Intel Dedicated GPU Discussion Thread

[SteelCrysis]

https://www.tomshardware.com/news/intels...orm-factor

As spotted by @momomo_us, Intel is preparing its Ponte Vecchio graphics cards for shipment in a reference validation platform (RVP), which are systems used by vendors to begin optimizing hardware and software. According to the listing at the Eurasian Economic Commission, the graphics cards come in three flavors and a standard AIC form factor.
...
Intel splits the Xe Architecture up into three designs that each address different segments: Data center, consumer graphics cards, and AI use-cases (HP); integrated graphics on its processors (LP); and the high-tier Xe HPC for high performance computing, with the latter (Ponte Vecchio) being designed specifically for compute. The consumer version of the Xe Graphics card for gaming will lead the way in 2020, likely on the 10nm process.

[SteelCrysis]

https://www.tomshardware.com/news/intel-...rious-chip

We're all a little curious about what's in store for us with Intel Xe Graphics, and we've been following the company's @IntelGraphics Twitter channel closely. The latest tweet gave a hint of what might be on the horizon:

The object of interest is the image of the chip next to the AA battery. The tweet itself doesn't reveal much about the chip, nor does it have any inscriptions on it to guide us. All we have to go by is that this image was tweeted by @IntelGraphics, and that it's a big chip. Also, Raja Koduri has previously shown a wafer with what appears to be massive Xe HPC die, which would coincide with this final chip package.

[Update: @IntelGraphics deleted a previous tweet showing Raja and Jim with masks and replaced it with the above tweet focused just on the chip.]
...
Watching the image do the rounds on Twitter, it's clear that nobody has a definitive answer as to what this chip is. As we note in our Xe Graphics article, however, unless Intel is trolling everyone, our best guess is that this is the data center Xe HPC Graphics variant. We count 2,688 pads on the LGA package, which is significantly more than sockets LGA2066 and LGA1200, but also fewer than the current Xeon chips use for FCLGA3647.

We can also get a good estimate of the total package size, thanks to the AA battery. The battery measures 49.2-50.5mm long and is 13.5-14.5mm in diameter. Based on that, the package is approximately the same area as six AA batteries, measuring about 80 x 52 mm. Or if you want to be more precise, the package is somewhere between 77.6 x 50 mm and 80 x 52 mm based on the relative sizes of the battery and package (plus or minus another 2mm). Given the size, it's again unlikely this is a consumer oriented product.

One thing is clear: Intel is taking the "go big or go home" mentality with this chip. Time will tell what it is exactly, what it can do, and whether it can compete with AMD and Nvidia when it comes to the best graphics cards.

https://www.techpowerup.com/266513/intel...-xe-hp-gpu

The Intel Graphics Twitter account was on fire today, because they posted an update on the development of the Xe graphics processor, mentioning that samples are ready and packed up in quite an interesting package. The processor in question was discovered to be a Xe-HP GPU variant with an estimated die size of 3700 mm², which means we sure are talking about a multi-chip package here. How we concluded that it is the Xe-HP GPU, is by words of Raja Koduri, senior vice president, chief architect, general manager for Architecture, Graphics, and Software at Intel. He made a tweet, which was later deleted, that says this processor is a "baap of all", meaning "father of all" when translated from Hindi.

Mr. Koduri previously tweeted a photo of the Intel Graphics team at India, which has been working on the same "baap of all" GPU, which suggests this is a Xe-HP chip. It seems that this is not the version of the GPU made for HPC workloads (this is reserved for the Xe-HPC GPU), this model could be a direct competitor to offers like NVIDIA Quadro or AMD Radeon Pro. We can't wait to learn more about Intel's Xe GPUs, so stay tuned.

[SteelCrysis]

https://www.techpowerup.com/267309/intel...eekbenched

Intel's ambitious Xe graphics architecture is expected to make its first commercial debut as an iGPU that's part of the company's 11th gen Core "Tiger Lake" mobile processors, but it already received a non-commercial distribution as a discrete GPU called the DG1, with Intel shipping it to its independent software vendor ecosystem partners to begin exploratory work on Xe. One such ISV paired the card with a Core i7-8700 processor, and put it through Geekbench. While the Geekbench device identification doesn't mention "DG1," we lean toward the possibility looking at its 96 EU configuration, and 1.50 GHz clock speed, and 3 GB memory.

