Intel Xe2 Lunar Lake Graphics Performance Disappoints On Linux

[qarium]

to me its really funny that everytime we get REAL news [email protected] had to buy the hardware himself....

well well well... everytime good news would come the companies send him a sample... every time bad news come he has to spend 1400€... for a shit notebook. sorry i have to say.

other reviewers never make any negative press because they never have the resources to buy shit hardware to make a negative review.

[Kano]

I think when Intel is forced to update the processors twice a year then Linux support is very limited.

I would like to see some stable diffusion benchmarks on the same systems (best with some optimization hints) and an additional laptop with latest Nvidia gfx (at least 8 GB VRAM, best 16 GB) on Linux and Windows.

The new thing with the V chips is the stacked gfx and ram… Well does not matter that much it seems especially when you compare it to AMD.

But the price for both are too high for the delivered performance. The NPUs are slow compared to normal GPUs by a huge factor and you pay more or less the same.

[coder]

The NPUs are slow compared to normal GPUs by a huge factor

The biggest advantage of NPUs is in their energy efficiency, which is a big deal for phones and laptops.

[AdrianBc]

Thanks. I knew there were multiple, but I forgot if the other was Denver 2 or what it was called.

I wish there were more information about them. I don't know how we can know just how good they really were (or weren't), but they did stop making them and switched back to licensing ARM's IP.

Was Carmel another VLIW-like core, like Denver?

Yes.

I have an NVIDIA Jetson Xavier board, with 6 Carmel cores, a Volta GPU and a maximum 15 W power consumption.

Because it has a quite different internal architecture than the Arm cores and it uses some kind of translation from the Arm ISA to internal microoperations, the ratio between its performance and that of a mostly equivalent on average Cortex-A75 can vary a lot depending on the benchmark, so for some things Cortex-A75 can be much faster, while for others Carmel can be much faster.

At the time of its launch, Xavier was much faster than any other available Arm-based single-board computer, in good part because it had a wider memory interface, with much higher throughput. For any tasks suitable for a GPU, obviously Volta was much faster than any Arm Mali GPU.

However, Xavier was already overpriced and its successor, Orin, is very overpriced, so Orin is worthwhile only for automotive applications that need expensive certified components and only for those of these applications that have an unavoidable dependency on its relatively big Ampere GPU.

Xavier had little competition, but for Orin there are many alternatives, e.g. a lot of very cheap computers with Intel Alder Lake N/Amston Lake CPUs, or with Cortex-A76 (RK3588 or Raspberry Pi), and starting with this year there are also cheaper SBCs with Cortex-A78 using SoCs from MediaTek or Qualcomm (previously the SBCs with MediaTek or Qualcomm SoCs using big Arm cores had minimum prices not much less than $500, but now there are models with a price around $200 or even down to $100 when a smaller memory is enough; even so, the MediaTek or Qualcomm SoCs are disadvantaged in comparison with the Intel alternatives that have the same price by their poor documentation and software compatibility).

[coder]

Yes.

I have an NVIDIA Jetson Xavier board, with 6 Carmel cores, a Volta GPU and a maximum 15 W power consumption.

Because it has a quite different internal architecture than the Arm cores and it uses some kind of translation from the Arm ISA to internal microoperations, the ratio between its performance and that of a mostly equivalent on average Cortex-A75 can vary a lot depending on the benchmark, so for some things Cortex-A75 can be much faster, while for others Carmel can be much faster.

At the time of its launch, Xavier was much faster than any other available Arm-based single-board computer, in good part because it had a wider memory interface, with much higher throughput. For any tasks suitable for a GPU, obviously Volta was much faster than any Arm Mali GPU.

Awesome! Thanks for sharing!

Xavier had little competition, but for Orin there are many alternatives, e.g. a lot of very cheap computers with Intel Alder Lake N/Amston Lake CPUs, or with Cortex-A76 (RK3588 or Raspberry Pi), and starting with this year there are also cheaper SBCs with Cortex-A78 using SoCs from MediaTek or Qualcomm (previously the SBCs with MediaTek or Qualcomm SoCs using big Arm cores had minimum prices not much less than $500, but now there are models with a price around $200 or even down to $100 when a smaller memory is enough;

So, even Orin NX is not competitive?

https://developer.download.nvidia.com/assets/embedded/secure/jetson/orin_nx/docs/Jetson_Orin_NX_DS-10712-001_v0.5.pdf?G54XHHR7wUMA3PdybPt1jbdES9ElNxoZA_kedcMSZTafTJwrMHk5mZ7g3h-JbZcy8jwKAIEOmbcYgMYL5Ui7YlDlzzt0ayH5nfBla5a-IHcl2L7RnasHLVvw5n1ZSnK7aJsBajqA7RWKUEu0nj4lZTMukcxJ83w37bRpRzLYFOGvfut977RkRS6lqFMbUA==

50 int8 TOPS (dense) is still a lot, and not a level of performance you could get in a SoC until Snapdragon X, Lunar Lake, and Strix Point launched. That only happened this year, yet Orin NX is already like 2.5 years old!

[AdrianBc]

Awesome! Thanks for sharing!


