online. The SoC reportedly features a new "Travis + Gelas" CPU architecture with eight cores. The leak also alleges a big clock speed boost for the SoC with the Dimensity 9500's two big cores alleged to reach up to 4 GHz.

www.notebookcheck.net

(Ignoring predictions of the upcoming Dimensity chip and focusing on current gen it’s 14% slower in GB - of course not sure at what power level the Dimensity was tested at)

It would be more useful to look at the GB6 Single Core subtest scores.

OnePlus PJZ110 vs OPPO PKB110 - Geekbench

PJZ110 is the device with 8 Elite.
PKB110 is the device with Dimensity 9400.

We can see that 8 Elite has a huge lead in Ray Tracer and Object Remover subtests.

 

[Jimmyjames]

We finally have some good scores from an 8 Elite soc on Geekbench AI. 62017 for quantized score is the highest I've seen.

OnePlus PJZ110 - Geekbench

Benchmark results for an OnePlus PJZ110 with a Qualcomm ARMv8 processor.

browser.geekbench.com

Some of the accuracy scores are bad though. 56% for machine translation.

 

[dada_dave]

It would be more useful to look at the GB6 Single Core subtest scores.

OnePlus PJZ110 vs OPPO PKB110 - Geekbench

PJZ110 is the device with 8 Elite.
PKB110 is the device with Dimensity 9400.

We can see that 8 Elite has a huge lead in Ray Tracer and Object Remover subtests.

Thanks! Given the wide variance in GB results (here the Dimensity and Elite are significantly closer than in NBC's tests), I had been hoping to do the violin plots but was having trouble pulling together lots of different Dimensity processors together since search by processor seems to be broken (at least for Dimensity, other times it works). Maybe just doing it by device is better anyway - more consistent. Not sure when I'll get around to it.

Do you know if Geekerwan publishes its SPEC results anywhere? It would be nice to see if there are any similar subtest outliers in SPEC that simply average out.

I wonder why Ray Tracer and Object Remover are such outliers though? Something to do with the FP performance that doesn't show up as much in SPEC (which has its own Ray Tracing and Image manipulation subtests)? FP is where the Oryon-L pulls ahead slightly even in SPEC17, just not by anywhere near as much as in GB6.

 

[leman]

We finally have some good scores from an 8 Elite soc on Geekbench AI. 62017 for quantized score is the highest I've seen.

OnePlus PJZ110 - Geekbench

Benchmark results for an OnePlus PJZ110 with a Qualcomm ARMv8 processor.

browser.geekbench.com

Some of the accuracy scores are bad though. 56% for machine translation.
View attachment 33319

Yeah, the quantized results are weird. Certainly an impressive showing for some of the image analysis tests, but the low accuracy on other tests is worrying. It looks like Qualcomm is using some pretty aggressive optimizations here which are not always correct. Great performance potential though on low-precision workloads. Excellent half-precision results as well.

 

[dada_dave]

Qualcomm expanding data center/server team, hiring away from Intel.

Qualcomm hires Intel's Xeon architect to lead development of server CPUs

Qualcomm's datacenter effort begins to take shape.

www.tomshardware.com

Presumably they will be fulfilling Nuvia’s original premise of using Orion cores for servers - could be off-the-shelf or custom. I could also have filed this under “The Fall of Intel” thread …

 

[Cmaier]

Qualcomm expanding data center/server team, hiring away from Intel.

Qualcomm hires Intel's Xeon architect to lead development of server CPUs

Qualcomm's datacenter effort begins to take shape.

www.tomshardware.com

Presumably they will be fulfilling Nuvia’s original premise of using Orion cores for servers - could be off-the-shelf or custom. I could also have filed this under “The Fall of Intel” thread …

he worked on itanium. what could go wrong?

 

[dada_dave]

Qualcomm Project Glymur surfaces in new leaks with possible Snapdragon X2 Ultra Premium also spotted

Two of Qualcomm's next-generation Snapdragon X series chipsets have appeared in early leaks. With Qualcomm having already teased that Oryon V3-based will arrive this year, it now seems that something called the Snapdragon X2 Ultra Premium could also be in the pipeline.

www.notebookcheck.net

I’m a little tired so I’m not sure how to read the above. Previously it’s been reported that Qualcomm was going to release two chips this time around. It’s not clear to me if this “Ultra Premium” chip represents a third one or a possible name of the second, presumably bigger one.

 

[Yoused]

Here is an image that seems to suggest three versions of the SoC

linked site

It looks from the chart as though Qualcomm may be going big.little in some of the second gen. The linked site is difficult to read, as the text is mostly Hangul.