Introduction

Intel Arc "Alchemist" A770 is the company's first major attempt at high-performance gaming, which is what the Xe-HPG graphics architecture is all about, we'll take a closer look in this review. Up until "Alchemist," Intel was mainly focused on giving its client processors a decent integrated graphics solution with Xe, that could handle the rich user interfaces of modern operating systems, as well as web-browsing with increasing amounts of raster graphics and video acceleration. Intel's entry to the high performance discrete graphics segment is fueled mainly by the growth in PC gaming, and rising margins in discrete GPUs by established players NVIDIA and AMD, which has the potential to spike each time crypto-currency mining experiences a cycle. Intel's biggest strides toward discrete GPUs were taken in the thick of the GPU supply crisis, as gamers were ready to give up on the PC platform, due to unreasonably high prices from crypto-mining and scalping. If Intel can establish itself as one of the good guys to gamers, demonstrating that it can deliver a good product, with good supply, then it will unsettle the NVIDIA-AMD duopoly, and establish Intel as a key player in this segment.



Intel Xe-HPG is also a co-product of Intel's renewed push in the accelerated computing space, where it's seeing rivals NVIDIA and AMD release HPC processors to compete in the high-margin AI processing space. Intel developed the Xe-HP architecture and the "Ponte Vecchio" HPC processor. This also meant that the company had a formidable, highly scalable SIMD architecture that it could build discrete GPUs with, if it can just innovate the necessary raster graphics and ray tracing hardware, and build a software stack around it. Intel knows that it cannot compete in the high-end discrete GPU space without being contemporary with its feature-set, and so the entire product-stack in the Arc "Alchemist" family meets DirectX 12 Ultimate spec, including real-time ray tracing, mesh shaders, variable-rate shading, and sampler feedback. Intel also knows that native-resolution rendering is sliding out of the realm of possibility for mainstream GPU price-points, and given that the other companies developed features such as DLSS and FSR, it needed one such of its own—XeSS. Intel "Alchemist" is technologically closer to NVIDIA GPUs than AMD.

There is a greater role of fixed-function hardware in the ray tracing architecture, including SIMD innovations such as shader reordering; and the GPU has AI-acceleration hardware just like NVIDIA has Tensor cores. The AI hardware is used not just for denoising, but also to accelerate XeSS. Intel's philosophy with the software-stack of its GPUs is similar to that of AMD and NVIDIA: to keep things as "open" and accessible to other brands as possible. OneAPI is the overarching software foundation across Intel's accelerated processing group, just like CUDA, while features such as XeSS are kept accessible to other brands too, via alternative means such as compute shaders.

The higher-end of the Arc "Alchemist" desktop graphics card family is led by the Arc 7-series, with two SKUs, that we both review today: the Arc A750 and the Arc A770 Limited Edition. The A770 maxes out the ACM-G10 silicon that both these SKUs are based on with all 32 Xe cores enabled, and comes with 8 or 16 GB of memory, while the A750 enables 28 out of 32 Xe cores, and comes with 8 GB of memory. Both cards feature the full Xe feature-stack, including XMX AI acceleration, DirectX 12 Ultimate, XeSS, etc., and target the same category of gamers—the performance segment, which plays at resolutions of 1080p or 1440p, with high to maximum details. Intel is targeting the middle-of-the-market segment, as this where most gamers shop for a serious gaming graphics card. NVIDIA sells its popular GeForce RTX 3060, and AMD offers the Radeon RX 6600 XT. In this review we have with us the Arc A770 Limited Edition. This card offers 32 Xe Cores, which work out to 512 execution units, or 4,096 unified shaders. Its 16 GB of memory ticks at 17.5 Gbps across a 256-bit wide memory bus. Intel is pricing the 16 GB version of the A770 at $350. There's also an 8 GB variant starting at $330.

Architecture

Intel's Xe-HPG "Alchemist" graphics architecture sees its biggest implementation to date with the ACM-G10 silicon powering the A770 and A750. Built on the 6 nm process at TSMC, the ACM-G10 measures 406 mm² in die-area, and packs 21.7 billion transistors. Much like NVIDIA and AMD, Intel has innovated its own hierarchy for the number-crunching machinery of its GPUs, and differentiates SKUs by changing the number of indivisible groups of these to meet performance targets. The ACM-G10 silicon features a PCI-Express 4.0 x16 host interface, a 256-bit wide GDDR6 memory interface, the Xe Media Engine and Xe Display Engine, among a global dispatch processor, and memory fabric that's cushioned by L2 cache. The main SIMD component top-level organization is the Render Slice. Each of these is a self-contained unit with all the number-crunching and raster graphics hardware a GPU needs.


The ACM-G10 features eight such Render Slices. Each of these features four blocks of indivisible processing machinery, called Xe Cores; four Ray Tracing Units (RT units), and DirectX 12 Ultimate optimized raster-graphics units that includes four Samplers, Tessellation geometry processors, 16 ROPs, and 32 TMUs. Since there are 8 Render Slices, the silicon physically has 16 RT units, 128 ROPs, and 256 TMUs.


The Xe Core is the indivisible computation core, with sixteen 256-bit Vector Engines (execution units), sixteen 1024-bit XMX Matrix Engines, and 192 KB of L1 cache. Each Vector Engine has eight each of FP and INT units, besides a register file. Two adjacent Vector Engines share a Thread Control unit to share execution waves. There are 16 VEs per Xe Core, 4 Xe Cores per Render Slice, and 8 Render Slices on the ACM-G10 silicon, so we have 512 execution units in all, each with 8 FP/INT execution stages, which logically work out to 4,096 unified shaders. The Arc A770 is configured with all 32 Xe Cores, so it gets 4,096 shaders. The A750 gets 28 Xe Cores, or 7 out of 8 Render Slices, or 448 execution units, amounting to 3,584 unified shaders.


