News Intel GPUs - Falcon Shores cancelled

[SolidQ]

Tell me what the **** did AMD do these last 6 years HUH? All they did was follow Nvidia.

They both follow, only nvidia missing is similar to AFMF

 

[adroc_thurston]

and better arch than AMD

what crack are you smoking.
BMG has PPA as terribad as ACM had.

All they did was follow Nvidia.

They PPAmaxed. They will continue PPAmaxing.

 

[SolidQ]

Seems AMD need UDNA for compete with NV and Intel

https://twitter.com/x/status/1867290424802783558

 

[linkgoron]

Seems AMD need UDNA for compete with NV and Intel

https://twitter.com/x/status/1867290424802783558

6800xt is more than 4 years old by now.

 

[Win2012R2]

it's more likely Nvidia and even AMD are selling cards at more than 200% markup

Then they'd be lying to the stock market and execs go to jail.

It would probably be their craziest investment

Not after McAfee, not by along shot...

Xe3 is coming next year, the point is Intel has yearly GPU arch updates.

Intel Arc was launched on 30 Mar 2022, so hardly yearly.

if Intel cancels or reduces funding for their GPU department, intel client cannot survive.

They can save lots of money by shutting down their own GPU efforts and licensing it from AMD (and using their own fabs to make them) - I doubt regulators would allow NVIDIA getting it. This market is unlikely to sustain 3 players with one being so dominant in hearts and minds (and wallets!).

 

[DeathReborn]

6800xt is more than 4 years old by now.

Not to mention it has no AI boosting hardware inside.

 

[blckgrffn]

Then they'd be lying to the stock market and execs go to jail.

No, just take a second here.

What I am reading - and definitely believing is that the markup over BoM is like 200% - so a $300 “costs” $100 in hard costs, maybe or maybe not including packaging and transport to retail outlets.

THEN, depending on what’s going on in the company, overhead is allocated to the sku.

So if Intel is only breaking even on these hard costs and for every $600 card Nvidia sells they are only investing $200 you can see where the trouble steps in.

And it’s super easy to see this with how a 4070 and the B580 have the same investment into memory and even board power handling given their similar power envelopes. So the cooler is the same price. Then it comes down to how big the die is and the prices negotiated on it… then integration and boxing and transport and bam. Intel spends $200 and gets a $260 sku and Nvidia spends $200 and gets a $600 sku.

And then ongoing R&D and software and advertising and partner incentives and and and.

See what a disaster that is?

In the mean time we benefit from Intel choosing to manifest this (more dubious looking the more I think about it and the math) strategy into reality.

The funny math Intel is doing might be loading the iGPU tech with most of the r&d cost and ongoing support overhead.

 

[Hans Gruber]

You guys need to stop doing the financial math and be happy Intel fixed their GPU business. Eventually, Intel will be making GPU's in their own fab. At that point the cost'/price is like throwing a dart at a board. Intel will soon be eating AMD's lunch. I estimate Intel could take 5% of Nvidia's market share and most of AMD's in the next 2 or 3 years. It also means much better iGPU's for desktop CPU's and laptops.

I am interested to see how well the B580 overclocks. That's a new feature that was not available on Alchemist. The memory Ocing is usually golden with GDDR6.

 

[Win2012R2]

the markup over BoM is like 200% - so a $300 “costs” $100 in hard costs

What you are describing is called "gross margin", as public companies they declare that to the stock market - they don't break it down by product line, but we know that DC product get bigger margins than consumer, plus there are unavoidable fixed costs per unit, so when NVIDIA was saying they had 60% gross margin that means consumer stuff would defo have less - with possible exception of 4090. Other big overheads like R&D, marketing, leather jackets all taken into account later - that's not a gross margin.

There is no way a brand new GPU with a die size of near 300 mm on a fairly modern TSMC N4 process with 12 GB of VRAM can have "hard costs" of $100 - absolutely no chance.

The funny math Intel is doing might be loading the iGPU tech with most of the r&d cost and ongoing support overhead.

Here is the funny thing - those iGPUs don't make Intel money, they make them by selling CPU - and if design of those iGPUs is as bad (area wise) as it is for desktop then it's costing Intel money, plus R&D and inability to offset it with some higher end models. For a bean counter it is very attractive target to cut totally by licensing designs.

You guys need to stop doing the financial math and be happy Intel fixed their GPU business

It's not fixed if finance does not work, and if it isn't working then chances are a broken clock MLID will be proven correct with Intel canning discrete GPUs for good, this calculation is important when deciding whether to gamble and save 50 bucks or just hold out a bit to get The Way It's Meant To Be Played thingy with 12 GB VRAM.

 

[ajsdkflsdjfio]

B580 should be 10%-ish cheaper to manufacture than the 4070 Super, which has an MSRP of $600. So either this is sold at a loss, or Jensen is robbing gamers blind.

Why compare it to the 4070 Super which is likely overpriced for its die size, what about the 7700xt? The 7700xt currently retails at 400$ with a chiplet architecture and a main die size of 200 mm^2 alone, with a total die size of 350mm^2. Although the extra 150mm is on TSMC n6, I'd say the packaging costs for this MCM design make up for the slightly cheaper n6 node for the MCD dies. So unless the 7700xt is selling at barely breakeven pricing, I'd say it's quite possible the b580 isn't losing intel money.

