Intel’s New GPU: Xe3 Architecture Changes, Handheld Gaming CPUs, & XeSS3

The greatest change to Xe3 is that it’s simply bigger, with render slices scaling as much as extra Xe cores per slice, a rise in L1 cache from 192KB to 256KB, a big improve in L2 cache, and extra registers which are higher utilized. 

Micro benchmarks present important enhancements in occluded primitives culling for pointless triangles when rendering sport scenes along with enhancements in anisotropic filtering. 

Its variable register allocation and register modifications additionally purpose to unclog the pipeline in order that the {hardware} will be higher utilized, as one of many greatest issues with Arc in its present Xe2 and Battlemage implementation has been that there’s loads of {hardware}, however it’s not getting used correctly. This is a mixture of each {hardware} points, like with fastened operate models within the structure, and driver points, which it has been slowly addressing. Some of this included shifting off of emulation of issues like execute oblique beforehand to remove overhead. 

For Xe3, Intel famous to us a few of its driver enhancements and software program management panel focus as properly, all of which ought to profit the corporate because it strikes towards its eventual dGPU Celestial GPUs.

This accompanies numerous different bulletins associated to its Panther Lake cell options and laptop computer {hardware}, plus some “AI” and NPU {hardware}. 

We’re principally going to deal with the IP block of Xe3 and the structure and received’t be as centered on the product aspect for laptops. 

Although this isn’t a dGPU half, it’s seemingly that this strategy will both be immediately discovered within the subsequent dGPU or will at the least point out which course Intel goes.

Intel was clear that this isn’t precisely Celestial, which is the structure following within the Alchemist – Battlemage – Celestial – Druid lineup. Intel famous that “Xe3P” will comply with Xe3. The “P” unironically stands for “Plus,” exhibiting previous Intel habits die onerous. Intel didn’t verify this, however the impression we received is that Xe3P would be the “real” Celestial GPUs, whereas this Xe3 makes main modifications that seemingly set the stage for it.

Overview of Announcements

Intel had a number of bulletins to share with the press for at present. For our protection, we’re focusing virtually solely on the Xe3 modifications and micro benchmarks. We’ll cowl a few of the different information as properly, like efficiency/Watt enhancements and XeSS modifications, however we’re not going to get into the NPU and AI processing modifications at present. There’s sufficient to speak about simply with the stuff that’ll have an effect on client desktop parts sooner or later (plus the instant affect to laptops).

All of this follows the announcement that NVIDIA is investing in Intel to construct its personal cell elements with them later, however there’s no information on that matter at present. This is all Intel’s {hardware}.

Naming Confusion

Briefly on the naming: Intel admitted its naming mixture of Xe for IP and Alchemist / Battlemage / Celestial / Druid for branding has been complicated. It was cautious to notice that these elements should not Celestial and the impression we received was that they don’t need to burn the title on an incremental enchancment previous to a pending main overhaul. Intel is sticking with “Arc B-Series” for the Panther Lake cell elements, however is shifting to the Xe3 structure. Xe3P will seemingly be Celestial or desktop elements later.

Xe3 IP GPU Block

Intel particularly talked about designing Xe3 to scale to bigger configuration sizes, which might be excellent news for anybody who desires to see one thing higher-end than a B580-class card sooner or later.

Let’s get into micro benchmarks first, then take a look at the block diagram. 

This is a chart of micro benchmarks, that are workloads designed to focus on extraordinarily particular features or behaviors on a product. A 2x enchancment right here received’t equal a 2x enchancment in most real-world functions, however these permit us to see the place the enhancements are showing. Intel printed these for Xe2 additionally.

In Xe3 for “depth writes,” Intel says it noticed a 7.4x relative efficiency enchancment normalized to clock frequency. We’re not sure, however our understanding is that this isn’t remoted for configuration dimension. This implies that this isn’t an ideal comparability for the reason that Xe core rely is totally different between Xe2 and Xe3 in these exams. This 7.4x enchancment outstrips the change in configuration dimension, although.

