intel Intel Arc Pro GPUs Announced

[BachChain]

Summary

In what could either be interpreted as confidence or desperation for their, uhh, controversial thus far Arc GPU lineup, Intel is forging ahead with the official announcement of a number of Pro-oriented SKUs. The announcement includes the mobile A30M, single-slot A40, and dual-slot A50. Features include ray-tracing and AI acceleration. However, given that the current top model is limited to 6GB of ram, a 4.0 x 8 PCIe interface, and a 75w peak TDP, one should probably temper their performance expectations. No specific dates have been given for general availability beyond "later this year".

 

Quotes

Quote

All the newly announced GPUs feature built-in ray tracing hardware, machine learning capabilities and industry-first AV1 hardware encoding acceleration. Google’s royalty-free, open source alternative to HEVC, AV1 hasn’t gained a lot of traction on the web so far despite promises from Netflix and YouTube, with its main use being in Google’s Duo video calling despite beating HEVC for compression quality. It’s always been very slow to encode, however, so a good hardware accelerator and Intel’s backing could see it take off.

Quote

We do have to wonder about the drivers situation with Arc Pro. There have been numerous documented problems with the Arc A380 cards, and Intel has already released several driver updates — which can ultimately get Windows into a state where the drivers won't even install, according to Gamers' Nexus. Other aspects of the drivers, like Smooth Sync, are broken. But perhaps that's because Intel has been focusing its driver efforts on the professional side of things? We hope that's the case, because as bad as the consumer drivers are, professionals will be far less willing to deal with broken support.

 

My thoughts

Well, at the very least, it seems like they haven't decided to can the whole thing just yet. Hopefully somebody finds enough of a use for these things that Intel continues the product line.

 

Sources

https://www.tomshardware.com/news/intel-unveils-arc-pro-siggraph-2022

https://www.intel.com/content/www/us/en/products/docs/arc-discrete-graphics/a-series/workstation/overview.html

[Quackers101]

I think their videos did hint at this, or AI to using AV1 on their GPUs to some other workloads.

Also a lot of fun stuff this siggraph, many delicious papers.

[ToboRobot]

vGPU support?  As Luke mentioned on the WAN show, vGPU support will sell these...

[porina]

Aww, was hoping they're finally selling the gaming models in the West. Pro/enterprise uses are easier than consumer since they're more focused than gaming. 

[IkeaGnome]

3 minutes ago, porina said:

Aww, was hoping they're finally selling the gaming models in the West. Pro/enterprise uses are easier than consumer since they're more focused than gaming. 

I'm wondering if this has to do with hardware and driver issues on the gaming lineup. What are the chances that the issues that (if I remember correctly) are said to be hardware and can't be fixed with drivers is a non problem with pro/enterprise uses. Push that line up forward, recoup some money and try again with knowledge of what went wrong in the consumer hardware last time.

[ToboRobot]

1 minute ago, IkeaGnome said:

I'm wondering if this has to do with hardware and driver issues on the gaming lineup. What are the chances that the issues that (if I remember correctly) are said to be hardware and can't be fixed with drivers is a non problem with pro/enterprise uses. Push that line up forward, recoup some money and try again with knowledge of what went wrong in the consumer hardware last time.

Enterprise customers aren't going to be happy with imperfect drivers or hardware with faults that aren't fixable.  Unless there are enterprise features that are enabled...

[IkeaGnome]

18 minutes ago, ToboRobot said:

Enterprise customers aren't going to be happy with imperfect drivers or hardware with faults that aren't fixable.  Unless there are enterprise features that are enabled...

For sure. I'm just wondering if the hardware issues are with something not related to enterprise.

[ToboRobot]

12 minutes ago, IkeaGnome said:

For sure. I'm just wondering if the hardware issues are with something not related to enterprise.

What core function of a gaming GPU isn't needed fore the enterprise models?

[IkeaGnome]

27 minutes ago, ToboRobot said:

What core function of a gaming GPU isn't needed fore the enterprise models?

Not sure. That's why I threw it out there as an open ended idea.

