Intel Arc GPU Updates: Raja Koduri and Team on OEM Graphics Cards, Driver Features, Workload Balancing, and Much More

Gaming

As Intel gears up for the launch of its push into the discrete gaming GPU market, information about the first wave of products has started to come out. We already know that the company has created a new brand called Arc for these GPUs, and that at least four generations have been planned and mapped out already. Now expected in early 2022, the first-generation Intel Arc graphics cards, codenamed ‘Alchemist’, will be based on the Xe graphics architecture – more specifically, the Xe HPG implementation, which was adapted specifically for the gaming PC segment. Beyond this, Intel continues its ambitions in the data centre and high-performance computing spaces for AI and machine learning, with higher-end Xe processors that share the same architectural underpinnings, as well as entry-level integrated graphics based on Xe LP.

While the company is still holding a few secrets, it will soon be competing with Nvidia and AMD, which have up until now had a duopoly in the gaming GPU market. This comes at a time when cryptocurrency mining is still resulting in massive stock shortages and price spikes everywhere, and global semiconductor supplies are stretched thin.

Gadgets 360 recently had a chance to participate in a global roundtable conversation, and we asked some questions to Intel’s Raja Koduri, SVP and GM of Accelerated Computing Systems and Graphics, as well as Roger Chandler, VP and GM of Client Graphics Products and Solutions, and Jeff McVeigh, VP and GM of Data Centre XPU Products and Solutions. Here are most important takeaways for gamers and enthusiasts about what we can expect as the official Intel Arc launch draws nearer.

How does Intel want to position itself as a major player in the discrete GPU market? How long are you expecting to take to win significant market share from current major players?

Roger Chandler: The way we’re positioning the product is that Intel brings the full platform. There are opportunities to be providing platform value that’s going to be great for users. Intel is not a stranger to graphics. We’ve been in graphics for years, and over the past couple of years we’ve quadrupled the performance of our integrated graphics, so you actually have integrated graphics products that are basically performing at the same level as some entry-level discrete products. As we as we approach the market, we’re not going to be shy, we’re very aggressive, but it’s going to take time. It’s one step at a time, but we’re in it for the long haul. This is a very big bet for us and we’re going to take our best shot. We’re very excited.

How does Intel intend to establish its discrete graphics ecosystem? Will it work with partner brands?

Roger Chandler: Intel has a long history of working with the Taiwanese ODM community as well as OEMs around the world. So, at a platform level we are extending those relationships, and we are modifying them because it’s different with consumer-based discrete graphics products. But we actually found that a lot of the relationships and processes were extremely applicable to this line of product. Also, that’s one of the reasons we did DG1, our prior discrete graphics product launched last year, so we can actually make sure all of these pipelines, connections and relationships are really solid. There’s the ODMs and the OEMs, but I think also it’s important to call up the ISV ecosystem, and Intel has decades of experience working with ISVs and game developers.

Will Intel make reference cards, or will add-in board partners only have custom designs?

Roger Chandler: We’re providing the designs that are needed to remove friction for our partners to really accelerate and differentiate themselves on top of our products. Now whether or not it’s comparable to what’s in the market today, we’re not going to go into those details.

Was the presentation of Xe HPG at Architecture Day representative of the maximum configuration for Alchemist?

Roger Chandler: The architecture is designed to be scalable, from Xe LP all the way up to HPG high-performance gaming. We’re not disclosing all the details about the maximum extent by which this could scale. There is a limit and then you start getting into different segments that we’re looking at as well. We want to make sure that we can enter new markets as the opportunities arise.

Will Intel Arc graphics cards be optimised for professional graphics software? Are there plans for a separate graphics card line specifically for pro graphics?

Roger Chandler: There are a host of workstation-class, creator-class professional application vendors that we work closely with and we’re making sure our products work extremely well with their solutions. So yes, we will have support for those workloads and platforms. We’re not going into all the details and specifics yet, of applying products to those particular markets, but the product is capable.

Is Intel confident that it will have enough supply of Arc GPUs are launch time?

Raja Koduri: The response to our announcements at architecture day from our OEM customers and end users has been phenomenal. Everyone is welcoming us, so that’s great news. We’ll be coming in as a third player. I’ll always be very cautious, when the demand is so high and when the market is so hard. I can always use more supply. So I’m not going to say I have enough supply in this high-demand market. I think every one of my competitors will say the same thing right now.