We requested Intel what “depth writes” means. The firm advised us that it’s associated to high-Z culling and that this bar represents higher primitives culling within the pipeline, that means culling of unseen triangles and geometry sooner within the pipeline in order to not waste assets rendering unseen objects in-game. An instance may be if a constructing is obstructing a participant — there’s no level rendering the participant if it may’t be seen. Culling isn’t new and batching primitives in ways in which remove occluded primitives has been round endlessly, however this reveals that there’s nonetheless loads of floor to achieve right here for Intel. This will lead to higher utilization of assets and allocating them to extra productive work. Intel advised us that the development to this course of is disproportionately useful, that means that it ought to have an effect in gaming efficiency that will be extra noticeable than different enhancements. We’d count on this to hold over to future Celestial dGPU elements as properly.

The “High Register Pressure Shader” part additionally noticed a big uplift in micro benchmarks at 1.9x to three.1x. Scattered reads improved by 2.7x on the relative scale of time, with Intel noting to us that this has to do with utilizing samplers to learn knowledge scattered throughout one thing like a texture (versus a well-organized knowledge set).

Mesh rendering can be proven right here, with Intel telling us that Xe2 had already supplied a proof of idea round bettering mesh shading. Intel famous that this micro benchmark is consultant of workloads the place a number of polygons are current, telling us that the uplift comes from a bigger cache and extra environment friendly use of its registers. Culling additionally contributes.

Quickly, Intel additionally noticed uplift in anisotropic filtering, which is the previous operate that helps enhance smoothness of textures and objects proportionate to the view frustum’s angle. Ray-Triangle intersection additionally improved by 2x within the microbenchmarks on the relative scale, which is noteworthy since Xe2 already benefitted from comparatively giant ray tracing enhancements.

Looking back at the Xe2 micro benchmarks, Intel then highlighted Draw XI and Compute Dispatch XI primarily. At the time it talked to us about this chart, Intel advised us that this was resulting from implementation of native execute oblique assist for oblique draw and dispatch, versus its Xe1 emulation of those features.

Block Diagram

Time to get into block diagrams for the way the brand new Panther Lake Xe3 block is constructed. This reveals a 12 Xe-core configuration as the utmost dimension introduced for cell, with this configuration carrying 16MB of L2 cache, 2 geometry pipelines, 12 samplers, and 4 pixel backends. The L2 cache is noteworthy right here.

This is the brand new Xe3 render slice. A render slice is Intel’s terminology that defines a block on the GPU containing Xe cores. For reference, the Battlemage B580 with Xe2 has 20 Xe cores on 5 render slices, so every slice is only one a part of the overall GPU.

The Xe2 slice had 4 Xe cores each, with Xe3 shifting to six Xe cores per render slice. Intel additionally intends to scale-up the configuration dimension on cell gadgets to a most of 12 Xe cores (or 2x render slices, up from 8 Xe cores on a previous 2-slice configuration).

The Xe3 render slice reveals that every Xe core has 8 vector engines, which is unchanged from Xe2 cores; nonetheless, Intel is rising the cache dimension in Xe3. Intel’s Tom Petersen acknowledged, “The first thing we’ve done is increase the size of our L2. By increasing the size of the L2 from 8MB to 16MB, we reduced the traffic that hits the memory interface. That’s important because the memory interface is typically one of the most precious resources on a graphics chip. We can see anywhere between 17% and 36% traffic reduction heading towards memory, which has a significant performance effect on these different applications.”

Looking at Intel’s first-party outcomes, it presents the development within the type of relative visitors on the SoC material (within the vertical axis) in opposition to a baseline 8MB L2 cache. Cyberpunk with RT confirmed a 19% discount, Black Myth rasterized confirmed a 36% discount, and the rasterized Steel Nomad check confirmed a 17% discount.

