Intel Arc - discrete GPUs get a name and date - TSMC confirmed!

[porina]

 

Summary

Intel have now revealed the name of their discrete GPU offerings, and will come under Intel ARC branding. They also give a window of Q1 2022 so they should be on the shelves no later than March next year. Feature wise they support DX12 Ultimate, so in that sense should be competitive against AMD and nvidia offerings. Actual price / performance will remain to be seen closer to the time.

 

Under the Arc umbrella we have 4 generational names in the pipeline. Starting with Alchemist, follows Battlemage, Celestial and Druid.

 

Quotes

Quote

After several months of various teasers, Intel is finally starting to put the band together for their first high-performance discrete GPUs and video cards. This morning the company is kicking its pre-launch marketing game into high gear by announcing a new brand name that these video cards will sold under: Arc. As well, the company is finally giving us our first real (albeit wide) launch window for the hardware. The first Arc video cards, based on the "Alchemist" generation of hardware, will be released in the first quarter of 2022, kicking off Intel’s formal foray into high-performance discrete consumer graphics for desktop and mobile.

 

My thoughts

Intel have been working on Xe graphics and we knew that included a discrete offering too. The end of the road to a product is in sight, although the Window extending to March next year might be a bit long for those hoping to see quick relief to the current GPU shortages. The codenames used are also interesting. To me, they're kinda generic fantasy/RPG like. Anyone know if there a more specific inspiration for those?

 

Update added 19 Aug

 

Quote

Ultimately, in conjunction with the use of TSMC’s N6 process, Intel is reporting that they’ve improved both their power efficiency (performance-per-watt) and their clockspeeds at a given voltage by 50% compared to Xe-LP. Note that this is the sum total of all of their improvements – process, logic, circuit, and architecture – so it’s not clear how much of this comes from the jump to TSMC N6 from Intel 10SF, and how much comes from other optimizations. But either way, Intel will need everything they can get.

 

Meanwhile, a 50% clockspeed uplift over Xe-LP would put Intel’s clockspeeds at roughly around 2GHz. That would be somewhere in between NVIDIA and AMD, who tend to have slightly lower and higher clockspeeds respectively. Or, to put things in terms of raw throughput, this would give a theoretical 8 slice Alchemist design 16.4 TFLOPS of single precision (FP32) shader/compute throughput, which would be almost 8x the rated throughput of the discrete Xe-LP DG1 video card.

Intel confirms it is using TSMC N6 process. Compared to the well known N7, this is not expected to give more performance but improves density.

 

Anandtech estimate peak FP32 performance to be 16.4 TFLOPS. For context with other GPUs, this is close to a 3060 Ti at 16.2 TFLOPS, sandwiched between 3070 at 20.3 TFLOPS and 3060 at 12.7 TFLOPS. Note FP32 performance does not necessarily scale with gaming performance so it may deviate from this.

 

Sources

https://www.anandtech.com/show/16886/intel-video-cards-get-a-brand-name-arc-coming-q1-2022 

https://www.anandtech.com/show/16895/a-sneak-peek-at-intels-xe-hpg-gpu-architecture

 

[Justaphysicsnerd]

3 minutes ago, porina said:

Starting with Alchemist, follows Battlemage, Celestial and Druid.

seems like someone at Intel is into DnD

[StDragon]

I hope Intel has addressed the FPS micro-stutter that plagued their Iris Xe DG1 as reviewed from GN. Granted, it wasn't marketed for gaming specifically, but a low cost entry as a dGPU.

[TVwazhere]

27 minutes ago, porina said:

They also give a window of Q1 2022 so they should be on the shelves no later than March next year.

I'm usually optimistic, but I have a feeling this is going to be more paper-launch ish given the current state of chip manufacturing globally. 

[Guest]

Can't wait for it to be released. Would be fun to see 3 competitors. EZ win for the consumer.

Just gotta pray to God that the cards will be in stock.

[porina]

Looking at the article in more detail, there is also comment that Intel is looking at "artificial intelligence-driven super sampling" which could be a competitor to nvidia's DLSS.

 

There's still many unknowns.

 

Which fab will they get made at? Intel's own, or TSMC or elsewhere? Xe is already made on 10SF as part of Tiger Lake CPUs so I don't think it would be any worse than that. Technically it also exists on 14nm as part of Rocket Lake, but it seems extremely unlikely for 14nm to be used for any high volume performance parts looking forward. Given the timescale if it were Intel internal, I'd think it would be no worse than 10SF and a fair chance it will be on 7. Especially given that Alder Lake will be on 7 and also includes Xe graphics.