[porina]

1 hour ago, IkeaGnome said:

I'm wondering if this has to do with hardware and driver issues on the gaming lineup.

What "hardware" issues? The problems on the gaming side seem to be that they've simply not baked in all the game specific optimisations that red and green have been doing since forever. Totally underestimated the work needed to do that. It does seem like they have been working with professional software for longer, so that area is more developed.

 

As for launching it, Intel know they can't hold forever and need market experience. They have said in essence they're going to price it for the typical worst case game performance vs competition, so perf/$ could be a lot better vs competition for the titles they have optimised.

 

I'm still waiting to see what I can spend my discount voucher on from the earlier competition they ran. It was supposed to be $100 off and if that is allowed for the low end models I think it would be fun to play with.

[leadeater]

3 hours ago, IkeaGnome said:

I'm wondering if this has to do with hardware and driver issues on the gaming lineup. What are the chances that the issues that (if I remember correctly) are said to be hardware and can't be fixed with drivers is a non problem with pro/enterprise uses. Push that line up forward, recoup some money and try again with knowledge of what went wrong in the consumer hardware last time.

Enterprise customers are easier to work with and you can do remote support calls and collect logs in real time etc, if Intel needs to fix software issues that are applicable to both it'll get done so much faster via the enterprise channel than the consumer.

 

So there is at least "some" good news here for gamers, issues might get fixed faster.

[IkeaGnome]

1 hour ago, porina said:

What "hardware" issues?

I can't seem to find it again, but I could have sworn at some point I heard someone(I thought it was a GN video) talking about the driver issues also being deeper in the hardware. 

[porina]

11 minutes ago, IkeaGnome said:

I can't seem to find it again, but I could have sworn at some point I heard someone(I thought it was a GN video) talking about the driver issues also being deeper in the hardware. 

My impression is that the performance problems with older games/interfaces seemed to be due to difficulty getting work through the GPU efficiently, needing that driver level optimisation. Newer APIs don't seem to be as much a problem. So it does feel like it is more a driver problem than hardware. Maybe there's an argument for designing future hardware differently to make that easier or less of a problem somehow, although I'd suspect "B" is already well on the way and too late to make radical changes not already baked in, so it could be another gen if they need to react.

[hishnash]

For AV1 to succeed it needs to have very low power hardware decoding solutions that support higher resolutions and colour depth.  The best integrated AV1 decoding in mobile chips right now is limited to 8bit and even there the flagship Exynos 2200 is reported to drop a lot of frames at 4k 60 as the device heats up playing back these videos very fast.  

HEVC on the other hand has many existing hardware decoders that are low power enough to easily push 4k 60 10bit without dropping frames even in a mobile phone. This was clearly a design goal of the HEVC group as the main market for video consumption is mobile so power efficacy is very important. 

[Kisai]

1 hour ago, hishnash said:

For AV1 to succeed it needs to have very low power hardware decoding solutions that support higher resolutions and colour depth.  The best integrated AV1 decoding in mobile chips right now is limited to 8bit and even there the flagship Exynos 2200 is reported to drop a lot of frames at 4k 60 as the device heats up playing back these videos very fast.  

For AV1 to succeed, platforms like Netflix, Amazon (both Prime VOD and Twitch), Youtube/Stadia, have to push the issue. No hardware decoder, no access.

https://www.nexttv.com/news/did-roku-quietly-capitulate-on-av1-codec-support-with-its-october-streaming-stick-4k-release

 

And I'm not saying, youtube should suddenly stop using h.264. What I'm saying is that h.264 streams get capped to 480p (where performance parity is hit for both AV1 and H.264 on similar hardware in software.) To get any higher quality streams in H264, the user must have youtube premium, or use the more efficient AV1 codec. 

 

1 hour ago, hishnash said:

 

HEVC on the other hand has many existing hardware decoders that are low power enough to easily push 4k 60 10bit without dropping frames even in a mobile phone. This was clearly a design goal of the HEVC group as the main market for video consumption is mobile so power efficacy is very important. 

