The reveal of the Intel Xe HPG during Intel's Architecture Day 2020 was part of the virtual firehose of information on a ton of upcoming products and technologies. We've covered Tiger Lake and Xe LP Graphics elsewhere, and we have new details on Xe HP / HPC and the newly christened 10nm SuperFin process. But one of the products we're most interested in seeing and testing isn't coming until 2021. We've speculated on what Intel's Xe Graphics would bring to the dedicated GPU since we first heard about Team Blue's aspirations back in 2018. The latest details are far more promising than what we were previously expecting, so a delay until (hopefully earlier rather than later) 2021 isn't necessarily a bad thing. Intel's dedicated GPU plans are definitely going to be more potent than the DG1 demo at CES back in January.
Intel originally had plans for a single graphics architecture with two micro-architectures back in 2018. Two years later, the number of microarchitectures has doubled to four. Xe LP is for integrated graphics and entry-level solutions, Xe HP is for high-end compute and data center workloads, and Xe HPC is basically Xe HP on steroids (and possibly with a shrink to 7nm, sort of), targeting supercomputing exascale solutions. That leaves the fourth and final microarchitecture that Intel just announced during its Architecture Day briefings: Intel Xe HPG.
Given what we know of Xe HP, the bifurcation into two different parts makes a lot of sense. First, Xe HP of necessity has FP64 (64-bit floating point) support, along with tensor cores. These are common features in data center compute, deep learning, and AI environments, but they're also extra bloat that's not needed for a gaming GPU. Second, Xe HP will use HBM2e memory.
Intel's Raja Koduri joked about having scars on his back from trying to launch two different HBM (high bandwidth memory) GPUs into the consumer market, referring to his time at AMD with the Fiji / R9 Fury X and Vega 10 / RX Vega 64 product launches. Both performed okay, but the adoption of HBM inflated costs and the resulting prices of both GPU families, ultimately leading to less than optimal prices and performance. For data center workloads, the additional bandwidth that HBM2e brings to the table makes sense — Nvidia uses HBM2 in its Nvidia P100, V100, and A100 products — but for consumers, it's still too expensive.
Intel Xe HPG will address both of these items by removing the FP64 support (or at least trimming it way down) and by adopting GDDR6 memory. That's not the only change, however. Dumping FP64 (possibly tensor cores as well) and HBM gives Intel room to add in other features. Specifically, Intel confirmed that Xe HPG will support hardware ray tracing. That's a critical move, considering the AMD Big Navi / RDNA 2 and Nvidia RTX 3080 Ampere will both arrive in the next month or two with ray tracing, and in November the Sony PlayStation 5 and Microsoft Xbox Series X will launch with AMD GPUs that also support the feature. Trying to break into the dedicated graphics card market with a product that lacks features the established players support wouldn't go over well.
There are still plenty of unknowns. How many EUs will Xe HPG support? Xe HP appears to have up to 512 EUs per tile, or the equivalent of 4096 shader cores (ALUs) if we're comparing it with AMD and Nvidia GPUs. AMD's Big Navi meanwhile is expected to support up to 5120 shader cores, while Nvidia's Ampere could go as high as 8192 CUDA cores (but will probably come in below that mark). With the reworking of features, Xe HPG could certainly increase the number of EUs to stay competitive, and a configuration with 640 EUs (5120 ALUs) isn't out of the question.
The other major bombshell that Intel dropped is that it plans to utilize a third party for fabrication of Xe HPG. That means it won't have to use up its limited 10nm SuperFin capacity making gaming GPUs, and it will perhaps join AMD and Nvidia by using TSMC to manufacture Xe HPG. It's either that or Samsung; either way Intel gets access to 7nm fabrication technology, though 10nm SuperFin is perhaps similar or even superior in practice — the nanometer numbers are certainly prone to being used as marketing rather than indicating true feature sizes.
Despite everything we don't yet know about Xe HPG and Intel's enthusiast gaming aspirations, don't count Team Blue out just yet. Intel demonstrated 1-tile, 2-tile, and 4-tile variants of Xe HP running a compute workload, with almost perfect scaling based on the number of tiles. The test used early drivers, running at non-final clocks of 1300 MHz. Intel was able to deliver 10.6 TFLOPS (FP32) with a single tile and up to 42.3 TFLOPS with a 4-tile implementation.
Obviously, this isn't the same thing as playing a game. Still, strip the extra stuff out of Xe HP, add some more EUs, and clock it at 1.7 GHz and we're talking serious firepower. If Intel can do that with Xe HPG (and we're not saying it will, but merely that it could) and Intel would have a gaming graphics card capable of 17.4 TFLOPS. That's far more than Nvidia's current RTX 2080 Ti, and it could end up being competitive with even Big Navi and Ampere — assuming drivers, ray tracing performance, and everything else fall into place.
The drivers experience with Intel has certainly improved, at least, as we discuss further in the Xe LP article. We recently tested Ice Lake Gen11 graphics as well as other integrated graphics solutions on Intel (and AMD) CPUs from the past several years. While the older HD 4600 (Gen7) graphics failed to run several titles (it lacks DX12 and Vulkan support) and had very poor performance, UHD 630 and Iris Plus worked properly. Scaling to 20 times as many EUs as UHD 630 should do wonders for performance, naturally.
Intel's ability to scale to a 2-tile solution linked via EMIB might also solve the scaling problem we've seen with SLI and CrossFire. The two tiles would simply appear as a single GPU (maybe), meaning there's even potential for a 2-tile Xe HPG solution that could theoretically crush any current GPU. It wouldn't be cheap, and scaling in games probably wouldn't be nearly as good as what Intel showed with the compute demo above, but we'll wait and see.
We're not saying any of this will happen, of course, but even a 512 EU Xe HPG solution could be competitive. Maybe. The proverbial proof is in the eating of the pudding — ray traced pudding, naturally. The good news is that, whatever happens, the next year of GPUs is shaping up to be far more exciting than the past year. New products from AMD, Nvidia, and Intel are coming to market, all with ray tracing support, and any one of the big three could come out on top.