Intel additionally advised us that it has elevated its L1 Cache by 33%, noting a transfer from 192KB to 256KB. When we requested Tom Petersen which space of uplift he thought had essentially the most affect on total efficiency, he pointed us towards the register and thread modifications. Intel has elevated thread rely upwards of 25% relying on configuration and has moved to variable register allocation. Petersen famous that occupancy of the compute models (together with on Battlemage) beforehand wasn’t at all times excessive, regardless of them being out there for work, that means that there was extra GPU {hardware} current than was being correctly utilized by functions. Intel has centered on this in each drivers and {hardware}. He famous that earlier register allocation and thread rely decisions would “starve the pipeline if the shader used too many registers,” which is being addressed.

The ray tracing unit additionally received enhancements. Intel says it “slowed down dispatches of new rays while the sorting unit catches up,” citing out-of-order dispatch and triangle testing. The ray tracing unit enhancements appear to be largely attributed to asynchronous dispatch-test processes.

Intel additionally highlighted a brand new URB supervisor as a part of its fastened operate enhancements, which can be the place we discover the anisotropic filtering uplift. Petersen acknowledged this, “We also now have a new URB manager, which allows partial updates versus flushing the whole thing. Our URB is a structure where we pass results between our units inside of our GPU. It used to be somewhat of a serializing point; now we can actually use that partially without flushing each complex.”

Frame Inspection

We thought these subsequent couple slides had been fairly attention-grabbing as properly:

Intel confirmed a body on Xe3 versus Xe2. These should not normalized for configuration dimension, so it’s not an ideal comparability and it reveals a 12-core vs. 8-core configuration, disallowing an ideal like-for-like inspection. This is iso frequency and energy, so it’s at the least normalized there.

The horizontal axis is for API name execution, with the vertical axis being milliseconds of time to execute throughout a single body being drawn (larger is worse). This is for Cyberpunk 2077. 

Of word, Intel reveals an 8ms discount to Xe3 with the compute and pixel shader part towards the top, assigning a few of that uplift to the change to the variable registers and L1 cache dimension improve. We may also see that, based on Intel, the L2 advantages the render base move with a 0.39ms enchancment, preceded by the transfer to 10 threads (and variable registers) offering a 2.93ms enchancment within the pre-pass. 

More broadly, Petersen advised us in a name that the register allocation and variety of threads would starve the pipeline if the shader used too many registers beforehand, which is being partially addressed right here. He mentioned that the earlier structure might trigger a discount within the utilization of accessible compute assets resulting from common flushing of the pipeline resulting from common reallocation into reminiscence.

This picture is fairly cool and is a take a look at what really occurs in a body when it’s being drawn. We have a full video talking about this beforehand.

Power Delivery

Intel’s deal with energy supply and energy administration cites learnings from the MSI Claw (learn our overview) gadgets and principally comes within the type of guaranteeing correct useful resource allocation for energy finances between the CPU and GPU, which ought to profit laptop computer and handheld gadgets which have a restricted energy finances cut up between the 2. 

Intel famous that beforehand, an absence of software consciousness meant that the machine might generally divert an excessive amount of energy to the CPU, leaving the GPU bottlenecked on its energy restrict whereas the CPU supplied a stage of efficiency that wasn’t being kept-up with by the GPU. 

Intel gave the MSI Claw for example of a time this didn’t go properly.

The firm famous that it improved on this earlier within the 12 months with its Intelligent Bias Control v2 and is now introducing a v3 to construct upon that. 

Because the system was beforehand unaware of the applying being run, on this case a sport, Intel mentioned that software program and {hardware} wouldn’t accurately steadiness the workload between the CPU and GPU, leading to stuttering resulting from being energy starved.

“Intelligent Bias Control v2” took GPU heuristics and utilization metrics to then inform thread scheduling and useful resource project on the working system-level. Intel had beforehand marketed enhancements to 1% and 0.1% low metrics by way of higher body interval pacing on account of this transformation.