The Geekbench run only covers OpenCL performance of the selected device: "Intel® Gen12 Desktop Graphics Controller." The total score is 55373 points, with 3.53 Gpixels/s in "Sorbel," 1.30 Gpixels/sec in Histogram Equalization, 16 GFLOPs in SFFT, 1.62 GPixels/s in Gaussian Blur, 4.51 Msubwindows/s in Face Detection, 2.88 Gpixels/s in RAW, 327.4 Mpixels/s in DoF, and 13656 FPS in Particle Physics. These scores roughly match the 11 CU Radeon Vega iGPU found in AMD "Picasso" Ryzen 5 3400G processors.

[SteelCrysis]

https://www.tomshardware.com/news/intel-...uan-xe-gpu

Remember two months ago when Intel tweeted an image of a huge, mysterious chip? Well, we still don't know exactly what that chip is, but Intel is back again teasing something even bigger. Well, rather, Intel's Raja Koduri is here, showing off his (we quote) "big 'fabulous' package."

Of course, we still don't know exactly what this chip is, but there is room to speculate. Nor the last chip, nor this one are at all likely to be products that could also show up in non-socketed consumer graphics cards, competing with AMD and Nvidia in the ultra-high-end space of the best graphics cards. With chips this big, they're bound to be HPC products aimed at scientific and data center use, where performance vs. cost becomes a carefully balanced equation that absorbs extreme costs for individual parts.

It looks to measure about 80 by 80 mm (with a margin of error), and we would expect a huge GPU, along with numerous HBM2 or HMB2e stacks to be lurking on the interposer, underneath the heatspreader.

This huge chip is undoubtedly a creation based on Intel's Xe graphics architecture (as proven by this tweet showing the huge chip under water). The amount of silicon underneath that heatspreader would be so huge, it would no longer be economically viable for consumers in any way, shape, or form -- not that Xe graphics is meant to compete in the consumer space anyway.

Looking at the image of the GPU under water, you’ll notice the sheer number of springs keeping the cold plate pressed against the die surface. This shows that bowing becomes a huge issue when chip packages get this big. The cooler also mentions ATS-4T at the top (as spotted by Komachi Ensaka), possibly referring to Arctic Sound - 4 tiles.

Rumor has it that the big HPC Xe GPUs will come with tiled layouts, featuring either one, two, or four performance tiles. Based on the sizes, we reckon this is the biggest Xe GPU of them all with all four tiles, with Koduri's last tease showing off the two-tile product.

Nevertheless, only time will tell what this massive, socketed GPU really is, and who it is for. For all we know, Intel is simply showing off what it can build, instilling fear in both AMD and Nvidia.

https://www.techpowerup.com/269016/intel...us-package

Raja Koduri likes to show off his work in GPU engineering, way back from his days at AMD. His latest bragging rights is what he calls the "BFP" (big fabulous package), a large GPU package that's larger than even an AMD SP3 package, roughly 75 mm x 80 mm in size, looking at its "family" photo where a AA battery is used for scale. This could very well be the largest version of Xe "Ponte Vecchio." Koduri's caption reads "And..they let me hold peta ops in my palm (almost)," likely referring to the chip's tensor ops performance. He also posted pictures of Xe's various test platforms, from Intel's labs in Folsom, California.

[SteelCrysis]

https://www.techpowerup.com/270251/intel...management

It's interesting to note that Intel was specific about "CPU" when talking about 7 nm, meaning that Intel's foundry woes only affect its CPU product stack, and not a word was mentioned in the release about the company's discrete GPU and scalar compute processors that are being prototyped and validated. This is probably the biggest hint we'll ever get from Intel that the company's dGPUs are being designed for third-party foundries (such as Samsung or TSMC), and that the Xe dGPU product roadmap is disconnected from that of Intel's fabs.

[SteelCrysis]

https://www.tomshardware.com/news/intel-...in-20-days

Intel tweeted via its Intel Graphics Twitter account yesterday that the chipmaker will provide more information on Xe graphics in 20 days. Intel has since deleted the tweet, but we think the promise is still on.

There's a bit of mystery around Intel's tweet. The company has adopted the Xe moniker for both its integrated graphics and upcoming discrete graphics cards, but it's uncertain which product segment the chipmaker was referring to. Intel shared the tweet yesterday so the big day should be on August 13. Given the timing of the tweet and Intel's recent struggles with its 7nm and 10nm nodes, the event will likely be a presentation rather than a formal product launch.

Intel actually has a couple of events planned for the upcoming months. David Blythe, Senior Fellow and Director of Graphics Architecture at Intel, will deliver a speech about the Xe GPU architecture at Hot Chips this year on August 17. That's the closest event that coincides with Intel's tweet. It's plausible that we might get some new and juicy details on Intel's Xe graphics cards after all.

[SteelCrysis]

https://www.tomshardware.com/news/raja-k...-hot-chips

Over the past several months, Intel's Raja Koduri has been slowly revealing portions of the upcoming Xe Graphics architecture and lineup. The Xe family will span everything from low power integrated and entry level graphics solutions under the Xe LP Graphics brand, up through data center multi-chip solutions with die stacking. The latter is we're talking about.

As discussed last week, Intel's Xe HP Graphics will come in three variants. The base model, which has been shown several times, has a single tile with 512 EUs (Execution Units) and most likely two HMB2e stacks. Intel hasn't confirmed exact specs, but it did show performance scaling for a computational workload of the 1-tile, 2-tile, and 4-tile variants:
...
Lest anyone doubt that the 4-tile GPU doesn't actually exist and is merely a publicity stunt, Raja whipped out the large package and briefly flashed it at the camera during his Hot Chips presentation. And yes, it's really big — much bigger than any other chip package we've seen.

Whether the 4-tile Xe HP will ever be put into production, or if it's merely a test product while Intel prepares Xe HPC, aka Ponte Vecchio, is a different matter.

[SteelCrysis]

https://www.tomshardware.com/news/intel-...benchmarks

Intel Iris Xe Max graphics have shown up in some leaked benchmarks, and also indicate the possibility that Xe Max is the name for DG1, Intel's first discrete GPU.

In a SiSoftware entry, the GPU is shown with 96 execution units and 768 streaming processors. The benchmark also shows a clock speed of 1.55 GHz, a 1MB L2 cache, and 3GB of VRAM.

Also interesting is that the benchmark record shows the laptop used for testing is a Coffee Lake-based Intel reference platform. If it's a validation system, the CPU name may simply be wrong, but it's also a good possibility that this Iris Xe Max is DG1, a discrete GPU, and it wouldn't matter which CPU it's used with.
...
It seems most likely that Iris Xe Max graphics will debut alongside Tiger Lake-H series processors, but we'll find out for sure whenever it officially launches.

[SteelCrysis]

https://www.tomshardware.com/news/intels...xe-hp-gpus

Intel's Chief Architect Raja Koduri tweeted the first images of the company's new Xe HP GPUs that that are currently sampling to Intel's customers. The cards, built on the 10nm SuperFin process, are also being used as the software development vehicle for the Aurora Supercomputer. The exascale-class supercomputer has been delayed due to Intel's struggles with the 7nm node used in its Ponte Vecchio GPUs, but due to similarities between the cards, the new Xe-HP cards can be used to develop software for Aurora. The image above shows the already-announced XG310 Intel data center GPU flanked by the new Xe-Hp models.

[SteelCrysis]

https://www.tomshardware.com/news/intel-...-at-14-ghz

Listings in two benchmark databases reveal that Intel, or perhaps a company close to the chip giant, seems to be testing a rather mysterious unannounced graphics processor that has slightly more compute horsepower than the company's Iris Xe Max discrete GPU for notebooks.

Someone has submitted benchmark results of what is called the Intel Gen12 Desktop Graphics Controller featuring 128 execution units (EUs), clocked at 1.40 GHz, and equipped with 3.03GB of memory. Given the 'Gen12' designation, the GPU is based on Intel's Xe-LP microarchitecture as other Xe microarchitectures are referred to as Gen12.x in the drivers. Since Intel's Xe-LP EUs are 8-wide, 128 EUs may be compared to 1024 stream processors of AMD or Nvidia GPUs.

So far, benchmark results of the Intel Gen12 Desktop Graphics Controller with 128 EUs have been submitted to Geekbench 5 and SiSoftware databases, as found by TUM_APISAK/Twitter. The part scored 9,311 points in Geekbench 5 and 82 points in Sandra.
...
The Iris Xe Max with 96 EUs is advertised as a standalone graphics processor for laptops and Intel has never disclosed plans to make it available on the desktop market. Meanwhile, when Intel started its 'Odyssey' discrete GPU marketing campaign back in 2018, it clearly announced plans to offer graphics cards for desktops in 2020 and even demonstrated a small Xe Graphics DG1 development board at CES 2020. Keeping these facts in mind, we can speculate that the Intel Gen12 Desktop Graphics Controller with 128 EUs could be an entry-level desktop GPU.

Assuming that the Intel Gen12 Desktop Graphics Controller with 128 EUs runs at up to 1.4 GHz, its compute performance will be around 2.86 TFLOPS, about 13% higher when compared to that of the Iris Xe Max running at 1650 MHz (~2.53 TFLOPS). That said, it is hard to expect such a discrete GPU to beat gaming graphics cards running chips from AMD or Nvidia. Meanwhile, Intel could market it as a co-GPU for certain tasks in some niche markets.

Intel has not officially announced an Iris Xe-LP GPU with 128 EUs. In fact, it is not even guaranteed that the product will actually be released. Nonetheless, it looks like Intel has something in the works.

https://www.techpowerup.com/276100/intel...-geekbench

Intel is still in the work of testing and certification for their more complex graphics products based on their Xe microarchitectures, and that means that some tests are being done in well-known benchmarking platforms. Case in point, an Intel Xe GPU with a reported 128 EUs (1024 shading units), 3 GB of memory, and a clockspeed of 1.4 GHz ran through Geekbench, where it scored an... interesting 9,311 points in the OpenCL test.

This is more likely than not an engineering sample, considering that Intel's Iris Xe MAX wrapped in its Tiger Lake package can score up to 23,000 points. It is currently unclear if this particular Xe manifestation is running on Intel's Xe-LP or Xe-HPG architecture. This might be Intel's DG-2 product, which offers higher performance than their DG-1 discrete graphics card that is only available for system integrators.

[SteelCrysis]

https://www.extremetech.com/computing/31...ed-version

The first thing I want to say here is that the Intel Iris Xe Max is not marketed as a gaming GPU, even though it’s a discrete solution. It is, however, identical to the well-received 11th Gen integrated GPU, and should therefore have at least some basic gaming capabilities. Our sister site PCMag recently took a laptop equipped with an Iris Xe Max for a spin, and what they found was rather odd.

The Iris Xe Max is Intel’s first discrete GPU in many years and it sports 768 cores, 48 TMUs, and 24 ROPS (768:48:24). It uses LPDDR4X to keep power consumption down, instead of GDDR, and it offers a 128-bit memory bus, with a 4GB RAM buffer. Intel’s latest full-fat integrated GPU, the Xe Graphics G7, is built into chips like the Core i7-1185G7, also have a 768:48:24 configuration.
...
These results are odd. We’re comparing two GPUs based on identical GPU cores. The discrete card has every advantage — dedicated RAM, higher core clocks, higher overall memory bandwidth — and yet it can’t win. Why not?

Let me say upfront that the answer might be something banal, like “Dell shipped the laptop with an early graphics driver,” or “A firmware update improved CPU and GPU performance by tweaking throttle points.” A more interesting possibility is that we’re seeing the impact of differences in power management and thermal targets between the Dell and the Intel whitebook, or, a performance-limiting characteristic of LPDDR4X. The penalty would have to be enormous — the Intel whitebook has a large lead in some of these tests. This seems unlikely.
...
When it comes to OEM laptops, Intel allows manufacturers to target a wide range of skin temperatures and thermal targets. Many of Intel’s chips can be configured for a TDP of 12W – 28W depending on how the manufacturer wants to position a specific system. Manufacturers also have wide latitude in terms of the cooling solutions they build into their hardware and the degree to which they specifically optimize component placement to ensure the fastest system with the least amount of throttling possible.
...
This PCMag benchmark from Ice Lake’s debut makes the point. Two of the three Core i7-8565U-powered machines score almost identically. The third is much faster. It’s possible that Dell picked less aggressive thermal and power targets than Intel did, either for the CPU, GPU, or both. It’s also possible that Intel’s Core i7-1185G7 is more aggressive when it comes to maintaining boost clocks and runs faster in 720p for that reason. I checked other reviews on Tiger Lake laptops, and the GPU performance on those systems was only a little slower (1-2 fps) than the whitebook.

One of the reasons I suspect there might be a power or thermal issue in play here is because of what happened back when Intel launched Core M. There were three Core M CPUs to choose from, and folks soon noticed that the lowest-end CPU sometimes ran faster than the higher-end chips. While the higher-end chips had a higher boost clock, they hit their thermal trip points much more often and had to throttle back. It’s possible that the dGPU is either ramping up and down, or that other thermal limits in the chassis keep its overall clock rate lower.
...
This is not a terribly strong debut for the Iris Xe Max. Intel wants to sell this card as an AI and machine learning accelerator, and it has a technology called Deep Link built into the chip that will allow the iGPU and dGPU to team up on certain workloads. At the same time, Tiger Lake’s biggest performance jump is in graphics, where Intel managed to take the overall performance lead from AMD. It would be incoherent to simultaneously praise Tiger Lake’s integrated performance while pretending the Iris Xe Max isn’t built from the same silicon and shouldn’t offer at least the same performance.

Hopefully, there’s a combination of improvements that will at least bring the performance of the Iris Xe Max up to match the Core i7-1185G7’s baseline. Intel isn’t marketing the Iris Xe Max as a gaming GPU, and we can’t evaluate it as an AI/ML accelerator using Deep Link yet, because the feature isn’t supported in any software, but customers who buy Iris Xe Max silicon shouldn’t be penalized with slower GPU performance compared with integrated users, even if that’s a relatively modest amount in the grand scheme of things.

[SteelCrysis]

https://www.techpowerup.com/276811/intel...8-gb-gddr6

Intel's return to discrete gaming GPUs may have had a modest beginning with the Iris Xe MAX, but the company is looking to take a real stab at the gaming market. Driver code from the latest 100.9126 graphics driver, and OEM data-sheets pieced together by VideoCardz, reveal that its next attempt will be substantially bigger. Called "DG2," and based on the Xe-HPG graphics architecture, a derivative of Xe targeting gaming graphics, the new GPU allegedly features 512 Xe execution units. To put this number into perspective, the Iris Xe MAX features 96, as does the Iris Xe iGPU found in Intel's "Tiger Lake" mobile processors. The upcoming 11th Gen Core "Rocket Lake-S" is rumored to have a Xe-based iGPU with 48. Subject to comparable clock speeds, this alone amounts to a roughly 5x compute power uplift over DG1, 10x over the "Rocket Lake-S" iGPU. 512 EUs convert to 4,096 programmable shaders.

A leaked OEM data-sheet referencing the DG2 also mentions a rather contemporary video memory setup, with 8 GB of GDDR6 memory. While the Iris Xe MAX is built on Intel's homebrew 10 nm SuperFin node, Intel announced that its Xe-HPG chips will use third-party foundries. With these specs, Intel potentially has a GPU to target competitive e-sports gaming (where the money is). Sponsorship of major e-sports clans could help with the popularity of Intel Graphics. With enough beans on the pole, Intel could finally invest in scaling up the architecture to even higher client graphics market segments. As for availability, VideoCardz predicts a launch roughly coinciding with that of Intel's "Tiger Lake-H" mobile processor series, possibly slated for mid-2021.