The XMX Matrix Engine is extremely capable matrix-multiplication fixed-function hardware that can accelerate AI deep-learning neural net building and training. It's also a highly capable math accelerator. Intel originally designed this for the Xe-HP AI processors, but it finds client applications in Arc Graphics, where it is leveraged for ray tracing denoising and to accelerate features such as XeSS. There are 16 XMX units per Xe Core, and 64 per Render Slice, 512 across the ACM-G10 silicon. Each XMX unit can handle 128 FP16 or BF16 operations per clock; up to 256 INT8 ops/clock and 512 INT4 ops/clock. The XMX-optimized native XeSS code is an order of magnitude faster than the industry-standard DP4a codepath of XeSS, as you'll see in our testing.


When it comes to real time ray tracing, Intel's Xe-HPG architecture has technological-parity with NVIDIA RTX, due to its heavy reliance on fixed-function hardware for ray intersection, BVH, and AI-based denoising.


There are several optimizations that further reduce the burden of ray tracing operations on the main SIMD machinery, such as shader execution reordering which optimizes shader work threads for streamlined execution among the SIMD units. NVIDIA is only now implementing such a feature, with its GeForce "Ada" architecture. There's a special component in each Xe Core that reorders shader threads. It's essentially a very intelligent dispatch unit. Intel refers to its ray tracing architecture as Asynchronous.


With Moore's Law tapering, despite Intel claiming otherwise, the writing is on the wall—rendering at native resolution is over, at least in the performance and mainstream GPU segments. High-quality super resolution features such as DLSS and FSR are helping NVIDIA and AMD shore up performance at minimal quality loss, by rendering games at lower resolutions than native to the display; and upscaling them intelligently, with minimal quality losses. Intel's take is XeSS. The company claims that this is a 2nd generation super-res technology, on par with DLSS 2 and FSR 2.0. The XeSS upscaling tech is as easily integrated with a game engine's rendering pipeline as TAA (or AMD FSR). The XeSS algorithm is executed as XMX-optimized AI on Arc GPUs, and as DP4a-compatible code on other GPU brands. The algorithm takes into account low-resolution frame data, motion vectors, and temporal data from previously output high-res frames, to reconstruct details, before passing on the high-res output to the game engine for post-processing and UI/HUD.


The Xe Display Engine is capable of up to two display outputs at 8K 60 Hz + HDR; up to four outputs at 4K 120 Hz + HDR, and up to four 1440p or 1080p with 360 Hz + HDR. VESA Adaptive Sync and Intel Smooth Sync features are supported. The latter is a feature that runs the GPU at its native frame-rate, while attempting to remove the screen-tearing from the display output. A typical Arc desktop graphics card has two each of DisplayPort 2.0 and HDMI 2.0b connections. The Xe Media Engine provides hardware-accelerated decoding and encoding of AV1, and accelerated decoding of H.265 HEVC, and H.264 AVC.


Besides VESA Adaptive Sync, the Xe Display Engine offers a software feature called Smooth Sync. This gives fixed refresh-rate monitors the ability to play games without V-Sync, rendering them at the highest possible FPS (and the least input latency), while attempting to eliminate the screen-tearing using a shader-based dithering filter pass. This is an extremely simple way to solve this problem, and we're surprised AMD and NVIDIA haven't tried it.

Packaging

The Card

The A770 is a beauty. It comes with a clean and stylish design that could be coming right out of Apple's design labs. The black matte surfaces are soft to the touch, thanks to some rubbery surface application. The rubber isn't too sticky, it's very nice to the touch. The back of the card has a backplate, but not one in the classical sense, more on that further down on this page.


Dimensions of the card are 27.0 x 11.5 cm, and it weighs 1087 g.


Installation requires two slots in your system.


Display connectivity includes three standard DisplayPort 2.0 ports and one HDMI 2.1.


The card has one 8-pin and one 6-pin power input. This configuration is rated for up to 300 W of power draw. If you look closely, you'll notice that the connectors are slightly mismatched in color. We confirmed with Intel that this will be fixed on future production runs, but some cards will make it in the retail market.


On the Arc A770, the RGB control functionality is provided through this external port. The included cable goes into a USB port in your motherboard. This is a surprisingly basic implementation, all competing cards send RGB commands over the PCIe bus to a microcontroller that's sitting on the GPU's I2C bus.

Teardown

Taking the Arc A770 apart is slightly more complicated than cards from other brands. While there's solid engineering everywhere, I suspect that the limited experience Intel has with building such cards is why some things are solved in a slightly less efficient way.

First, you have to take off the "backplate", which is a thin metal plate that's glued to the frame on the back of the card. Yup, glue isn't good for enthusiast end-users, I could have preferred a design with screws, as that's much easier to maintain. If you heat up the backplate with a heat gun, the glue will become soft and easy to remove.


Once the backplate is removed there's a bunch of Torx screws to take out, nothing out of the ordinary. Here we've encountered some sort of cooling plate that sits on top of the VRMs on the back—except there are no VRMs on this side, just an empty silkscreen on the PCB.


Now you can separate the cooler from the PCB assembly.


This metal reinforcement brace covers the PCB and helps strengthen it against bending.


Intel's cooler uses a large vapor-chamber baseplate paired with five heatpipes.

High-resolution PCB Pictures

These pictures are for the convenience of volt modders and people who would like to see all the finer details on the PCB. Feel free to link back to us and use these in your articles, videos or forum posts.


High-resolution versions are also available (front, back).