And yet, we're going to have to wait until 2033 at the MINIMUM for any patents to have been clear and expired on HEVC, and unfortunately the HEVC licensers are kinda head-up-their-butt. They want to be paid for every encoder, decoder, video, etc implementation of the codec. So in "hardware" that works, because it's one-and-done kind of license payment. The problem is platforms do not want to pay for every video transferred in h265 format.  Which is something that was attempted with h264 and that resulted in the content creators going with VP8 (which was previously used by flash, and part of the reason why the flash player was "the video player" for so long.)

https://web.archive.org/web/20150723075148/http://hevcadvance.com/pdf/RoyaltyRatesSummary.pdf

One of the patent pools wants 0.5% of ALL your revenue. Which funny, enough, was changed to

"and a waiving of royalties on content that is free to end users. " , which is why you see a lot of "videos made with IP holder's content must be free to watch by end users" and "members-only" content on Twitch and Youtube, if it were using HEVC would simply not be a thing. 

 

This is one of the strange issues that affects only one type of content (video) and and we'd have all been find sticking with H.262 (MPEG-2) if it were not for the desire to fit more content into the same space, and later HDR content which h.262 and h.263 and h.264 don't cover. It's not like there is any really innovative here at all. They're just taking advantage of things we've tried years ago and found the performance wall was too steep to implement at that time. 

 

In an ideal situation, there would be no patents on any video or audio compression codec's themselves. Any patent licensing would be applied directly to the encoder side of the equation. eg, If I'm youtube or twitch and transcoding on the fly, the patent licensing should apply to the actual real-time encoding of streams. This is at least, something that could be tracked. Trying to track content already encoded, or being decoded, is impossible to track except in hardware decoders alone. Content should not be paying royalties twice (eg encoding it once in Davinci/Adobe products, and then youtube encoding it again on the fly.)  We will never see such a situation because, patent holders on video have their heads too far up their butts to see that nobody will pay that blackmail to use it.

 

AV1 is basically the end-run around that head-up-the-butt mindset of the patent pools that make nothing.

 

[hishnash]

2 hours ago, Kisai said:

For AV1 to succeed, platforms like Netflix, Amazon (both Prime VOD and Twitch), Youtube/Stadia, have to push the issue. No hardware decoder, no access.

 

So most of these platforms (other than YouTube) are using HEVC for 4k and 4k HDR content.  

Devices like Roku streaming sticks can support 4k AV1 as they do not have the same power concerns as mobile phones.

For AV1 to be accepted by vendors like apple or even to be promoted by other phone vendors the power draw of decoding it needs to not be on pare with HEVC as in the end for most consumers the battery life of the phone is the most important metric, they are willing to use up 10% more mobile bandwidth if that gives them 2h more video playback. 

Maybe there will be a breakthrough in the AV1 hardware decoder space but I would also not be surprised if some companies are intentionally keeping their best silicon teams at arms length from it due to the patent pool possibly leading any tec they put into AV1 decoders/encoders to be used by others (not just for AV1). The last thing you want is your core IP crucial for the rest of your silicon being `infected` by the terms of the AV1 patent pool and thus accessible to everyone. I wander if this leads most silicon vendors to carefully avoid any IP they want to keep for themselves when it comes to building AV1 decoders.

On the licensing front I think the best balance would be to charge for software encoding the same as you charge for hardware encoding, eg a license fee for each install, not something that is per video encoded. 

[ouroesa]

7 hours ago, leadeater said:

Enterprise customers are easier to work with and you can do remote support calls and collect logs in real time etc, if Intel needs to fix software issues that are applicable to both it'll get done so much faster via the enterprise channel than the consumer.

 

So there is at least "some" good news here for gamers, issues might get fixed faster.

This. And they likely sell the hardware as part of a specific system which means there are much less variables (hardware configurations and software compatibility) so they can focus on those specific configurations to ensure very little issues and high degree of optimisation. They dont have to worry about the card wokring with every motherboard, memory config, CPU etc and nearly limitless applications - just the ones the enterprise clients use. If issues crop up, they are able to quickly identify the issue and patch it.  

[PineyCreek]

Well, as much as I hate to see hard work and potential ingenuity be left out in the cold, I'll believe it when I see it, i.e. when they get drivers that work so we can actually see how well it performs.  Maybe they put all their eggs in the Pro basket instead and this will be awesome.

