Intel Arc Xe3 "Celestial" GPU Reaches Pre-Silicon Validation, Tapeout Next

[AleksandarK]

In December, we reported that Intel's next‑generation Arc graphics cards, based on the Xe3 "Celestial" IP, are finished. Tom Petersen of Intel confirmed that the Xe3 IP is baked, meaning that basic media engines, Xe cores, XMX matrix engines, ray‑tracing engines, and other parts of the gaming GPU are already designed and most likely awaiting trial fabrication. Today, we learn that Intel has reached pre‑silicon validation, meaning that trial production is imminent. According to the X account @Haze2K1, which shared a snippet of Intel's milestones, a pre‑silicon hardware model of the Intel Arc Xe3 Celestial IP is being used to map out frequency and power usage in firmware. As a reminder, Intel's pre‑silicon validation platform enables OEM and IBV partners to boot and test new chip architectures months before any physical silicon is available, catching design issues much earlier in the development cycle.

Intel provides OEMs and IBVs access to a secure, cloud‑based environment that faithfully emulates hardware‑representative systems, allowing developers to validate firmware and software stacks from anywhere without the need for physical labs. Most likely, Intel is running massive emulations of hardware on FPGAs, which act as an ASIC chip—an Arc Xe3 GPU in this case. The pre‑silicon validation team is now optimizing the power‑frequency curve and the voltage in sleep, rest, and boost states, as well as their respective frequencies. With the Xe3 IP taking many forms, engineers are experimenting with every possible form factor, from mobile to discrete graphics. Additionally, data pathways depend on these frequency curves, which in turn rely on power states that allow voltage to spike up and down as the application requires. As this work is now complete, engineers are moving on to other areas for optimization, and once the silicon returns from volume production, it will be fully optimized. We expect the first trial of silicon soon, with volume production by the end of the year or in early 2026.



View at TechPowerUp Main Site | Source

 

[lexluthermiester]

What most of us what to know is when the newer higher end ARC GPU's will be coming to market.

 

[boppingXQWB]

Really excited to see what Intel will pull off with Celestial. Alchemist was pretty bad, but Battlemage is actually OK. Celestial on par (perf/silicon and perf/power) with AMD maybe hopefully pretty please?

 

[ZoneDymo]

Im really looking forward to this, high expectations (and yes as always, the price matters but man this COULD be such a win for Intel....which probably means they will screw it up where its just on par with the rest in price performance)

 

[Darmok N Jalad]

I guess that means the GPU division lives to frag another day. I kinda wondered what Tan was going to do here, because while dGPUs are a huge market, it's been a massive money loser for Intel so far. No way Intel made money on its investment into Battlemage with the pricing B570/580 has.

 

[john_]

It's good that they didn't abandoned GPUs. I mean, it would have been suicidal if they where doing something like that.
But Intel needs it's manufacturing to become REALLY competitive. When they have healthy manufacturing behind their products, they will be in a position to start a price war with AMD mostly and Nvidia in a lesser degree and offer really great VFM products to the sub $400 market.

I guess that means the GPU division lives to frag another day. I kinda wondered what Tan was going to do here, because while dGPUs are a huge market, it's been a massive money loser for Intel so far. No way Intel made money on its investment into Battlemage with the pricing B570/580 has.

Intel can't navigate the future without good GPUs. The CPU is not as important as it was in the past. And if you see Apple with it's huge M# APUs, AMD's HALO, Nvidia's Digits and probably in a couple of years something similar from Qualcomm for example, Intel will have no chance without a good GPU architecture.
I was saying it from the Alchemist era when people where sure that Intel will pull the plug, to avoid losing billions.

 

[Denver]

This could be intel's do-or-die moment in the dGPU market. If its next-gen offerings fail to be both competitive and profitable—a scenario I find highly unlikely—the company will likely scale back its dGPU division, retaining only the resources necessary to continue advancing its iGPUs.

 

[rattlehead99]

What most of us what to know is when the newer higher end ARC GPU's will be coming to market.

Most people don't buy high-end GPUs. And in general GPUs above the rtx 5080 are just not good for the market as a whole.

 

[Launcestonian]

In light of this news, guess no B750 or 770 coming anytime soon.

 

[lexluthermiester]

And in general GPUs above the rtx 5080 are just not good for the market as a whole.

Are you kidding?

when the newer highER end ARC GPU's

Yup, that's what I said.

In light of this news, guess no B750 or 770 coming anytime soon.

Sadly, that might be the case.

