Today has been a long time coming. When Intel announced it was getting into the graphics business a few years ago, interest was high across the industry. It’s been nearly two decades since we had a third player in the graphics industry and even longer since we had a third player with deep pockets and funding to match the major players. Today, Intel is sharing new details on its Xe line of products and officially launching these new GPUs. Parts will debut for the lower-end mobile family first before higher-end solutions are available later this summer.
The basic building block of the chip is the Xe core. The Xe Core replaces the EUs as Intel’s fundamental “unit” of graphics computing. Each features 16 256-bit vector engines and 16 1024-bit matrix engines for hardware math. This XMX Matrix Engine can perform 128 FP16/BF16 operations per clock, 256 INT8 operations, or 512 INT4/INT2 operations per clock. GPU manufacturers have been steadily adding the ability to run limited precision workloads at high speed as a means of improving neural net power efficiency and increasing overall performance.
When GPGPU workloads launched, single-precision and double-precision FP32 and FP64 calculations were the major facets that we talked about. While these remain important for some workloads, data types like BF16 and INT8 have become more important in recent years.
If you use a standard multiply-accumulate, Xe can perform up to 16 ops/clock. DP4a boosts that to 64 ops/clock (DP4a is essential to Intel technologies like XeSS for AI-boosted upscaling). With XMX, the same chip can perform up to 64 operations / clock when using INT8. The GPU is capable of co-issuing floating point, integer, and matrix workloads simultaneously.
Media Decode Engine, New Features
Intel’s new media decode engine has some capabilities we haven’t seen in-market before. Arc GPUs will be the first cards with AV1 hardware encode support, with support for up to 8K / 12-bit HDR hardware decoding and up to 8K / 10-bit HDR encoding. Display connection technology is limited to HDMI 2.0b and DP 1.4a, though Intel does note that its hardware is DisplayPort “2.010G ready,” implying that manufacturers might be able to update support post-launch? This bit was unclear.
This iteration of Intel’s graphics chip can support up to 2x 8K60 displays, 4x 4K120 displays, or up to one 1080p or 1440p monitor running at up to 360Hz. Apart from AV1 support, this is pretty standard stuff. Intel, for example, also supports Adaptive Sync. That’s the generic term for the FreeSync and G-Sync technologies AMD and Nvidia already field. But it also supports two new technologies — Speed Sync and Smooth Sync. Both are meant to combat tearing, but they do it differently. Speed Sync delivers frames more quickly, supposedly reducing latency. Smooth Shift, meanwhile, blurs torn frames together to avoid the characteristic jagged edge of a V-Sync tear.
Smooth Sync’s schtick is that it blurs this boundary and shifts the frame location, avoiding the appearance of tearing altogether. Tearing isn’t a problem for most games or gamers these days, but if you are an e-sports player gaming with V-Sync off, Intel thinks these features might help slice a bit of latency off your game.
Overall, these are nifty features. It’s not clear what the total impact of them will be on anyone’s median gaming experience, but more developments that prevent tearing and reduce latency are good things.
Product Lineup
Here’s the Xe mobile product family. There will be three families of cards and five SKUs — two Arc 3, one Arc 5, and two Arc 7.
One thing that sticks out about these cards is their relatively small buses. AMD compensates for small memory buses by using large L3 caches. Intel doesn’t seem to think it needs to take this step. The low VRAM on lower-end solutions also means these cards are unlikely to get much use out of their ray tracing capabilities. The Arc 5 is better positioned, but we’re awful curious about how fast the GDDR6 is on these GPUs. Fast VRAM with a narrow memory bus has been a viable strategy before. The desktop versions that eventually debut might also have better specs.
Intel is claiming that the Arc 3 can offer up to 2x the performance of its current integrated Xe graphics. While that’s not going to challenge high-end cards from Nvidia or AMD it might be enough to give them a plausible lower-end part. Intel is broadly thought to have gotten into graphics partly because it wanted to absorb more of the available margin on notebooks, so a relatively cheap dGPU with good performance characteristics is a smart way to do it.
Intel Deep Link
Intel is also debuting some new technologies related to CPU-GPU sharing. Some of these we’ve seen before, but a few of them are genuinely new. First up is a feature called Dynamic Power Sharing. As the name implies, this allows the system to split power allocation between the CPU and GPU depending on which component needs more power in any given moment. AMD has a similar system in Ryzen.
Hyper Encode, Intel’s second feature, is an interesting one. It allows the system to leverage all of the graphics processors in the chassis to handle GPU workloads. If you’re encoding with Intel’s OpenVPL API, which is part of OpenVINO, a software designer can write code that will spread encoding across both the dGPU and iGPU. This is a smart way to make use of iGPU silicon in systems with discrete cards.
Intel’s Hyper Compute offers the same capability, but applied to content creation workloads. This would be potentially useful for content creators or AI upscalers.
Intel expects overall hardware availability for Arc 3 chips to begin more-or-less now, with higher-end cards coming later this year. The company is tiptoeing its way into the graphics market, but this is probably a smart strategy. It doesn’t have the same brand cachet as Nvidia or even AMD, and it doesn’t have a track record of high-end parts in this market. Intel’s first-gen products don’t have to be barn burners in order to compete, especially not in the current market. They just need to offer reasonably good price/performance and competitive feature sets. So far, it looks like Intel is delivering, though the relatively limited memory buses on these cards do raise some concerns. We’ll see how much they impact performance when silicon becomes available.
Now Read: