Intel Launches First Arc GPUs For Mobile, Details Architecture & Features

After many months of anticipation, today is the day that Intel officially launches its first Arc discrete graphics cards, and also gives us most of the architecture information we’ve been clamoring for. There’s been a fair bit we’ve known of Arc leading up to this point, but now we finally get to understand how everything comes together, and what special abilities the first “Alchemist” cards will bring to our PCs.

The first Arc launch revolves around notebooks, with Intel going for a bottom-up approach with its stack. The Arc 3 series is effectively “Available Now”, although we don’t expect any actual availability for upwards of a month. In early summer, both the Arc 5 and Arc 7 series will round out the mid-range and top-end. Intel hasn’t confirmed a specific date for the first desktop GPUs, but we do know it will be in Q2.

While Arc 3, based on the number alone, may not seem too impressive from a gaming standpoint, Intel does in fact target “enhanced gaming” for the chip. Arc 5 upgrades users to “advanced gaming”, while Arc 7 targets “high-performance gaming”. That being said, the focus isn’t entirely on gaming with Arc. Intel has prepared these GPUs to also be effective at creation, and even AI and deep-learning.

Intel’s Arc Notable Features

Before we get into other details, we feel compelled to highlight at the top of this article a unique feature that Intel’s packing into Arc: AV1 video encode. When Intel announced this feature during a pre-brief, we had to do a double-take. Both AMD and NVIDIA support AV1 decode, but not encode. Intel becomes the first to offer acceleration for both AV1 decode and encode, and we’re excited to see it.

AV1 is primed to become a leading format for streaming video, so it’d be great for the likes of game streamers and Netflix alike. In fact, Netflix is one of the founding companies behind the format, with others including Amazon, Google, Cisco, Mozilla, Microsoft, and – wouldn’t you know it – Intel. Others joined the alliance since then, including AMD, Arm, NVIDIA, Xilinx, and Adobe.

What’s perhaps even more exciting about the inclusion of accelerated AV1 encode is the fact that five solutions are already being prepared to support the feature soon. That includes Adobe’s Premiere Pro, Blackmagic Design’s DaVinci Resolve, HandBrake, XSplit, and FFMPEG, which acts as a backend to many solutions, like OBS Studio, and VLC.

Also on the encode and decode front, Intel fully supports both for VP9, AVC, HEVC, as well as AV1, with up to 8K/60 12-bit decode, and 8K 10-bit HDR encode. Intel’s media engine has been pretty solid for a long while, so we’re interested to see how performance on its higher-performance discrete cards will end up comparing to the competition – especially since we see AMD and NVIDIA perform a fair bit different from one solution to the next.

AV1 accelerated encode is the primary feature Intel’s offering with Arc that its competitors are not, and beyond that, you can expect to see many feature sets shared between them all, including software solutions. Intel’s been working hard on its entire suite of Arc-related utilities (some of which are mentioned later), so we’ll see soon enough if they manage to improve upon the competitor solutions.

On the display front, many specs are pretty much expected, but it is nice to see confirmation that Arc is capable of powering both high-resolution and high-refresh displays. With support for HDMI 2.0b and DisplayPort 1.4a/2.0, Arc can drive dual 8K/60, or quad 4K/120 monitors. Benefiting high-refresh gaming, it can also support 360Hz at 1080 and 1440p.

Multiple frame smoothing features are offered on Arc, as well, including the well-known Adaptive Sync, which syncs the display’s refresh rate. Also available is Speed Sync, which always shows the latest frame with no tearing, as well as a third option, called Smooth Sync, which uses a dithering filter to blur screen tears, and works best with fixed refresh rates.

As you’d probably expect, Arc also supports all of the DirectX 12 Ultimate features, including variable rate shading, mesh shading, sampler feedback, and ray tracing (also supported with Vulkan RT). Intel’s going about its RT capabilities a bit differently than the others, so we’re really interested in seeing how the implementation benefits (or holds back) its performance, and compares to the competition – especially since AMD’s and NVIDIA’s RT solutions currently offer notably different levels of performance.

