News Intel GPUs - waiting for B770

[Aapje]

There is a much bigger risk that a product targeting the high-end will run into a big bottleneck and will have to be sold for very little, than a lower-end product. There is inherently less risk in making a small die that costs less.

The A770 was arguably already too ambitious, requiring 406 mm2 of N6 to achieve the same results as 237 mm2 of N7 (Radeon 6650 XT). For Intel, it's more about creating a proper architecture that works and then they can think about scaling it up to larger sizes.

Don't forget that AMD effectively did the same thing by limiting RDNA 1 to the 5700 XT tier and only scaling up for RDNA 2 and 3.

I personally think that best case for Intel Arc is if they sell Battlemage without a loss and then start making a profit with Celestial. And that they gradually increase the size of the chips.

 

[DavidC1]

In this case you are not talking about them taking a x80/90 class and making them x70. x70 chips are still profitable.

By the time it's released they are taking what may be a x60 Ti to a x50. The fixed costs start dominating at the price range and margin gets really low.

And there's another factor to consider. @Exist50 said their GPUs have moved to a 2 year schedule when the revenue collapsed. So the successor is basically 2027.

 

[Mopetar]

Since both AMD and NVidia seem to be allergic to releasing anything for under $200, Intel could actually gain some market share and good will by making smaller GPUs targeted at the lower end of the market. Even some people who'd normally stick with AMD/NV might grab a low cost Intel card just to mess with it as a spare.

Until they can execute making smaller GPUs, there's not much point to them making larger ones. They still have a ways to go on the software front and it's difficult to ask anyone but your most diehard loyalists to fork over $700+ for something that's just a bit above the beta phase.

Their biggest problem seems to be the arrogance of thinking that because they're Intel they should just be on top of the market. They lack anyone who will figure out where they could grow market share and build a base and just have people that will bang their head against a wall trying to act like they own the market.

 

[igor_kavinski]

Either way Aurora is half the efficiency of Frontier in HPL. Just wild.

Intel should swallow their pride and do the following configuration changes in Aurora:

Disable HT.

Clock the CPUs and GPUs down to the efficient part of the v/f curve.

Create better drivers/software to eek out more performance from their hardware.

Stop trying too hard to beat AMD. I mean, come on. Now it's just embarrassing. Just admit that you have different strengths and capitalize on that, instead of trying to hide your weaknesses.

Yes, Aurora may become pretty pedestrian in performance this way but at least it will be efficient enough to not be a total loss in terms of long term costs of power and cooling.

 

[igor_kavinski]

Their biggest problem seems to be the arrogance of thinking that because they're Intel they should just be on top of the market. They lack anyone who will figure out where they could grow market share and build a base and just have people that will bang their head against a wall trying to act like they own the market.

This! Indeed their biggest and worst problem that they need to realize and identify. Only then can they think of a solution.

 

[Hitman928]

Intel should swallow their pride and do the following configuration changes in Aurora:

Disable HT.

Clock the CPUs and GPUs down to the efficient part of the v/f curve.

Create better drivers/software to eek out more performance from their hardware.

Stop trying too hard to beat AMD. I mean, come on. Now it's just embarrassing. Just admit that you have different strengths and capitalize on that, instead of trying to hide your weaknesses.

Yes, Aurora may become pretty pedestrian in performance this way but at least it will be efficient enough to not be a total loss in terms of long term costs of power and cooling.

They have a contractual obligation to hit a certain performance level.

 

[adroc_thurston]

Disable HT.

it's an accelerated system. doesn't matter.

Clock the CPUs and GPUs down to the efficient part of the v/f curve.

already did so, this isn't desktop.

Create better drivers/software to eek out more performance from their hardware.

They certainly tried.
Gen12 is just ass.

Just admit that you have different strengths

They do not have any (well, DAOS, but that's not relevant for HPL).

 

[igor_kavinski]

They have a contractual obligation to hit a certain performance level.

Too bad. They are screwed then.

 

[DAPUNISHER]

Too bad. They are screwed then.

Someone said Intel had to take a $300 million loss a couple of years ago over the delivery issues.

 

[igor_kavinski]

Someone said Intel had to take a $300 million loss a couple of years ago over the delivery issues.

Prime Contracting No Longer One Of Intel’s HPC Aspirations

Intel has spent the past nine months reorganizing itself in the wake of Pat Gelsinger becoming its chief executive officer in January, including new

www.nextplatform.com

In its fourth quarter, the Intel Federal division, which is the prime contractor for the Aurora system, is taking a $300 million write-down.

 

[Hitman928]

Someone said Intel had to take a $300 million loss a couple of years ago over the delivery issues.

Yep. I'm curious what the consequences are if they can't get Aurora to reach the contracted performance level. Maybe they have a certain amount of payment they've held back that the DOE will just keep, or is it something else?

 

[moinmoin]

Yep. I'm curious what the consequences are if they can't get Aurora to reach the contracted performance level. Maybe they have a certain amount of payment they've held back that the DOE will just keep, or is it something else?

Is the DOE even paying anything anymore at this point? Aside electricity bills I mean.

 

[igor_kavinski]

Is the DOE even paying anything anymore at this point? Aside electricity bills I mean.

They don't need to pay if it's down coz the cooling broke down

 

[DrMrLordX]

Yep. I'm curious what the consequences are if they can't get Aurora to reach the contracted performance level.