The new v3 model of this provides E-core first scheduling, which is self-explanatory in that E-cores get scheduling first when gaming. This sounds worse, and usually could be, however Intel says that the top result’s lowered energy diversion to the CPU by utilizing decrease energy cores previous to P-cores, freeing-up extra of the shared whole energy finances to go towards the GPU as a substitute. In GPU-bound eventualities, like many video games significantly on handheld gadgets, this can be a higher end result than burning energy on a element that isn’t as burdened. 

This comparability between Panther Lake and the prior era of this bias management answer reveals that peaks in energy utilization have smoothed-out whereas the GPU energy consumption has leveled to be extra predictable. Reminder: This is a first-party exams. The GPU can be getting extra whole energy finances as a proportion than beforehand, whereas decreasing CPU energy in trade. For GPU-bound eventualities particularly, this must be a greater end result. It may assist in some CPU-bound eventualities as properly.

XeSS Multi-Frame Generation and Other Changes

Intel additionally introduced XeSS 3, which incorporates XeSS-Multi-Frame Generation (or XeSS-MFG). A couple of extra letters and so they’ll have the entire alphabet. 

XeSS-MFG is conceptually just like NVIDIA’s MFG. XeSS-MFG takes 2 actual frames to calculate optical move networks utilizing movement vectors and the depth buffer, then makes use of that info to generate as much as 3 frames between the two actual frames. The frames are then displayed so as and paced in a technique to reduce animation error. We even have a separate deep-dive on our new animation error testing methodology.

The new “XeSS Frame Generation Override” setting within the driver software program permits the consumer to set 2x, 3x, or 4x mode.

Intel offered just a few timelines of a single body. One at native, after which a number of with varied ranges of XeSS know-how in use. The shorter the period that the body is on the X-axis, the much less time the body took to finish. The high half of every reveals directions and the underside half reveals when the geometry pipeline is lively.

Compared to native, the raster, RT, and denoise sections of the body are shorter on the XeSS 3 timeline resulting from rendering at a decrease decision. The first purple part represents XeSS-SR to carry out the upscaling. The second purple block begins with the optical move portion of body gen, adopted by 3 body era operations.

It looks like Intel’s argument is that the complete body gen course of takes much less time than drawing one actual body, and is due to this fact higher or one thing, however this completely ignores picture high quality. We’ve proven with each AMD FMF and NVIDIA MFG that the picture high quality sacrifice isn’t at all times value it. Sometimes it’s, however it’s not at all times so simple as being that means. Intel acknowledged that these frames upscaled with XeSS-SR are the identical high quality as native, which is unlikely. Intel acknowledged: “That frame is as good as the prior picture, the native frame. But it’s actually being run quicker.” We doubt this might be broadly true and can consider in a while dGPUs. It was bullshit when NVIDIA claimed it, too. The high quality will be good, however is never pretty much as good.

Intel had another side-by-sides that we take challenge with, and that together with nonetheless having watermarks on the video means we’ll skip them and simply check it ourselves later. 

Intel referred to the body gen course of as trying into the long run. NVIDIA CEO Jensen Huang has mentioned related issues about NVIDIA’s body era. Both of them are fallacious, as a result of all present strategies of body era rely solely on completed frames and engine knowledge. These frames already existed and will have been displayed as a substitute of holding them to run the body era in between. That isn’t trying into the long run, that’s interpolating between two sequential snapshots of the current or close to current. Until a predictive technique of body era comes out, none of those applied sciences look into or generate “the future,” they at greatest interpolate the previous. And that’s high-quality, however we’d actually prefer it if these firms might get their shit collectively and cease saying that they generate the long run. 

MFG represented on benchmark charts has been a significant and ongoing controversy and misrepresentation of efficiency on NVIDIA’s aspect of issues. Intel dedicated to counting on base raster efficiency with out body era because the baseline for efficiency and mentioned that, when it publishes numbers together with upscaling or body gen, these might be supplied as supplemental to the bottom metric. We suppose this can be a higher steadiness of selling the potential with out completely misrepresenting the fact.

Intel additionally talked a few new model of PresentMon that features a few modifications, partly accounting for body era know-how.

---