 

Ignoring power and looking at performance, that could be estimated factoring in two major assumptions. Is it the same as current Xe or will there be some generational improvement? Current Xe doesn't do RT does it? The iGPU graphical power doesn't make sense to do that. So is it kinda like 16 vs 20 series nvidia, or a bigger jump between them? Secondly, how big could Intel go on a GPU? Someone more motivated than I could look at Tiger Lake die shots, work out the Xe part, and then pick out the scalable parts of it. How big could it go assuming die sizes comparable to the competition?

 

8 minutes ago, TVwazhere said:

I'm usually optimistic, but I have a feeling this is going to be more paper-launch ish given the current state of chip manufacturing globally. 

As I question above, who's fabs will it be made on? If they want to make a splash on launch, ensuring supply using internal capacity might be a starting point. Also it is quite far out, potentially over 7 months away. Suppose if we work the other way, what needs to be done when to launch a product? Hardware design should be largely settled and they should be looking at optimising and executing volume manufacturing.

 

Note a 2022Q1 date allows them to get the Alder Lake launch out of the way first.

[Guest]

4 minutes ago, porina said:

As I question above, who's fabs will it be made on?

TSMC 3nm node?

Intel locks down all remaining TSMC 3nm production capacity, boxing out AMD and Apple | TechRadar

 

[8tg]

12 minutes ago, TVwazhere said:

I'm usually optimistic, but I have a feeling this is going to be more paper-launch ish given the current state of chip manufacturing globally. 

The difference between intel and AMD/Nvidia is that intel has their own fabs. They’re having very little in the way of supply chain issues because they’re not as reliant on other companies, and they’re also simply the biggest and were able to quickly snag up any available resources left.

 

I don’t expect to see supply issues limited by production but by vendors selling large quantities to scalpers and miners at first.

[porina]

2 minutes ago, sordinary said:

TSMC 3nm node?

Timing doesn't work out. Arc is expected to launch Q1 next year, with TSMC 3nm not going into volume production until later in that year. Also I'm not sure Intel would put what is unlikely to be leading edge GPUs on what is definitely a leading edge process. Far more valuable if Intel had access to TSMC 3nm they'd use it on something that mattered more like server CPUs.

[igormp]

Oh boy, I have really high expectations for those GPUs on the prosumer market. Intel has most of the software stack already in place to face nvidia head on on a market that AMD has simply forgone.

[Hairless Monkey Boy]

52 minutes ago, porina said:

Under the Arc umbrella we have 4 generational names in the pipeline. Starting with Alchemist, follows Battlemage, Celestial and Druid.

I love this.

[Master Disaster]

Whoever named these GPUs 100% watches LTT/J2C and TechTubers in general and has seen how much shit gets thrown at the terrible Intel product names.

 

One question though, when they say Q1 don't they mean Fiscal Q1? So that would be Apr - Jun, not Jan - Mar, right?

[porina]

2 minutes ago, Master Disaster said:

One question though, when they say Q1 don't they mean Fiscal Q1? So that would be Apr - Jun, not Jan - Mar, right?

I think general product communications are calendar years. Compare to previous if in doubt. Leave the FY talk to financial types.

[pas008]

has it been said if they are using aib or doing it all themselves?

[BuckGup]

12 minutes ago, pas008 said:

has it been said if they are using aib or doing it all themselves?

As of now I've only ever seen Intel made variants of the card but that would be expected during the testing phase

[MageTank]

2 hours ago, Justaphysicsnerd said:

seems like someone at Intel is into DnD

If only my character didn't die last Friday.

 

RIP Iron Thunderclinch. You spammed thunderclap in the face of a vampire lord one too many times.

 

1 hour ago, porina said:

Which fab will they get made at? Intel's own, or TSMC or elsewhere?

This is honestly the million dollar question. What fab are they going to use, and how much fab space are they going to dedicate to producing these? Intel has already shown us they struggle at meeting the demands for their own processors and chipsets on their own fabs. Can they honestly meet the high demands of this particular market if they stick solely to their own fabs for GPU's? If so, do they have fabs capable of matching Samsung/TSMC while offering similar yields?

[AlexGoesHigh]

It basically says nothing about the cards performance but this shows Arc (probably the top card) running some games and a custom demo at the end.

27 minutes ago, pas008 said:

has it been said if they are using aib or doing it all themselves?

