Leak: Intel is Planning a 400-500W Top-End GPU to Challenge AMD, Nvidia
For more than a year, Intel has kept the exact details of its next-generation Xe GPUs tightly under its vest. We’ve gotten a few updates, like those about Ponte Vecchio, but the company has mostly let its advancing integrated graphics speak for itself (Ice Lake, which is based on Xe’s immediate predecessor, offered substantially improved graphics performance compared to Skylake).
A new leak from Digital Trends, however, claims to have a good deal more information about Intel’s graphics plan. According to their data, Intel is planning to bring cards to market based on several designs. The cards will feature either 1, 2, or 4 tiles, with each tile apparently constructed from 128 EUs. Each EU appears to operate on eight threads simultaneously, which means we compare Intel threads when comparing the total number of GPU cores to an AMD or Nvidia chip. That works out to 1024, 2048, or 4096 threads, in a GPU 128, 256, or 512 EUs wide. Intel will apparently use a tiled architecture to hit these kinds of densities. Tiled, in this sense, appears to be a reference to the physical hardware design, not a reference to tiled rendering.
This slide refers to the form factor, size, power consumption, and form factors that Intel will be launching around Xe. Not all the terminology could be defined, however. SDV = Software Development Vehicle. RVP = Reference Validation Platform. Wilson City seems to be a new type of Intel PCH, and Sawtooth Pass is an existing 2S Intel Xeon solution. It looks as though those 400-500W cards will be server parts only, however — there’s no sign Intel intends to ship those in consumer form factors and they require 48V input voltage. These are clearly server parts, at least for now.
It’s still impossible to draw any conclusions about what performance on these cards will look like, but we know Intel is willing to push the outer edge of the thermal envelope to hit its performance targets. The 300W target doesn’t establish that — AMD has also been willing to ship in these kinds of brackets — but 500W, even for a high-end server card, is really pushing it.
Problem is, you could use that fact to make two different arguments: 1). That Intel can’t match Nvidia / AMD power efficiency and will therefore rely on maximum power draw in HPC workloads, or 2). That Intel is willing to ship a part at maximum power draw to make a name for itself and its high-performance card. Either of those (or both) is a valid argument, based on what we know right now.
The leak claims to confirm additional facts, including Xe’s use of PCIe 4.0 (implying we’ll see Intel support the standard at some point before it launches the hardware) and its HBM2 support. HBM2 is a bit less surprising in this segment, and the GPU will use HBM2e. DigitalTrends think Intel will use Foveros 3D stacking for the memory support, which seems quite possible given how the company is advancing its interconnect technology.
I’m genuinely curious to see what Intel brings to the table, as far as new 3D graphics capability. Ice Lake’s GPU is significantly better than anything the company has shipped before and I’d love to see new competition in the PC GPU space. It seems unlikely that Intel’s 400-500W GPU will hit the consumer space, but stranger things have been known to happen.