Intel's GPU Business AXG Split and Distributed to Client and Enterprise Businesses, Plans "Alchemist+" and "Battlemage" Launches Over Next 2 Years

[btarunr]

Intel in its Q4-2022 Financial Results presentation confirmed that it has split Accelerated Computing Group (AXG), the business that designs both client GPUs and scalar compute processors, and distributed its IP among the CCG (client computing group), and DCAI (data-center and artificial-intelligence group). CCG is Intel's largest breadwinner, and the group behind its Intel Core client processors. This business was receiving iGPU IP from AXG, and will now oversee all its client graphics portfolio. The DCAI group will handle the company's Data Center GPU lineup, with products such as "Ponte Vecchio."

Over 2023, Intel is expected to update its Arc Graphics series with the refreshed "Alchemist+" graphics architecture. Very little is known about "Alchemist+," except that it is rumored to introduced a more advanced third-party foundry node than its current TSMC 6 nm node; and while retaining the same IP as the current "Alchemist" GPUs, scale upward—either in terms of Xe core counts; or performance, by increasing clock-speeds. 2024 will see Intel introduce "Battlemage," its second client graphics architecture for discrete GPUs. "Battlemage" is expected to introduce new architectural changes, utilize an even newer foundry node, and significantly increase performance. Raja Koduri hinted in late 2022 that he still wants to build GPUs in the 200-300 W power-range, without implying that this constitutes "mid-range," so we'll have to wait and see how "Battlemage" improves performance/Watt over "Alchemist+."



Image Courtesy: VideoCardz

View at TechPowerUp Main Site

 

[Daven]

Rumble…rumble…rumble…

 

[Vayra86]

Oh wow. Raja figured out for the second (?) time that you shouldn't build a multi purpose GPU and transplant it for gaming.

The hard way.

 

[wNotyarD]

Alchemist+? Why?

 

[Vayra86]

Alchemist+? Why?

They think its viable with a shrink. Its fucking hilarious, still approaching shit software support with better hardware, as if they've got the former under control. They clearly want to salvage whatever they can, but this is an exercise in futility.

Still though, future plans do mean they have to keep at the driver updates, are they going to keep pace? There's enough work, at least I guess the idea is, as it was initially, to make the Intel dGPU fairy tale last long enough for us to forget older APIs exist or until they can just brute force past all the TDRs and blue screens or no signals? And that is with the weird assumption current day APIs are all working fine on their GPUs as well. Which they don't.

 

[Dr. Dro]

Alchemist+? Why?

Most sensible explanation is that Battlemage isn't ready (either silicon design/performance target or software), and they need to release something to make investors happy.

 

[TheinsanegamerN]

Most sensible explanation is that Battlemage isn't ready (either silicon design/performance target or software), and they need to release something to make investors happy.

Enterprise has long lifecycles for support, and alchemist already exists. The same companies that continued to buy ivy bridge laptops when skylake existed, and are the reason said chips are produced long after their replacements hit the market.

 

[Dristun]

Still though, future plans do mean they have to keep at the driver updates, are they going to keep pace? There's enough work, at least I guess the idea is, as it was initially, to make the Intel dGPU fairy tale last long enough for us to forget older APIs exist or until they can just brute force past all the TDRs and blue screens or no signals? And that is with the weird assumption current day APIs are all working fine on their GPUs as well. Which they don't.

They now have a flood of bug reports coming in daily via discord and github. Also in my experience the most broken is actually Vulkan, not older APIs. Though I have to add that at the current pace of 2-3 games fixed per release they won't catch up this year, and cutting staff to save money doesn't bode well either.
As for scaling — if they have a new node that will allow them to clock higher at the same power target (I know it doesn't always work like that!), A770 does gain nicely, which one can see right now by breaching the power limit thanks to broken drivers, lol. If they can get to Timespy-level optimizations across the board in at least new games, at 2.8ghz it's firmly in the highly-overclocked 3070 territory instead of competing with 3060 non-ti. And they're willing to sell this at 350$ right now while throwing AAA games into the deal. It's not too bad, but they really need to work on the drivers 24/7.

 

[Dr. Dro]

Enterprise has long lifecycles for support, and alchemist already exists. The same companies that continued to buy ivy bridge laptops when skylake existed, and are the reason said chips are produced long after their replacements hit the market.