The DG1 which only ships in prebuilt are made by the same OEM as the prebuilt, like an ASUS prebuilt with a DG1, they made themselves the card that the GPU sits on, so OEM should supply the graphics cards for Arc.

[Quackers101]

a rumor of intel getting an DLSS developer on their team or contribution for having their own intel DLSS feature? oh oh

[porina]

1 hour ago, MageTank said:

This is honestly the million dollar question. What fab are they going to use, and how much fab space are they going to dedicate to producing these? Intel has already shown us they struggle at meeting the demands for their own processors and chipsets on their own fabs. Can they honestly meet the high demands of this particular market if they stick solely to their own fabs for GPU's? If so, do they have fabs capable of matching Samsung/TSMC while offering similar yields?

I tried to look up the exact statement, and Intel have recently said they are producing more 10nm wafers than 14nm wafers, marking that transition. I think it safe to assume it wont be Intel 14nm. Yield I don't think has ever been publicly stated so that's anyone's guess. My gut feeling is, if Intel make it internally, then it will be on either 10SF or 7. It depends on when they locked the design. Both 10SF and 7 should be performance competitive, with 7 having better power efficiency. I was surprised to see Ice Lake server was still on old 10nm. Presumably they locked that design long ago and couldn't switch it to newer 10nm when they became available. 

 

I'm not up to speed on speculation of availability of other fabs, and if Intel has had history with them. While Intel talked about going external earlier this year, the gap is too short for that to happen unless they had been in negotiation in secret for far longer before the public statements. So my personal guess is this wont be externally made.

 

Quick recap in case people can't keep up with Intel process names:

(older) 10nm as used in Ice Lake mobile and now Ice Lake server. Did not scale well with clock and hits a wall around 4 GHz. Not good.

10SF or SuperFin: Used in Tiger Lake. Unblocks the clock scaling and improves power efficiency.

Intel 7 (formerly known as 10SFE or SuperFin Enhanced): to be used for Alder Lake which will be released before Arc. Expected to bring further power efficiency over 10SF.

[WolframaticAlpha]

3 hours ago, porina said:

-snip-

I love the names. Nvidia and Intel have really good code names in the GPU space. Radeon needs to look at ryzen for their codenames.

 

Raja Koduri has already specified that Intel will support FSR. I am pretty sure they had the hardware ready, but we're waiting for the drivers to be fully written. My concern is that Intel should come up with something akin to a midgen refresh, ie higher performance. Bringing 2021 performance to 2022 won't work.

 

Intel's biggest opportunities are to make power efficient chips with good features. They will also need to polish up Quicksync to compete with NVENC. The biggest threat is the software. While Intel has had a good track record with software, GPU drivers are a whole different beast

 

Nvidia will probably launch the 3050 series and super series in response to it.

 

Intel will probably be using 10nm superfin or 7nm TSMC tiles for the actual GPU. 3nm will probably come to raptor lake first and then to Arc.

 

Intel is pretty strong in AI(vino), so I am expecting AI acceleration to be there, like tensor cores. Hopefully they use it more than Nvidia.

 

Maybe they would implement something akin to the ryzen and Radeon integration.

 

10SF is pretty close to TSMC non enhaced 7nm or samsung 8nm, so node size won't be an issue(fingers crossed).

 

I hope that intel opens up its RT tech and upscaling to other companies. It is the only way they would be able to defeat Nvidia. AMD's ray tracing is pretty much non existent on PC and FSR is not close the DLSS in image quality.

[porina]

18 minutes ago, WolframaticAlpha said:

Raja Koduri has already specified that Intel will support FSR.

Got a reference to that? All I could dig up was a tweet where he said they will look at it. Not quite they will support it.

 

Quote

I am pretty sure they had the hardware ready, but we're waiting for the drivers to be fully written.

I don't get the idea that Intel are somehow unable to make GPU drivers. The company that ships by far the most x86 connected GPUs doesn't know how to make GPU drivers? Now it might be fair to say they lack experience in higher gaming performance GPU drivers, and part of that might not even be on them but the game devs who will lack experience in fine tuning that area from the product not existing before. 

 

Quote

My concern is that Intel should come up with something akin to a midgen refresh, ie higher performance. Bringing 2021 performance to 2022 won't work.

Look at AMD from Polaris through to RDNA1. They were at least a step behind nvidia all the way. Specifically look at the comments AMD made about Polaris. They knew they weren't going to the high end with that, but they were addressing the mass market who buys a lot of devices lower in the performance stack. Intel would enter in a similar situation. How high up they can go remains to be seen, but if they can offer the right cards with price and performance to be competitive, they could shift some volume.