Laptops such as the MSI Prestige 14 Evo have prominent badges displaying specifications and marketing messages

Will Alchemist have any hardware or software lockouts to discourage mining and will we. How will we manage retail pricing to prevent all that.

Roger Chandler: This is a tough one to answer, but what I will say is that we are designing Intel Arc and the Alchemist family of products as gamer-first and creator-first. All the optimisations, the features, everything we’re doing, is really to make sure we can solve problems and deliver value to gamers and creators. As far as like software lockouts and things of that nature, we’re not designing this product or building any features at this point that specifically target miners. As far as actions we’re taking to avoid or lock them out, it’s a product that will be in the market and people will be able to buy it. It’s not a priority for us.

Raja Koduri: We’re not putting any extra work, yes.

Will Intel consider manufacturing its GPUs in-house, vs at third-party foundries?

Raja Koduri: Depending on the design points of the product, [we look at] what is the best technology that we have access to both internally and externally. We also look at capacity. We have to optimise, particularly for GPUs, which is a new line for Intel, [so] that we always have capacity, whether it’s internal or external.

If you look at our first-generation discrete card, the Xe Max, it was produced on Intel SuperFin technology. We’ll look at the roadmap published at Intel Accelerated, all the way to Angstrom era. I look at that roadmap with great interest, to see how I can leverage that.

Ponte Vecchio supports Xe Link, so can we expect Intel Arc to support this too?

Raja Koduri: At this point for Intel Arc our goal is to deliver the maximum performance possible in the socket. You know the PCIe form factor, MCM form factors, and the data centre market are different. For AI and HPC there is a tonne of compute; lots of Watts in a given box. So, the links, the scale-ups, the network bandwidth, are far more important. So, Xe Link is currently limited to our data centre products.

What’s the status of the support for XeSS in games? How many games will be supported on launch day?

Roger Chandler: XeSS has created a tremendous amount of excitement, and when we talked about it at Architecture Day, we actually showed the capabilities. We are actively working with dozens and dozens of studios right now. It’s only going to grow as well because really what motivates developers is the fact that they understand how many platforms support this feature. Our open approach is really driving more excitement than pretty much any other feature in the ecosystem right now. You should expect a very healthy collection of games to support this, and that will grow aggressively over time.

Will XeSS will be backward compatible with DG1 and Xe LP GPUs?

Roger Chandler: Yes, it should be. You will be able to experience the benefits of XeSS on our Gen 12 Xe-based graphics. This includes both the Tiger Lake as well as the DG1 platforms.

Intel demonstrated the DG1 discrete graphics card at CES 2021, before seeding it with developers for evaluation

Will there be a badge for PCs with Intel GPUs, similar to Core or Evo, that OEMs will use?

Roger Chandler:  It’s very important for brand awareness to start growing. Specifically how it will be applied on systems that use it? You should expect that it will be prominent. I can’t really go into the details of what you should expect to see on the shelf, because to be honest, we’re still kind of working through some of that. But we developed the brand and rolled it out, and we want to make sure that there is definitely user association with that and they understand that what it does bring to the table is great.

This brand actually has some meaning for us as well. We didn’t just want a word, and we feel that graphics really empowers gamers to experience stories, and it empowers creators to tell new stories. If you think about the word Arc, all stories have this arc that represents the plot and the characters. You should expect that to be the theme as we actually go to market with the product, and how we apply the brand.

Will PCs with an Intel CPU and GPU be optimised to balance workloads, system power and thermals, especially in laptops?

Roger Chandler: Deep Link is a technical marketing term to reference our platform’s capabilities [such as] the ability to use your integrated and discrete GPUs together, and another one is platform power [between] the CPU and the GPU. We have capabilities to balance the power pretty elegantly. If you have a workload that’s more CPU-intensive, we can divert a bit more power to the CPU and actually deliver better performance overall from a platform perspective. The products are going to work where users want the products to work, but that being said, we’re absolutely going to make sure these things run great together.

Jeff McVeigh: On the data centre side, it’s a very similar story. We are always tying together and leveraging what Xeon offers as well as our GPUs, and we bring that together with the unified programming model of oneAPI. Making sure that we’ve got the right balance is not just always offloading, but doing a productive offload so that you get the best performance for the power envelope that you’re dealing with. Our tool chain gives us a lot of advantage to do that. Workload focus is really where we can showcase the value of that full system, not just a CPU or GPU perspective.