[SteelCrysis]

https://www.techpowerup.com/277134/intel...-n7-report

Intel's first discrete gaming graphics card based on the Xe-HPG graphics architecture, will be built on a TSMC 7 nanometer silicon fabrication node, according to a Reuters report citing sources "familiar with the matter." The first such discrete GPU is being referred to internally by Intel as the DG2. Recent reports suggest that Intel will give the DG2 formidable specs, such as 4,096 unified shaders across 512 execution units, and 8 GB of GDDR6 video memory. Back in 2020, the company launched the DG1 under the Intel Iris Xe MAX marketing name, targeting only the mobile discrete GPU market. The DG1 has entry-level specs, with which Intel is eyeing the same pie as NVIDIA's fast-moving GeForce MX series mobile GPUs. Interestingly, the other major client of TSMC-N7 following Apple's transition to N5, is Intel's rival AMD.

[SteelCrysis]

https://www.extremetech.com/computing/31...d-7nm-node

According to Reuters, Intel intends to manufacture DG2 — its upcoming consumer graphics card — at TSMC, on an enhanced 7nm node that hasn’t even been named yet. TSMC has sold 7nm in three flavors — N7, N7P, and N7+. N7P was the original N7 node with additional performance enhancements, while N7+ introduced EUV lithography. The introduction of EUV was a major step all on its own. TSMC may have built a new 7nm node for its own purposes; the company occasionally introduces new variants of mature nodes. 7nm is still close enough to the leading edge to plausibly be refined in some fashion.

Alternately, Intel might have paid TSMC to implement a specific version of the node that suits its own goals. Reuters claims the node will be more advanced than Samsung’s 8N, which Nvidia tapped for Ampere.

[SteelCrysis]

https://www.extremetech.com/gaming/31953...n-20-years

Intel is finally back in the desktop graphics business, at least if you squint. The company has announced a partnership with certain PC OEMs to bring DG1 silicon to specific pre-built systems.

This isn’t exactly a full-on desktop graphics launch — that’ll come later in 2021 with the launch of DB2 — but Intel is still indisputably shipping at least a handful of discrete GPUs in the low end of the desktop market, for the first time in more than 20 years. The company announced it had partnered with “two ecosystem partners, including Asus” in its initial PR, but LegitReviews thinks the GPU featured in the image above is manufactured by Colorful. The other DG1 card identifies itself, ships without a fan, and is clearly an Asus-branded product.

Unfortunately for anyone hoping to play around with Intel’s latest desktop card and/or hoping to find a low-end GPU at a reasonable price, this GPU is OEM-only. Sometimes, OEM-only products will surface on secondary markets like eBay, but that’s not going to happen in this case. According to Intel, these GPUs will only work on very specific systems. LegitReviews inquired on this point and was told:
...
So, that’s that, then. It’s not clear why a motherboard would require a special UEFI to use a new GPU. Presumably DG2, when it arrives, will not have this problem.

It’s interesting to see Intel back in the graphics market because it’s been so long since we had an actual three-way fight. Once 3dfx died, the only company to offer any kind of competition to the ATI/Nvidia duopoly was PowerVR with the Kyro and Kyro II. While these GPUs were an interesting alternative to the Radeon and GeForce product lines, they did not find mainstream success and faded from the market.
...
It’s not unfair to be skeptical of Intel’s ability to launch a competitive GPU. We don’t even necessarily expect Intel’s first-generation cards to be all that great, objectively speaking. So long as they’re good enough to get a little traction somewhere in the market, Intel has an opportunity to iterate and improve the design. With a third player on the field, both AMD and Nvidia are presented with new challenges — but also, potentially, with new opportunities depending on how Intel’s presence impacts user GPU purchases.

All of this hinges on Intel building competitive products and being willing to stay the course over the long term. Incoming CEO Pat Gelsinger may have his own ideas about where to take the company. These DG1-equipped systems aren’t going to make a huge splash in the wider market, but we should know how effectively DG2 will compare against AMD and Nvidia before the end of the year. Given how hard AMD and Nvidia cards have both been to source, Intel could spin a modest GPU into a smash hit just by shipping it on-time, at MSRP.