 

Any possible next gen from nvidia and AMD will be end of 2022 unless they shorten their design cycles. Also that would be for top end parts, with lower end ones following much later.

 

Quote

I hope that intel opens up its RT tech and upscaling to other companies. It is the only way they would be able to defeat Nvidia. AMD's ray tracing is pretty much non existent on PC and FSR is not close the DLSS in image quality.

RT support isn't something to be shared. Arc will meet DX12 Ultimate, like AMD and nvidia. The upscaling is more interesting. Will game devs want to bother implementing a 3rd option? While FSR is clearly the weakest technically it has the simplicity advantage. If whatever Intel comes up with works closer to what nvidia does, then I'm sure game devs would want them to share a common interface to make their side easier. This also opens the door to a future version of FSR using more game information to give better results than dumb render output processing.

[WolframaticAlpha]

8 hours ago, porina said:

Got a reference to that? All I could dig up was a tweet where he said they will look at it. Not quite they will support it.

 

I don't get the idea that Intel are somehow unable to make GPU drivers. The company that ships by far the most x86 connected GPUs doesn't know how to make GPU drivers? Now it might be fair to say they lack experience in higher gaming performance GPU drivers, and part of that might not even be on them but the game devs who will lack experience in fine tuning that area from the product not existing before. 

 

Look at AMD from Polaris through to RDNA1. They were at least a step behind nvidia all the way. Specifically look at the comments AMD made about Polaris. They knew they weren't going to the high end with that, but they were addressing the mass market who buys a lot of devices lower in the performance stack. Intel would enter in a similar situation. How high up they can go remains to be seen, but if they can offer the right cards with price and performance to be competitive, they could shift some volume.

 

Any possible next gen from nvidia and AMD will be end of 2022 unless they shorten their design cycles. Also that would be for top end parts, with lower end ones following much later.

 

RT support isn't something to be shared. Arc will meet DX12 Ultimate, like AMD and nvidia. The upscaling is more interesting. Will game devs want to bother implementing a 3rd option? While FSR is clearly the weakest technically it has the simplicity advantage. If whatever Intel comes up with works closer to what nvidia does, then I'm sure game devs would want them to share a common interface to make their side easier. This also opens the door to a future version of FSR using more game information to give better results than dumb render output processing.

Intel will be able to make GPU drivers. They are probably trying to make really good drivers which are optimized for a lot of games. 

 

DX12 Ultimate might be great, but the driver needs to have optimizations for DX11 and Vulkan, since a lot of games are being written for those APIs too.

 

I was hoping that Intel make a flagship card, but they will probably have to start from the midrange. 2022 is a bit late and supply would have caught up with the demand for GPUs by then

[RejZoR]

@porina

It's not like you need specific support for FSR. It's done entirely through DirectCompute on the GPU (I frankly have no idea why AMD was hyping "FSR support drivers" when it's basically a game feature once implemented, not a graphic card feature). It's in the end entirely game feature, not a graphic card feature. It's why it just works on GTX 1060 as shown in presentation.

 

As for everyone saying FSR is just a dumb resampler. Well, it may be, but it may also be more than that. People spotted it's using Lanczos resampling. Which is pretty general purpose image resizing algorithm. There are more advanced ones like S-Spline Max developed by Benvista with few examples here:

https://www.benvista.com/photozoompro/examples

 

Take the Photoshop one as Lanczos basically. Quite dramatic difference. One could also use a neural based approach or even a dumber one that has better edge detection and reconstruction which isn't present in Lanczos. And since you can jerry rig almost any resampling method into it, possibilities are endless. This is just what AMD provided and there is no stopping them from developing their own method that's not as dumb. But even for it being a dumb solution at the moment, it delivers amazing results given how dumb it is and how easy it is to implement.

[leadeater]

5 hours ago, WolframaticAlpha said:

DX12 Ultimate might be great, but the driver needs to have optimizations for DX11 and Vulkan, since a lot of games are being written for those APIs too.

Current Intel drivers are already great with that. DX12 only gets a call out because marketing wise it's more important to mention, because that's what consumers hear and ask about. Not much more to that whole situation.

[Stahlmann]

18 hours ago, TVwazhere said:

I'm usually optimistic, but I have a feeling this is going to be more paper-launch ish given the current state of chip manufacturing globally. 

Important to note that intel is its own fab. Their current gen CPUs were mostly unaffected in availability. (partly also because most people wanted ryzen chips)

 

EDIT:

Lol forget it. This has already been posted here 50 times...