Enthusiast GPUs are now pushing into the 350-400W range. Will Alchemist consume as much power as possible to maximise performance, or is power efficiency more important?

Raja Koduri: Power efficiency always matters. Always. We focus on that, optimise on that, and with respect to the power envelopes, we are going to work with partners and we are going to enable them with various tools. You know our core architecture will be focused on different frequencies, different voltages, different choices. We leave a lot of choice with that. But, my short answer is efficiency still matters a lot.

Any key learnings from the first discrete product, Xe Max?

Roger Chandler: I think there are a lot of key learnings, because this was our first discrete graphics product, and we did it for that reason. We wanted to learn as much as possible. There were a number of things to make sure we could fine tune in regards to how Intel works with ODMs and OEMs. It was targeted as an entry-level graphics product, but having two GPUs that are similar in performance in the platform, such as a 96 execution unit Tiger Lake [integrated] GPU and DG1, we learned a lot about how they operate together and the workload behaviour. A lot of people are [now] creating and streaming videos. I think our software and driver teams discovered a lot of opportunities we didn’t realize were there when you combine these GPUs with a unified driver. That was [one of the] happy discoveries we made; I wish we had discovered it two years before, instead of nine months before.

Jeff McVeigh: We’ve also had our first server GPU in the market as well since November of last year (SG1), derived from the same microarchitecture as Xe Max, and we learned very similar things. We have deployments from Tencent and Alibaba today for applications like Android cloud gaming. So those are really some early beachheads that leverage some key technologies that we have around media processing and graphics, and that’s allowing us to now scale into much broader deployments with our follow-on products.

The upcoming ‘Alder Lake’ 12th Gen Core CPU family will feature integrated Xe graphics

Are you committed to a driver release schedule for GPUs?

Roger Chandler: Yes. Also, just you know, making sure [we] release updates associated with some of the major titles coming out as well, because there’s a lot of great work we can do there. So, we do plan to have a very nimble and dynamic as well as a regular release schedule for this.

Will driver software for Arc support features like game recording?

Roger Chandler: The user experience is important. Actually, this is an area we’re very excited by because we feel there are a lot of features and capabilities. These emerging usages, like game streaming and recording, are things that we are spending a lot of time making sure we understand, and if there are features and capabilities that will remove friction for users. We’re including it in our roadmap so you should expect an evolved set of capabilities that we’ll be rolling out with the product itself. It’s almost as important as the product itself in a lot of ways, because it is the entry point for a lot of users.

Where is Intel is most invested, between Arc consumer graphics and Ponte Vecchio for the data centre?

Raja Koduri: It’s less about where the competition is at and more about what users and workloads we want to enable first. We ship many hundreds of millions of units of integrated graphics every year. Our first target was to make their user experience way better with much higher performance, even in smaller form factors, [from] mobile to desktop. We invented technologies that scale from integrated graphics to the big 4K 240Hz displays that you all want to get to. How do you take advantage of the full platform; a CPU’s integrated graphics with discrete graphics? These are the kinds of things that we are passionate about.

The data centre, high-performance computing, and AI markets can’t have enough compute. So, us being able to bring in Ponte Vecchio, with the highest compute density and memory bandwidth, in the smallest package, is what Intel’s differentiation is. I won’t say much more than that. We solve problems that our competition isn’t focused on.

Final thoughts from Raja

Raja Koduri: This is exciting time to be at Intel. Since Pat [Gelsinger] coming back, the last six months have been highly energizing, for me personally and for all the engineers. We have product leadership and technology leadership as the key; the only thing that matters. And the world needs a tonne of manufacturing at scale. We are doubling down on manufacturing technology, but the core of our strategy is product leadership. [There is] demand for all forms of compute, whether it’s visual computing or accelerated general-purpose compute driving things like AI. It’s just massively exploding.

Our lives depend on computing. I’d like to have a conference call in fully immersive 3D within the next five years, and the amount of compute, the amount of bandwidth you’d need for that. So all the pieces of key technologies are coming together. The next five years will be so much more exciting than the last 10 years for the whole industry, in my opinion.

Some responses have been edited for length and clarity.