Wouldn't they just add more nodes until reaching the desired level of performance? At no additional cost to the DoE of course.

Is the DOE even paying anything anymore at this point? Aside electricity bills I mean.

Doesn't seem like it.

 

[igor_kavinski]

Wouldn't they just add more nodes until reaching the desired level of performance?

Don't know if the contract included estimated power consumption too. If it did, then Intel can't do that or they will be in breach of contract. I'm sure though that their lawyers will find a way to weasel out of such a clause.

 

[GodisanAtheist]

Is the DOE even paying anything anymore at this point? Aside electricity bills I mean.

-You mean the American taxpayer *suddenly pedantic*

 

[moinmoin]

-You mean the American taxpayer *suddenly pedantic*

Technically not for Aurora itself, that's legacy stuff.

(German taxpayers as well btw., with a further contribution from Israeli taxpayers. *levelling up pedantry*)

 

[Hitman928]

Is the DOE even paying anything anymore at this point? Aside electricity bills I mean.

Not sure, considering the build out and how expensive PVC must be to make, it doesn’t seem like $300M would be the total cost but given how things worked out, it might have been the contracted price.

Wouldn't they just add more nodes until reaching the desired level of performance? At no additional cost to the DoE of course.

No one knows the details outside of those involved, so I’m not sure what their options might be or what penalties might be written in.

 

[igor_kavinski]

Intel probably got the penalties waived off by saying, "We are INTEL! The one and only! If we can't do it, nobody can!".

 

[DavidC1]

Since both AMD and NVidia seem to be allergic to releasing anything for under $200, Intel could actually gain some market share and good will by making smaller GPUs targeted at the lower end of the market. Even some people who'd normally stick with AMD/NV might grab a low cost Intel card just to mess with it as a spare.

Market and volume share does not matter if you don't make money. And having only value products affect market perception.

These type of comments make me believe you and others seriously misunderstand Intel as a company and business matters. I believe others saying that the top guy, Gelsinger is the dam that stopped the company from abandoning the dGPU effort months ago, however every year they don't perform, both in performance and financials is a year closer to the inevitable - dGPU being abandoned.

Neither AMD nor Nvidia are "allergic" to super cheap GPUs. They simply don't need a reason to release it. Certainly not Nvidia, where they have absolute market share without such cheap parts, and making boatloads of money just on gaming graphics.

Having more expensive products also grant you a leverage in pricing for your low end ones. This is how Nvidia was able to out-price Iris Pro notebooks back in the day. The dGPU Nvidia combo was cheaper than "integrated" Iris Pro options, because the Nvidia dGPU was a cheapie they can leverage using the financial might and Iris Pro was a top end part for Intel.

How many people are willing to "play" with what others see as having next-to-nothing value? You think after the problems with Alchemist they can stand another TWO years with Battlemage ending up lower in the stack? They need to get the 64 Xe core version out with 4070 Super/Ti performance with all the bug fixes period. Not another warmed over A770. I'm pretty sure they'll be really happy with x50 lineup at best when one of the criticisms against the GPU market is that the real name of the GPUs should be shifted one level downwards, 90-->80, 80-->70, 70-->60, 60-->50, 50-->40.

 

[DAPUNISHER]

I'd really like to know what the feedback is from OEMs shipping prebuilts with ARC in them. Did tech support and returns exceed normal levels?

 

[igor_kavinski]

You know what would be really funny? That most ARC pre-builts are being used by non-gamers like office users and grandparents

 

[DAPUNISHER]

You know what would be really funny? That most ARC pre-builts are being used by non-gamers like office users and grandparents

It is well suited to the task. The daily driver experience is on point. If the grandkids want to fire up fortnite, minecraft, Overwatch 2, or the like, it has them covered too.

 

[Mopetar]

Market and volume share does not matter if you don't make money. And having only value products affect market perception.

Having crud products affects it even more. If you can't make a great small die product, you cannot hope to make a large die product that's even good.

Intel has a certain amount of fixed cost that they're going to spend on engineers just to develop their technology. They're competing against two companies that have been in the business for decades and have established customer bases.

Intel has none of the reputation that lets NVidia do what NVidia does. Why you even mention something like that is beyond me. Making a massive die that turns into a flop is a sure fire way to get shutdown. Make a lot of smaller dies that sell and let the company spread out their fixed costs and they can justify developing the next generation of GPUs where they will have an improved position and some loyal customers.

 

[igor_kavinski]

Rumor time!

Don't watch unless you want to waste time. Based on some hardwaretimes article and summarized below:

Some 12GB card (he's confused and refers it to as B980?) with following specs:

20 XE cores for 2560 shaders
boost clock between 2.8 and 3 GHz
8 MB of L2 cache
12 GB GDDR6 @ 20 GB/s
192 bit bus
480 GB/s total bandwidth

B770 16GB
32 XE cores for 4096 shaders
boost clock between 2.8 and 3 GHz
16 or 32 MB of L2 cache
16 GB GDDR6 @ 20-24 GB/s
256 bit bus
768 GB/s total bandwidth
TBP ~200W
$299 price

B980 same as above except maybe 5120 shaders and 225W TBP at $399.

B580 may deliver the same performance as A770 at a lower power limit of 150W and priced at $199.

Baked at TSMC at an optic resolution of 4 nanomeetahs!

Launch and availability possibly sometime in November or afterwards.