In addition to all of those advanced features, Intel is also delivering an upscaling feature (similar to AMD’s FSR and NVIDIA’s DLSS) called XeSS. The existence of this feature has been known about for a while, and everything we’ve seen from it so far looks really good. What we haven’t known yet is which games will support the feature, but now we’re finally given some names of the initial batch:

There’s a mix of new and old here, with the classic Shadow of the Tomb Raider included alongside the just-released Ghostwire: Tokyo. Naturally, Intel says that this is just the start, and as with the competition, we’ll see more and more games added to the support list over time.

You can’t release a GPU nowadays without having some of your own software to improve immediate functionality, and that’s what Intel is doing with its new “Control” application. This is going to be a one-stop shop for adjusting performance settings, streaming or recording, or for grabbing the latest drivers. Interestingly, Intel is going to let users enable an option to automatically update drivers without any user input – but you can of course go the manual route.

Users will be able to call upon Control with an Alt + I shortcut (hopefully configurable), while Alt + O will pop up an overlay. The included overlay will include relevant information to the game being played, such as frame rate, clocks, temperatures, fan speeds, and so forth.

Intel’s Arc Products & Performance

As covered earlier, Intel’s initial Arc launch is going to be with the Arc 3 series. These SKUs will include both the A350M and A370M, with the latter including 33% additional cores, a much higher clock speed, and likewise, a higher TDP. You’ll notice that Intel doesn’t give a single TDP value, but instead a range. The company notes that the graphics clock speeds it puts into these spec sheets will reflect what users will see at the lower TDP value, almost making it a worst-case scenario.

Both of the launch Arc 3 models have a 4GB frame buffer, and 64-bit memory bus width. That may seem modest, and it is, but these are the entry-level parts, and given the ultraportable notebooks these will find themselves in, that seems suitable enough.

Once Arc 5 and 7 launch, we’ll see expanded capabilities and potential, with the top-end A770M set to ship with 16GB of memory – which is absolutely huge for a mobile platform. To get anything more than 8GB with a current-gen NVIDIA solution, you’ll be going with the higher-end workstation parts (eg: RTX A4500 or A5500).

When we look over this specs sheet for Arc, one of the things that stands out most is the sheer difference between the number of cores vs. the competition. Each “Xe Core” comprises 16 individual cores (vector engines), so the top-end A770M will sport 512 total. That’s a far cry from the multiple thousands seen in AMD and NVIDIA chips, but ultimately, the resulting performance we see is what will be important.

To help paint a better picture of expected performance out of the launch Arc 3 series, Intel has provided comparisons against its own Iris Xe GPU found in its Core i7-1280P:

Compared to that previous-gen Xe GPU, the gains are not as large as we’d hope to see, but once again, this is the entry-level series, and there will be more to come. Even still, what Intel has shown is that 1080p/60 is going to be possible on Arc 3, and not just with lesser-demanding esports titles. Even in Destiny 2, with Medium settings, we should be able to expect an average of 60 FPS.

One thing that’s nice to see confirmation of from these results is that Arc’s Vulkan performance is good. That API is the only one Wolfenstein: Youngblood uses, and it happens to hit 78 FPS with Medium detail levels. Another great shooter, DOOM Eternal, hits 60+ FPS with High detail levels.

With esports titles, the performance can of course become even more impressive. According to Intel, games like Fortnite, GTA V, Rocket League, and Valorant will be able to hit around 100 FPS fairly easily, as long as detail settings are dialed in correctly.

We covered above that Intel’s XeSS will make its way into a decent collection of games at launch, but the company is working to ensure that many more are officially supported in general – a rather crucial endeavor. The following list includes a handful of games featured in the earlier XeSS slide, but notably has the ultra-popular Elden Ring added, as well as Total War: Warhammer III.

Gaming aside, Intel is also concentrating on creator workloads with this Arc launch, which is good to see, since it’s one of our favorite areas of focus. So far, primary focus has revolved around encoding, rather than heavier workloads like rendering, but that’s to be expected given that we’re dealing with the entry-level Arc parts. Still, what we see from encoding tools looks really good so far:

Compared to the Xe GPU found in the Core i7-12800H, encoding performance with Arc 3 is vastly improved. It’s not only with the transcoding that performance looks good, but also with advanced features like the Auto Reframe and Scene Edit Detect features found in Premiere Pro.