Agreed but, they also have the gaming market to cater to, and I imagine that there is investor expectation regarding a return on the Arc GPUs.

 

[Assimilator]

LMAO. They've only released a single generation and they're already on the tick/tock ++++++++ node shrink nonsense to cover up the fact that the underlying microarchitecture isn't being improved. What a joke.

Enterprise has long lifecycles for support, and alchemist already exists. The same companies that continued to buy ivy bridge laptops when skylake existed, and are the reason said chips are produced long after their replacements hit the market.

Nobody in the enterprise market is buying Intel GPUs, integrated or discrete.

 

[Fouquin]

Nobody in the enterprise market is buying Intel GPUs, integrated or discrete.

HPE Cray is. In case you're out of the loop: Aurora supercomputer at Argonne.

 

[Daven]

HPE Cray is. In case you're out of the loop: Aurora supercomputer at Argonne.

There has been no official announcement of Ponte Vecchio nor have SKUs been released. No one can buy them and the mythical Argonne supercomputer has yet to turn up on top500.org. I’m sorry but to continue to use a single customer that might not even have received their entire order yet after two years is hardly evidence of anything. Unless you can point me to any other customer outside of Argonne for Ponte Vecchio I will continue to treat this as a custom one time solution with general market availability cancelled.

 

[psyclist]

Oh wow. Raja figured out for the second (?) time that you shouldn't build a multi purpose GPU and transplant it for gaming.

The hard way.

Exactly...how do you not learn after Vega? I had a Vega 64 (under a WB) and it was as fast as a 2070 in most games. But she would gobble power to get that performance. Silly that Raja decided, for my next trick...Ill just do the same. One trick pony? we shall see!

 

[Fouquin]

There has been no official announcement of Ponte Vecchio

Are you having a laugh? They were announced last year as the GPU Max Series and are officially launched as of January 10th. Is a direct link to Intel's ARK and newsroom official enough for you?

No one can buy them and the mythical Argonne supercomputer has yet to turn up on top500.org.

Unfinished computer cant be benchmarked because it's unfinished, more at 11.

Unless you can point me to any other customer outside of Argonne for Ponte Vecchio I will continue to treat this as a custom one time solution with general market availability cancelled.

Nothing is stopping HPE Cray from marketing the blades to other customers EXCEPT that availability of spare blades is contingent on Aurora being fully built. Shockingly if you have an initial exclusive supply contract for a product, you have to deliver that product to that customer before you open orders to others. Weird, I know.

 

[wNotyarD]

Are you having a laugh?

Yes. They got "released" only now (check the link you posted yourself) and no one has ever seen them performing.

 

[Fouquin]

Yes. They got "released" only now (check the link you posted yourself) and no one has ever seen them performing.

Yes they have been both announced AND released. Which is the polar opposite of, "There has been no official announcement of Ponte Vecchio nor have SKUs been released."

 

[wNotyarD]

Yes they have been both announced AND released. Which is the polar opposite of, "There has been no official announcement of Ponte Vecchio nor have SKUs been released."

I'll concede that they've been announced. Over and over and over again.
Their SKUs release? Paper launch at best.

 

[Scrizz]

LMAO. They've only released a single generation and they're already on the tick/tock ++++++++ node shrink nonsense

um... what do you think other companies do? sit and wait a couple years?

 

[Count von Schwalbe]

LMAO. They've only released a single generation and they're already on the tick/tock ++++++++ node shrink nonsense to cover up the fact that the underlying microarchitecture isn't being improved

To be fair, it is a "true" die shrink unlike all the Skylake derivatives. And the Vega 64 gained majorly from a die shrink as the Radeon VII, with little architectural changes.

 

[wNotyarD]

To be fair, it is a "true" die shrink unlike all the Skylake derivatives. And the Vega 64 gained majorly from a die shrink as the Radeon VII, with little architectural changes.

From Vega 64 to VII, more than shrinking from GloFo 14 to TSMC 7 which enabled higher clocks, there was double HBM capacity and bus and more than double bandwidth. So it wasn't only the shrink at play.

 

[eidairaman1]

Oh wow. Raja figured out for the second (?) time that you shouldn't build a multi purpose GPU and transplant it for gaming.

The hard way.

When is strike 3 for him?

 

[wNotyarD]

When is strike 3 for him?

When will Nvidia hire him?