 

[Apocalypsee]

That's unexpected, I thought we see B750/B770 before seeing XE3. Anyhow I really hope they tweak the driver so it won't be so much of an overhead on lower end CPU/platform which where is their target market is.

 

[kondamin]

Intel can't navigate the future without good GPUs. The CPU is not as important as it was in the past. And if you see Apple with it's huge M# APUs, AMD's HALO, Nvidia's Digits and probably in a couple of years something similar from Qualcomm for example, Intel will have no chance without a good GPU architecture.
I was saying it from the Alchemist era when people where sure that Intel will pull the plug, to avoid losing billions.

they need to do something that changes the need for that memory you can't upgrade post purchase for these massive igpu's before that's going to gain serious traction

 

[john_]

This could be intel's do-or-die moment in the dGPU market. If its next-gen offerings fail to be both competitive and profitable—a scenario I find highly unlikely—the company will likely scale back its dGPU division, retaining only the resources necessary to continue advancing its iGPUs.

They will keep trying or they will end up with a platform, after 5-10 years, that will be awful for anything that needs 3D graphics. Nvidia will be pushing it's ARM platform, Nvidia and/or AMD could start making their GPUs work much better with their own platforms, by introducing a proprietary interface that accelerates their GPUs, I mean there are scenarios where GPUs take over and Intel without discrete GPUs are at the mercy of others.

That's unexpected, I thought we see B750/B770 before seeing XE3. Anyhow I really hope they tweak the driver so it won't be so much of an overhead on lower end CPU/platform which where is their target market is.

Probably they are doing what AMD is doing. AMD knows, from the RX 7000 series, that customers willing to pay over $800 will be going to Nvidia, so they ignore that market. At the same time Intel probably knows that people willing to pay over $300 will go for an Nvidia or an AMD card, so they stay under $300. Again for Intel, I believe the key is their manufacturing. If they start making GPU dies at half the cost others pay TSMC, without capacity restrains, they can flood the market at prices that AMD wouldn't be able to touch and Nvidia wouldn't be willing to touch.

they need to do something that changes the need for that memory you can't upgrade post purchase for these massive igpu's before that's going to gain serious traction

There are tricks, like SidePort memory ATi was using about 20 years ago and I really really really don't understand why they don't bring this back.
SidePort: On-Board GPU Memory - ATI Radeon Xpress 200: Performance, PCI Express & DX9 for Athlon 64
AMD could create a custom dimm slot and start selling GDDR5/6 VRAM dimms for example. If they could do it 20 years ago, they will be able to do it in a much better form today.

 

[Darmok N Jalad]

In light of this news, guess no B750 or 770 coming anytime soon.

I don't know if we'll ever see them. It would depend on if XE3 is ready soon enough. The initial effort for the high-end BM didn't pan out, from what I gather, so they would have had to do a redesign. Considering how long that takes, it might not be worth the time, effort and money.

 

[Wirko]

Ironically, a GPU in a "pre-silicon" state lives entirely in silicon. Lots of silicon. Or does Intel start the development with a wooden model?

 

[kondamin]

Ironically, a GPU in a "pre-silicon" state lives entirely in silicon. Lots of silicon. Or does Intel start the development with a wooden model?

They switched to clay after the node shrink to 10nm as the wood grain was starting to cause issues.

 

[Darmok N Jalad]

Ironically, a GPU in a "pre-silicon" state lives entirely in silicon. Lots of silicon. Or does Intel start the development with a wooden model?

By that you mean designed on a computer? Because it’s not been fabricated yet. Tape-out is next and that leads to engineering samples.

 

[nageme]

In light of this news, guess no B750 or 770 coming anytime soon.

I wonder what's up with that. B570 was only released 3-4 months ago.
If final C* chips are expected mid-late 2026, there's still time and reason for B7*.

 

[RandallFlagg]

Are you kidding?

Steam survey backs that up in spades.

If we define high end as X80 and higher, *only* going down to the *first* X80 series cards (a 3080), 47.07% of the market is spoken for before the *first* high end 3080 shows up. And that 3080 only commands 1.83% on steam. I then count 8 more low and midrange cards before the next high end card shows, a 4090 with 0.97%

Just on an offhand guess based on the pattern, you're likely talking about 90%+ of the market being made up of less than an X80 series card.

 

[lexluthermiester]

Steam survey backs that up in spades.

See edit. I should have only quoted part of the comment.

Just on an offhand guess based on the pattern, you're likely talking about 90%+ of the market being made up of less than an X80 series card.

100% agree.