Intel Arc(hitecture)

We’ve covered a number of architecture-related features here already, but now we’ll go into deeper detail about how those features are powered, and essentially how things come together.

With Xe HPG powering these Arc GPUs, the architecture is far different than Intel’s previous-generation GPUs. The media and display engines are all new, as is the graphics pipeline. It used to be fine to release a GPU with a primary focus on gaming, but Arc is launching with much more responsibility. Intel is taking both creator and AI seriously, and that’s reflected in this Alchemist architecture.

Each Intel Arc GPU will be built with “render slices”, with up to eight per. Each of these render slices include multiple “Xe-core”, each of which has 16 vector engines. Other pieces include functionality for rasterization, geometry, and the rendering backend.

While the render slices house all of the individual Xe core segments, Intel specs its Arc cards with those Xe-core measurements, instead of the overall render slice or individual vector engine count. That’s how a GPU like the A350M comes to be, as it has 96 total vector engines, which is effectively 1.5 render slices. Intel’s top-end part, A770M, has an effective 512 vector engines. Each Xe-core has its own ray tracing, sampler, and threading units, while every two share the same pixel backend. Intel claims that this modern architecture offers 1.5x perf/watt over the 11th-gen Core Xe-LP GPU.

The “XVE” part of the Xe-core represents the vector engines, which handle traditional shader processing, while the “XMX” addition handles matrix multiplies and accumulation, suited for AI-based workloads. To further boost performance, each Xe-core includes 192KB of shared cache. This can be partitioned into L1 cache, or shared into local memory, which is dynamically handled depending on what the particular workload needs.

Each vector engine in an Xe-core has a dedicated FP unit, which can operate with 16x FP32 ops/clock, or 32 FP16 ops/clock. An important step-up over previous Xe releases, Arc will support INT8 at 64 ops/clock. The matrix engine supports the most common AI workloads with support for BF16, FP16, INT8, and surprisingly, even INT4/2. Ultimately, the XMX engine can compute at 256 ops/clock, which Intel says offers a 16x boost to AI inferencing using INT8.

As the next slide highlights well, each pair of vector engines are physically paired together with each XMX engine to work in parallel, which is done to improve efficiency, especially with regards to the co-issuing of instructions being run between them.

One of the major targets of the matrix engine is XeSS, which we covered above. The more efficient an AI architecture, the better the outcome of game upscaling, relying on a convolutional neural network to deliver a convincing end result. This kind of technology wasn’t even thought about a handful of years ago, but now all three graphics vendors are competing hard against once another to deliver the more impressive result. We’re interested in seeing how Arc delivers that in real use.

Final Thoughts

It feels like we’ve been waiting a while for Intel’s Arc to arrive, and now it’s here. Well, effectively. We still believe it will take weeks for notebooks to start popping up, and a couple of months before we see the launch of the desktop counterparts. Based on everything we’ve learned from this launch though, there are many reasons to be excited, or at least intrigued.

Techgage has been around for 17 years at this point, and in all of that time, almost all of our gaming and creator-related GPU content has revolved around two vendors: AMD and NVIDIA. With Arc, that will soon be expanded to three vendors, highlighting a huge shift in the industry. We look forward to seeing how Arc GPUs will handle typical workloads we’ve become accustomed to running – anything from gaming to encoding to rendering.

In a way, it feels like Intel has taken a while to deliver these first products, but at the same time, with such an ambitious launch, we’re glad the company has taken its time. Graphics have always been a difficult task to tackle, but recent years have complicated matters greatly, especially with the introduction of GPU-accelerated creative and AI workloads.

Ultimately, we’re really glad to see a serious third competitor in the discrete GPU market. Even if Intel doesn’t target AMD’s or NVIDIA’s top-end options from the get-go, what we’re seeing so far paints a pretty picture of the future. It’s going to be really interesting to see not just how it lives up at launch, but how it improves over the first year, and beyond.