[Cyrus]

Maybe its relevant in this thread but no graphics virtualization for Arc...

Graphics Virtualization Technologies Supported on Each Intel®...

[LAwLz]

22 hours ago, hishnash said:

For AV1 to succeed it needs to have very low power hardware decoding solutions that support higher resolutions and colour depth.

We already got that.

 

22 hours ago, hishnash said:

The best integrated AV1 decoding in mobile chips right now is limited to 8bit and even there the flagship Exynos 2200 is reported to drop a lot of frames at 4k 60 as the device heats up playing back these videos very fast.  

This is false. I happen to have an Exynos 2200 device right in front of me currently playing a 4K 60 FPS AV1 video in hardware. The device is not heating up at all and I have 0 dropped or delayed frames after 5 minutes of playback. I doubt that will change even if I let it go for 50 minutes. It's honestly as cold as when it is idling.

The Exynos 2200 also supports 10 bit color depth in hardware.

 

 

22 hours ago, hishnash said:

HEVC on the other hand has many existing hardware decoders that are low power enough to easily push 4k 60 10bit without dropping frames even in a mobile phone. This was clearly a design goal of the HEVC group as the main market for video consumption is mobile so power efficacy is very important. 

The same can be said for AV1.

It was literally designed in collaboration with hardware makers to optimize it for hardware decoding. Several aspects of the spec changed in order to make it easier and cheaper to implement in hardware. 

 

The "problem" is that AV1 is that some players are dragging their feet with implementing it. Most notably Qualcomm and Apple.

 

 

18 hours ago, hishnash said:

Maybe there will be a breakthrough in the AV1 hardware decoder space but I would also not be surprised if some companies are intentionally keeping their best silicon teams at arms length from it due to the patent pool possibly leading any tec they put into AV1 decoders/encoders to be used by others (not just for AV1). The last thing you want is your core IP crucial for the rest of your silicon being `infected` by the terms of the AV1 patent pool and thus accessible to everyone. I wander if this leads most silicon vendors to carefully avoid any IP they want to keep for themselves when it comes to building AV1 decoders.

This entire section is such a load of rubbish.

There is 0 risk of "silicon being infested by the terms of the AV1 patent tool". You just don't understand what you are talking about.

The patent pool means that you are not allowed to sue anyone using AV1 specific patents. That's it.

 

Even IF what you said was true, then it still would not make any sense because all big players (except Apple and Qualcomm, which both are working on it) would have fallen trap for the "infestation" already.

If you are worried about your CPU architecture being disclosed (which it doesn't) from using AV1 then it doesn't matter if you implement AV1 poorly or not. It would not make any difference even if it worked like that, which it doesn't.

 

Besides, the fact that Intel, AMD, Nvidia, Samsung, Google, MediaTek, Broadcom, Realtek and many more have all implemented it already should give you a pretty good indicator that there is no worry about some chip being "infected" with AV1 licensing. 

 

Your entire post is bullshit. It is quite frankly upsetting to see you post this drivel. 

[hishnash]

3 minutes ago, LAwLz said:

The patent pool means that you are not allowed to sue anyone using AV1 specific patents. That's it.

 

That is what a mean, if a vendor like Qualcomm or Apple make a custom AV1 encoder/decoder the patent pool means others cant be sued for used IP that is used within that decoder/encoder since the patent pool then includes the patents that are released to that encoder/decoder. 
 

5 minutes ago, LAwLz said:

If you are worried about your CPU architecture being disclosed (which it doesn't) from using AV1 then it doesn't matter if you implement AV1 poorly or not. It would not make any difference even if it worked like that, which it doesn't.

It does since you can implement a poor AV1 encoder/decoder by using other IP (maybe off the self ARM controller cores for example... see the vendors that have already implemented AV1). 

 

6 minutes ago, LAwLz said:

Besides, the fact that Intel, AMD, Nvidia, Samsung, Google, MediaTek, Broadcom, Realtek and many more have all implemented it already should give you a pretty good indicator that there is no worry about some chip being "infected" with AV1 licensing. 