So, even Orin NX is not competitive?

https://developer.download.nvidia.com/assets/embedded/secure/jetson/orin_nx/docs/Jetson_Orin_NX_DS-10712-001_v0.5.pdf?G54XHHR7wUMA3PdybPt1jbdES9ElNxoZA_ked cMSZTafTJwrMHk5mZ7g3h-JbZcy8jwKAIEOmbcYgMYL5Ui7YlDlzzt0ayH5nfBla5a-IHcl2L7RnasHLVvw5n1ZSnK7aJsBajqA7RWKUEu0nj4lZTMukc xJ83w37bRpRzLYFOGvfut977RkRS6lqFMbUA==

50 int8 TOPS (dense) is still a lot, and not a level of performance you could get in a SoC until Snapdragon X, Lunar Lake, and Strix Point launched. That only happened this year, yet Orin NX is already like 2.5 years old!

There are various small computers that use Orin modules, but those are even more expensive than the NVIDIA developer kits.

The cheapest is the Jetson Orin Nano Developer Kit, at $500, but this is extremely crippled, with very low CPU and GPU clock frequencies and small memory. Even the video encoders are disabled. For $500, you can buy a complete computer with AMD Phoenix with much bigger memory, with a CPU that is many times faster and also with a GPU that is much faster and with fast hardware video encoders. The Phoenix GPU has 3/4 of the functional units of the NVIDIA Ampere from Orin Nano, but it has a clock frequency several times higher.

Even at the launch of Orin, in 2022 (Orin has become available for buying much later than its first announcement, and for a long time only the $2000 developer kit has been available), there were many alternatives with better performance per dollar, e.g. small computers with AMD Rembrandt or Intel NUCs with Alder Lake CPUs.

There are a few applications where the high price of Orin may be accepted, if they have an unavoidable dependence on NVIDIA CUDA or on some other software provided by NVIDIA. For such cases, NVIDIA provides excellent technical documentation for Orin, at a level far beyond what is provided by companies like Qualcomm, MediaTek or Rockchip. The level of detail of the NVIDIA Orin documentation is similar to that of the documentation published for microcontrollers by companies like NXP, ST, Infineon or Renesas, which also sell products to the automotive market for which Orin is intended, but all those microcontrollers use much slower and usually much older Arm cores, while Orin has Cortex-A78 cores with a decent performance and a relatively big Ampere GPU. Many of the cheapest smartphones are still using SoCs with Cortex-A78 cores.

[coder]

Even at the launch of Orin, in 2022 (Orin has become available for buying much later than its first announcement, and for a long time only the $2000 developer kit has been available), there were many alternatives with better performance per dollar, e.g. small computers with AMD Rembrandt or Intel NUCs with Alder Lake CPUs.

Better perf/$ I don't dispute. However, it was that 50 TOPS that meant you'd have to pay the premium, if you needed that level of performance.

I have an Intel N97 and I've done some perf/W testing on it, so I do know what they're capable of. Not much, in the way of AI performance, but otherwise good perf/$ and reasonably efficient. Software support is obviously excellent.

while Orin has Cortex-A78 cores with a decent performance and a relatively big Ampere GPU. Many of the cheapest smartphones are still using SoCs with Cortex-A78 cores.

I think a rough rule-of-thumb is that A76 performs somewhat like a Sandybridge laptop, while A78 is fairly close to Gracemont or a Skylake U-series laptop SoC.

[leledumbo]

Looks like they use one driver one card (generation) approach. If I'm not mistaken, AMD driver mostly uses generic code for all cards, while the card specific bits are autogenerated. This results in day 0/day 1 optimum performance.

[coder]

Looks like they use one driver one card (generation) approach.

No. They previously used the i915 for like 1.5 decades' worth of iGPUs. Making a break at Xe (gen 12) makes sense, based on how much they changed in that generation.

The Xe driver is also being used for Xe2. So, even going forward, it looks like it won't be one driver per generation. That would be pretty nuts.

[sophisticles]

What a complete disaster !!

Strengthens my belief Intel will actually be bought by either ARM or Qualcomm. They're the only two CPU designers that could teach Intel after 50 years how to make a chip that is both power efficient AND performant.

And on top of all that bullshit by Pat Gelsinger about how their fabs are going to fab the world's supply of chips and how Intel's "Angstrom" process was so nifty. Now they're trying to spin it off. The banks and the shareholders are already pressuring Intel to sell itself for parts to get that sweet sweet shareholder "return". Perhaps that's because they all know Intel is now incapable of building good chips anymore and making money from that. Better to break up Intel and sell the parts off for quick cash.

I don't know what you're smoking but it must be quite good.

Qualcomm will never get regulatory approval to buy Intel for the same reason that NVIDIA failed to buy ARM.

As for ARM buying Intel, early this year Intel sold 1.5 million shares if ARM that it had, so I don't that happening either, not to mention the regulatory hurdles.

Intel is scheduled to receive 8.5 billion dollars from the U.S. government by years end and if China moves against Taiwan like many are expecting to happen, Intel will be the only game in town for fab services for AMD and NVIDIA.

I have a decent amount of Intel shares and I do not want them to be broken up and bought out, I want them to stay the course and in a few years I will be rich.