Intel Reveals Full Details for Its Arc A-Series Mobile Lineup

[harlan4096]

Desktop GPUs still coming at a later date

Today marks the day when Intel Arc Alchemist officially enters the dedicated graphics card fray, turning this two-pony race into a three-way battle for supremacy. Or at least, that's the story we'd like to tell. Instead, we're getting the first full description of Intel's Arc A-Series mobile GPU lineup, with the first laptops sporting the new GPUs slated to arrive in April. But desktop users need not worry: Arc graphics cards for desktops are still coming, just not first. If you're not up to speed on Arc, we'll recap some of the details, but our main Arc Alchemist hub contains more low-level details.

Similar to its CPU stack, Intel's Arc Alchemist will have multiple tiers of performance: Arc 3, Arc 5, and Arc 7. The various models will be based off two chip designs, one with up to 96 Xe Vector Engines (XVE) and the other with up to 512 XVE. Note that XVE is the new name for what was formerly called an EU (Execution Unit), though the new XVE makes plenty of changes and certainly warrants a change in name.

Two Physical Arc Chips

The smaller chip is called ACM-G11, and it will launch first, with laptops expected to arrive shortly using the new Arc 3 branded GPUs. ACM-G10 uses a significantly larger chip and Intel expects laptops using the Arc 5 and Arc 7 GPUs to arrive by early summer.

In most respects, ACM-G10 has four times the hardware of ACM-G11: 4X the Xe-cores, ray tracing units, and L2 cache. The memory subsystem however is 2.67X as wide, with a maximum 256-bit bus compared to a maximum 96-bit bus on the smaller chip, and the PCIe slot interface is twice as wide at x16 vs. x8. The media and display capabilities meanwhile are equivalent between the two GPUs, so all Arc graphics solutions will have the same dual Xe Media Engines (MFX) and four display engines.

To quickly recap, each Xe-core contains 16 XVEs, and another 16 XMX units. The XVE vector engines can do 16 FP32, 32 FP16, and 64 INT8 operations per clock. The XMX matrix engines meanwhile can do 128 FP16/BF16, 256 INT8, or 512 INT4/INT2 operations per clock.

We'll get to clock speeds in a bit, but these can vary quite a lot, particularly in mobile form factors. We suspect Intel will break 2GHz on desktop cards, but the mobile parts appear to top out at around 1.55GHz on the smaller chip and 1.65GHz on the larger chip. Do the math and the smaller ACM-G11 should have a peak throughput of over 3 TFLOPS FP32, with 25 TFLOPS of FP16 deep learning capability. The larger ACM-G10 will more than quadruple those figures, hitting peak throughput of 13.5 TFLOPS FP32 and 108 TFLOPS FP16.

Besides the vector and matrix engines, each Xe-core also contains a single ray tracing unit (RTU). Intel still hasn't provided any details of what sort of ray tracing performance we can expect from Arc, but even the largest chip will top out at 32 RTUs. In contrast, AMD's RX 6900 XT has 80 ray accelerators and Nvidia's GeForce RTX 3090 Ti comes with 84 RT cores that are each about 1.75X as fast as its first-generation RTX 20-series RT cores.

We know from our GPU benchmarks that Nvidia's current RT cores are much faster than AMD's ray accelerators, so unless Intel's RTUs are substantially more capable than the AMD and Nvidia equivalents, we don't expect a lot from Arc in terms of ray tracing performance. 

Intel does hope to make up for any lack in RT performance with the matrix cores. Similar to Nvidia's tensor cores, these can dramatically accelerate AI workloads, and Intel has its open-source XeSS (Xe Super Sampling) upscaling algorithm slated to launch this summer. Unlike Nvidia's DLSS, XeSS will also have fail-back support for DP4a instructions, which are included on Pascal and later Nvidia GPUs and RDNA and later AMD GPUs.

How it will fare in terms of performance and image quality remains to be seen, though Intel has offered up a video showcasing the quality of XeSS 1080p to 4K upscaling vs. native 1080p in the upcoming sci-fi game Dolmen.

Arc A-Series Mobile Graphics Solutions

With the hardware overview out of the way, here are the specific A-Series mobile solutions that Intel will be launching in the coming months. The Arc 3 A350M and A370M are available "now" — laptops with these GPUs should be on sale in early April. The midrange Arc 5 A550M basically doubles most of the hardware features, but with a much lower graphics clock. Finally, the Arc 7 A730M and A770M are two higher-end options, with the latter being the full-fat variant of the ACM-G10 chip.

It's a bit odd to see some of the other specs for the announced mobile chips. Both the Arc 3 solutions only use a 64-bit memory interface, which feels decidedly low-end. The ACM-G11 chip could have easily supported 6GB on a wider 96-bit interface, though that would have increased power consumption a bit. There's also a very large 400MHz gap between the A350M and A370M, which means the A370M should in practice be around 35% faster than the lesser A350M.

If anything, the A550M makes even less sense. With a graphics clock of just 900MHz, it has a theoretical performance level of just 3.7 TFLOPS FP32, only 16% faster than the A370M despite having twice the cores, memory, and bandwidth. The A730M bumps the core and memory aspects up 50%, with a 22% increase in clocks as well, giving it potentially 83% more performance than the Arc 5 chip. Finally, there's another big gap between the A730M and A770M, with the latter delivering up to twice the compute performance.

We can't say for certain that these are all the mobile Arc solutions we're likely to see, though there's clearly room for other configurations should there be demand for them. We also don't know without testing how Arc will actually stack up against the existing Nvidia and AMD GPU incumbents, so it might be better for Intel to just focus on a few models rather than a wide range of options, some of which are likely to be unpopular.
...

Continue Reading