It's not the entier chip just the AV1 encoder/decoder that is affected by the licensing pool. The fear of the pool is with respect to using your latest controller and data pipeline IP in the encoder as that will allow anyone else to use it in their encoder/decoders. That is not really a concern for a vendor that just buys off the self ARM core IP or a big GPU vendor as the GPU is not part of the decoder/encoder and any management solution they are using will also be an off the self arm core.  

 

[BuckGup]

Lol why does their marketing sound 100% translated

[LAwLz]

16 minutes ago, hishnash said:

That is what a mean, if a vendor like Qualcomm or Apple make a custom AV1 encoder/decoder the patent pool means others cant be sued for used IP that is used within that decoder/encoder since the patent pool then includes the patents that are released to that encoder/decoder. 

You can't just copy a hardware implementation like that. Even if Apple handed a super high resolution die shot to Samsung or whatever, they wouldn't be able to do anything with it.

But again, pretty much all companies except Apple and Qualcomm (which are both working on it) have already released their hardware designs and it has not been a problem for them.

The issue you are talking about simply does not exist. It's entirely imaginary.

 

 

16 minutes ago, hishnash said:

It does since you can implement a poor AV1 encoder/decoder by using other IP (maybe off the self ARM controller cores for example... see the vendors that have already implemented AV1). 

What do you mean exactly?

Yes you can implement AV1 by using other peoples' designs, but only if you pay them for approval and get access to their designs. For example this is Chips&Media's business plan. They design IP such as the Wave510A which can be used in other companies' chips.

Those designs are however protected by more than just patents. For example they are protected by license agreements and copyright. Samsung can not just take the Wave510A design and implement in their Exynos line-up without paying for it. 

So even IF for example Samsung could "copy Apple's homework" once they release a media engine with AV1 support (again, you can't do this in practice), they would not be allowed to do so legally.

 

 

16 minutes ago, hishnash said:

The fear of the pool is with respect to using your latest controller and data pipeline IP in the encoder as that will allow anyone else to use it in their encoder/decoders.

This is something you have made up. It's simply not true and nobody except you has this fear.

 

I think we have had a similar discussion before and I might have said that a company could copy the design from someone else, but if I recall correctly the context was that:

1) Copying the hardware design would not be possible anyway, because you need a lot of details that are pretty much impossible to get just by analysing the finished hardware.

2) It was a response to you saying someone could get sued for implementing AV1 in hardware.

 

 

The entire point of the AV1 patent pool is that you won't get sued. Nobody will come after Apple for implementing AV1 because if they did then they would be forfeiting their rights to use any and all AV1 patents, which means they could be sued back by everyone in the AOMedia Alliance.

 

But trust me, neither Qualcomm nor Apple have the fears that you are describing. They are not worried that they will get sued for implementing AV1. They are not worried that the patent pool will "infect" other parts of their chips. They are not worried that the power consumption is too high. They are not worried someone will copy their hard work if they design their own implementation.

 

The reason they are dragging their feet are because both are heavily invested in HEVC.

Apple has invested a ton of resources into HEVC for things like HEIC and don't feel any hurry to change formats again.

Qualcomm is one of the lead developers (and one of the companies that gets massive amount of license revenue) from HEVC, so they are not in any hurry to change either.

 

Both of them will change sooner or later though. It's just that they aren't in as big of a hurry as the rest of the industry.

[IkeaGnome]

On 8/8/2022 at 1:39 PM, porina said:

My impression is that the performance problems with older games/interfaces seemed to be due to difficulty getting work through the GPU efficiently, needing that driver level optimisation. Newer APIs don't seem to be as much a problem. So it does feel like it is more a driver problem than hardware. Maybe there's an argument for designing future hardware differently to make that easier or less of a problem somehow, although I'd suspect "B" is already well on the way and too late to make radical changes not already baked in, so it could be another gen if they need to react.

Just to follow up on where I heard this. WAN show citing MLID. Do with that what you will. July 29th episode. Around the 11:30 mark. On mobile or I’d time stamp it for you.