Intel plans to make discrete graphics cards, and yes, that's exciting. But so far, it has maintained a tight seal on the details.
Digital Trends has obtained parts of an internal presentation of the Intel data center group that give us our first real glimpse of what Intel Xe is capable of (codenamed "Arctic Sound"). The presentation details the features that were in effect at the beginning of 2019, although Intel may have changed some of its plans since then.
These details show that Intel is serious about attacking Nvidia and AMD from every possible angle, and provides our best look at GPUs so far. The company clearly intends to go big with the new line of graphics cards, especially one with TDP(Power of thermal design, for its acronym in English), 500W, the most we have seen of any manufacturer.
Intel declined to comment when we contacted a spokesperson for an answer.
Intel Xe graphics use “tiled” chiplets, probably with 128 execution units each
Xe is the unifying architecture of Intel on all its new graphics cards, and the slides provide new information about Intel's design.
The documentation shows that Intel Xe GPUs will use "tiled" modules. Intel does not call these modules “chiplets” in the documentation, but the company revealed in November that its Xe cards will use a multi-matrix system, packaged together by Foveros 3D stacking.
The technique is similar to that developed by AMD in its Zen processors (which makes sense: consider who has hired Intel). But in graphics, this approach differs from how AMD and Nvidia cards are designed.
The mentioned cards include a GPU of a mosaic at the bottom of the stack, a two mosaic card and a maximized four mosaic card.
The documentation does not indicate how many units of execution (UE) will be included in each tile, but the mosaic counts are aligned with a mid-2019 controller leak that lists three Intel GPUs and their corresponding UEs: 128, 256 and 512. Assume that each tile has 128 UEs, the 384-UE configuration is missing. That aligns with the configuration of three missing tabs of the filtered slides we receive.
We assume that Xe use the same basic architecture shared by the previous Irisu Plus (Generation 11) graphics of the company, since Xe (Generation 12) arrive only one year later. The Intel Gen 11 GPUs contain a division, in which there were eight “sub-divisions,” each with eight UEs for a total of 64.
This is the basis of Xe, which begin to join several divisions into a single package. For the same math, a single mosaic will have two of these sectors (or 128 EU). The two-tile GPU, then, will present four slices (or 256 UE) and the four-tile GPU will get eight slices (or 512 UE).
Intel has been investing in multiple connection technologies that could allow these mosaics to work with high efficiency, known as EMIB (Embedded Multi-Die Interconnect Bridge), perhaps the "co-EMIB" mentioned by Intel last summer. It is a technology that Intel debut in the revolutionary but unfortunate Kaby Lake-G chips of 2018.
The thermal design power ranges between 75W and 500W
Intel has at least three different cards in process, with a TDP ranging from 75W to 500W. These numbers represent the full range of graphics, from basic level consumer cards to servers for data centers.
Let's address them one at a time, starting from the bottom. The base TDP of 75W to 150W applies only to cards with a single box (and, presumably, 128 UE). These seem the most suitable for consumer systems, and are aligned with the preview we have seen so far of a card called “DG1”.
At CES 2020, Intel revealed the DG1-SDV (software development vehicle), a discrete desktop graphics card. It did not have an external power connector, which indicated that it was probably a 75W card. That is a coincidence with the SDV card of a mosaic that appears first in the previous box.
It is difficult to say how much the 150W part of DG1-SDV could differ. However, a 150W TDP will be thermally in tune with the Nvidia RTX 2060 (160W consumption) and the AMD RX 5600XT (160W consumption, after a recent BIOS update).
Despite the striking design of the DG1 cover, Intel insisted that these were only for developers and software vendors. The cards listed in the table as “RVP” (reference validation platform) may be products that we expect Intel to sell to individuals and system developers. So far it is unknown how much the RVP and SDV will resemble each other, especially if Intel is preparing versions of these GPUs for both the consumer and the server.
Beyond these options, which can be correlated with consumer products, Intel seems to have more extreme Intel Xe graphics cards. Both need more power than a typical home PC can provide.
First is a two-tile GPU with a TDP of 300W. If this were sold to the players, it would easily exceed the energy consumption of current top-level graphics cards. Its TDP has a power of 50w more than the Nvidia RTX 2080 Ti, which already consumes a lot of energy.
Judging only by TDP, this product is likely to be a competitor to the 280W RTX Titan GPU or the 300W Tesla V100, the oldest data center card in Nvidia. It is likely that a part of the workstation complies with the second pillar in Intel's strategy, labeled "high power." Intel defines these as products made for activities such as media transcoding and analysis.
Intel's most powerful Intel Xe GPU may not be available as consumer electronics.
The real high-performance solution is the 4-tile graphics card, from 400W to 500W located at the top of the line. This model consumes much more energy than any consumer video card of the current generation, and more than the current data center, start-up cards.
This card also specifies 48V of power. That is only provided on a server's power supplies, which effectively confirms that the most powerful Intel Xe card does not appear as part of the consumer. The 48V power may be what allows Intel to reach 500W. While power supplies for more extreme games can handle a 500W video card, most cannot.
In November 2019, Intel announced “Ponte Vecchio,” what the company called the “first GPU exascale” for data centers. It is a 7 nm card that uses a series of technologies with chiplet connection to scale to that power.
It certainly seems like an option for this 4-piece 500W card, although Intel says that Ponte Vecchio will not come out until 2021. It was also reported at that time that Ponte Vecchio will use a Compute eXpress Link (CXL) interface through a PCIe 5.0 connection and a Foveros pack of eight chips.
Xe uses HBM2e memory and is compatible with PCIe 4.0
Rumors have circulated about Intel using expensive high-bandwidth memory (HBM) over conventional GDDR5 or GDDR6. According to our documentation, the rumors are true. The last important graphics card to use HBM2 was the AMD Radeon VII, although subsequent Radeon cards have been passed to GDDR6, such as Nvidia GPUs.
The documentation specifies that Xe use HBM2e, which is the latest evolution of technology. It aligns well with SK Hynix and Samsung's announcement that HBM2e parts will be released in 2020. The documents also detail how memory is designed, connect directly to the GPU package and use multiple 3D RAM arrays stacked on each other. others.
Although not mentioned, Xe is likely to use Foveros 3D stacking to interconnect between multiple matrices and also to bring the memory closer to the matrix.
The least surprising thing to confirm about Intel Xe is the PCIe 4 compatibility. The 2019 AMD Radeon cards are compatible with the latest PCIe generation, and we expect Nvidia to adjust to that as well in 2020.
GPUs will target each segment
Intel has created a unique architecture that scales from its integrated graphics for thin laptops to high-performance computing for data engineering and machine learning. That has always been the message from Intel, and the documentation we have received confirms it.
Discrete game cards are only a small part of the product line. The slides show Intel's plans for numerous uses that include multimedia processing and delivery, remote graphics (games), media analysis, augmented reality and immersive virtual reality, machine learning and high performance computing.
Where does this leave us?
It will be easy to take a look at these specifications and draw broad conclusions from them. But there are many things that we still don't know. It is too early to say if Intel's ambitious bet on discrete graphics cards affect AMD and Nvidia. We probably won't hear more official details about Xe until Computex 2020.
Even so, it is clear that Intel is not limping in the release of its first discrete graphics cards. The company compete with Nvidia and AMD in all key markets: entry level, midrange and HPC.
If it is possible to establish a foothold in only one of these three areas, Nvidia and AMD will have a strong new competitor. A third option apart from the current duopoly will involve better options at more aggressive prices. Who doesn't want that?