Intel ARC GPUs, Xe Architecture for dGPUs

Why ? I mean, granted it's a new uarch, but they produce drivers for igp for a while now. It's not like it's suddenly new to them.

 

The driver development for Gen 9 and newer has seemed pretty lively this year. It's all unified and the packaging has been changing. Updates seem as often as AMD and NV. I even found that Gen 9.5 Vulkan is working with Dolphin and that was previously broken AFAIK.

 

In my opinion the problem is going to be variability at the supposed performance level. No one really cares if an IGP in a particular game is terrible, so Intel's IGP track-record is mostly an unknown. Intel can get away with poor performance in IGP because the PC enthusiast crowd (forums, "journalists") doesn't really care.



That's over a year old. This is very recent:



Some nice improvements there over 14 months or so. And then there's the Shadow of the Tomb Raider epic fail.

 

DG2/Arc 512 EUs @2.1GHz leak on Sisoftware

 

If the leaks are accurate I'll probably snag a 75W 128EU card if it supports HDMI 2.1 for under $200. Need something to scratch a 4K@120hz itch for game streaming to the living room tv.

 

Intel's Vulkan Linux Driver Adds Experimental Mesh Shader Support For DG2/Alchemist

 

Intel Arc Alchemist engineering sample pictured

 

https://wccftech.com/next-gen-intel-arc-gaming-gpu-to-utilize-multi-chiplet-design/

Next-Gen Intel Arc Gaming GPUs To Utilize Multi-Chiplet Design As Hinted Within Patent

 

That's going to be one complex shader compiler to create position only shaders from the normal ones. Even then it will still have a hell of a lot of redundancy.

I wonder how this interacts with nanite.

 

Can we all remember how Raja was saying the Vega would beat the 1080? I love Raja and he's a great guy to hang with, but let's temper our expectations a little.

If he's saying 3070 I'm guessing it might beat a 3050.

 

Hey now, Vega 64 does trade blows with 1080. It's just that it was kinda sad that it wasn't up with 1080 Ti.
https://www.pcgameshardware.de/Graf...te-Vergleich-Benchmarks-GPU-Historie-1382238/

 

https://videocardz.com/newz/with-49...st-launch-rumors-about-a-delay-keep-piling-up

 

I'm sure every IHV shader compiler is already capable of eliminating code that doesn't contribute to any output, that's a basic optimisation. So if you disable all the other outputs (just as if you commented them out in the source) the compiler will do the right thing, no further complexity required.

 

Intel is porting Linux drivers for Arc to Chrome OS and there is recent work on VFIO (NVIDIA too). Dreaming for a bit, this could mean Chromebooks with Arc dGPUs are coming, which passthrough the dGPU to Borealis (Steam OS) and possibly third party VMs (Windows, maybe an Epic Linux distro?).

 

I am shocked, SHOCKED I say!

I'd even bet a US nickel that the initial launch will be more of a token launch and those cards will cost more than double their MSRP if you can get them.

 

Just speculation but a possibility could be that Intel will focus more on OEM and system integrator sales via in house boards rather than retail DIY with AiB partners.

 

That might be a tougher sale than selling direct to consumers.

OEMs don't need graphics cards for most PC buyers, integrated graphics are good enough. That means for the majority of Intel based systems there's no need for a discrete graphics card leaving only AMD requiring a discrete graphics card for many of their systems. However, with AMD increasing the presence of APUs with integrated graphics, discrete is also being made less relevant for the basic PC buyers (corporate or casual consumers) of AMD systems.

Where discrete is more desirable for OEM builders is gaming PCs. But then you are going into direct competition with NV and AMD who are both well represented in that space. And with the release of the 6500 XT and 3050, both NV and AMD cover a lot of space in that area from budget casual gaming machines all the way up to enthusiast gaming machines.

At that point, Intel would see itself in the same position as it would if it was selling to consumers. However, OEMs are less likely to buy product that isn't already established, well supported, and with a good industry reputation (look at how slow uptake of AMD's Opteron was despite it being superior to Intel CPUs at the time, or how Ryzen didn't really take off until Ryzen 2).

While Intel's CPUs are well proven, there's no track record for their GPUs. That means that OEMs won't know what levels of support costs Intel discrete GPU products might incur. While corporate and casual PC buyers don't really care about GPU driver support in games and thus aren't going to generate a lot of support tickets if an integrated Intel GPU doesn't work correctly in X game, that changes when you have buyers that are buying the PC with the intention of running games on it.

Basically, consumers are far more willing to take a chance on something new than OEMs are.

The only real leverage that Intel would have with OEMs for a discrete GPU would be if they sold it at a massive discount compared to an equally performing GPU from either AMD or NV. Basically Intel would likely have no choice but to sell it at a significant loss if they were to attempt to get an OEM to adopt it in a significant way.

Basically, OEMs will likely have a small number of products available with Intel discrete GPUs in order to test the waters, but are unlikely to want to do more than that unless there was a significant incentive for them to adopt it more widely.

Regards,
SB

 

Intel's come out and stated Notebooks Q1 (essentially means OEMs) with desktop following in Q2 -

I don't seem OEMs being a tougher sell, not for Intel. Intel has very strong OEM relations, way more so than AMD (especially at the point you are referencing) and even Nvidia. Intel also has more engineering resources and I'm guessing will help essentially develop the entire platform OEMs/ODMs can then build off of while also providing a lot of engineering support. This was even an issue that AMD themselves admitted they lacked in when trying to pull back into this market with Ryzen and the slow adoption. Conversely Intel likely doesn't have the same relation with the AiB partners for GPUs and ecosystem there that AMD and Nvidia enjoy. They are not going to have Intel specific focused AiBs for one at this point that both AMD and Nvidia have for one.

Also I think you might underestimate how many GPUs get to market with OEM and system integrator sales vs. DIY retail. I'm not referring to "basic desktop" sales. I've discussed this issue before with respect to why a lot of gamers are actually still able to get GPUs despite the mining induced retail situation. The online enthusiast community is heavily slanted (almost all) towards DIY retail which skews impressions of the overall market. Nvidia for instance was taking flak in these discussion circles for their CES presentation but there's a reason why it typically has more to do with mobile/notebooks and not retail focused on say the desktop cards.

If anything the enthusiast community would be a tougher break through. Just look at discussions everywhere including on here in which how a sizable chunk of people are fervently already entrenched with either AMD or Nvidia in terms of their personnel preference. I'd argue the consumer that's more so buying into prebuilt gaming PCs and/or laptops are going to be more brand agnostic than the retail DIY crowd. It's also interesting their has been statements from Intel that suggests a part of their marketing strategy will be selling on the "experience" instead of the a more numbers focused (eg. avg fps versus) strategy. The latter is more how DIY retail is marketed to while the former is more so OEMs.

Having a focus on OEMs would also act has Intel's crypto hedge essentially without relying on just coincidentally hardware (lack of) suitability (AMD's cache vs memory bandwidth) or usage restrictions (Nvidia's hash rate limiting).

 

Intel announced that they will totally produce 4 million Alchemist GPUs in 2022 (OEM laptops + AIBs). Even with a conservative 50/50 split (which I believe will be more like 70 laptop / 30 AIB, but let's assume 50/50 for easy calculation), it's only 2 millions cards in the DIY market. Last year, according to JPA, AIBs sold 40 millions cards. Thus 2 millions units is 5% market share. Nothing to see here folk, it's nearly rounding error...