Besides, this is what it takes to play catch up to AMD and Nvidia in the graphics segment. As long as battlemage isn't actively losing money (minus R&D), it helps to keep the graphics division alive to continue to iterate and improve to maybe be more competitive next time around and also more profitable. The battlemage launch is already 10x better than the alchemist launch in all ways, performance, die size (the a770 was 400mm^2 of n6 and get's trounced on by the 272 mm^2 b580), power consumption, and drivers. Even if they aren't 100% competitive this time around they have made huge leaps of progress to the point where depending on Celestial or Druid, we might actually be deciding between three graphics card makers instead of two in the future.

 

[gdansk]

better arch than AMD

It's competing with Navi 22... It is not better.
We can give them credit for pricing it appropriately and using a few billion extra transistors on both useful and seldom-used features.

 

[adroc_thurston]

Seems AMD need UDNA for compete with NV and Intel

que? they have more GPGPU mss that Intel would ever get.

 

[Kepler_L2]

Why compare it to the 4070 Super which is likely overpriced for its die size, what about the 7700xt? The 7700xt currently retails at 400$ with a chiplet architecture and a main die size of 200 mm^2 alone, with a total die size of 350mm^2. Although the extra 150mm is on TSMC n6, I'd say the packaging costs for this MCM design make up for the slightly cheaper n6 node for the MCD dies. So unless the 7700xt is selling at barely breakeven pricing, I'd say it's quite possible the b580 isn't losing intel money.

Besides, this is what it takes to play catch up to AMD and Nvidia in the graphics segment. As long as battlemage isn't actively losing money (minus R&D), it helps to keep the graphics division alive to continue to iterate and improve to maybe be more competitive next time around and also more profitable. The battlemage launch is already 10x better than the alchemist launch in all ways, performance, die size (the a770 was 400mm^2 of n6 and get's trounced on by the 272 mm^2 b580), power consumption, and drivers. Even if they aren't 100% competitive this time around they have made huge leaps of progress to the point where depending on Celestial or Druid, we might actually be deciding between three graphics card makers instead of two in the future.

RDNA3 isn't exactly a profit making machine...

 

[511]

Apparently in long term it is not a huge of amount
> 7.87 Billion
> $3Billion in Defence
> Tax credit on $100 Billion Investment (this is huge )

 

[511]

In terms of Fabs this is nothing. TSMC regularly spend 30-35 billion annually on R and D and Fabs with a better, cheaper and more efficient workforce.

Taiwan work culture is different from American so there is cost difference due to that TSMC america will charge you more for the same process

Over a period of 4 years, I would say Intel needs 100 billion to keep up with TSMC on top of their own funding as a result of waste.

Intels numbers of 10% faster than a RTX 4060 are definitely on the optimistic side. $250 pricing is fair based on risk and performance on this chip. WIll be a loss on this product at this price. Too bad because aside from the financials seems like a decent product.

I think they are selling with 0 Margin for the die but the R&D is shared between Lunar lake dGPU

 

[511]

It's competing with Navi 22... It is not better.
We can give them credit for pricing it appropriately and using a few billion extra transistors on both useful and seldom-used features.

The Tx count is matched with 4060 it's the Libraries they are using ( classic Intel using Relaxed pitch libraries for design)

 

[ajsdkflsdjfio]

RDNA3 isn't exactly a profit making machine...

Yea that's Nvidia, regardless I doubt RDNA3 is selling at 0% margin.. Like I said... "unless the 7700xt is selling at barely breakeven pricing, I'd say it's quite possible the b580 isn't losing intel money."

B580 should be 10%-ish cheaper to manufacture than the 4070 Super, which has an MSRP of $600. So either this is sold at a loss, or Jensen is robbing gamers blind.

Your comparison with 4070 was extremely flawed and implied that intel was somehow selling a 500 dollar + gpu at 250$, Nvidia isn't the only graphics card brand to draw data from...

 

[poke01]

by arch I meant RTRT and XMX/AI. I know the PPA is rubbish. If Intel doesn’t fix that by celestial then that’s a problem, means dGPU won’t be the money maker for a very long time.

 

[GTracing]

Yea that's Nvidia, regardless I doubt RDNA3 is selling at 0% margin.. Like I said... "unless the 7700xt is selling at barely breakeven pricing, I'd say it's quite possible the b580 isn't losing intel money."

Your comparison with 4070 was extremely flawed and implied that intel was somehow selling a 500 dollar + gpu at 250$

According to AMD's Q3 report, gaming is currently barely turning a profit.

Not that I think Battlemage is selling at a loss, but I agree with others that the margins can't be very big.

 

[gdansk]

by arch I meant RTRT and XMX/AI

That's like saying Bulldozer had a better architecture than Gulftown because it has AVX and FMA. Nice in their niche but not key to the architecture. The phrase you're looking for is "feature-rich". But even then I'm not sure it's accurate.
Their general performance is still only on par in most respects to a 4 year old Navi even when blessed with more memory bandwidth, N5 vs N7, and higher clocks.

 

[ajsdkflsdjfio]

According to AMD's Q3 report, gaming is currently barely turning a profit.

Not that I think Battlemage is selling at a loss, but I agree with others that the margins can't be very big.

I agree, it's probably close to break even or even a slight loss but it's not some crazy high loss. It's what it needs to be, another stopgap product release for intel to gather data and improve drivers etc for their future graphics generations. If it gains market share through its competitive pricing that's just icing on the cake.

 

[Tigerick]

For those really wants to know, Intel last reported financial of AXG group was in Q4 2022. With revenue of $247M, Intel occurred operating loss of $441M, yep almost double loss of revenue. The division is money loser, that's why Intel merged AXG into DCAI group....

 

[GTracing]

For those really wants to know, Intel last reported financial of AXG group was in Q4 2022. With revenue of $247M, Intel occurred operating loss of $441M, yep almost double loss of revenue. The division is money loser, that's why Intel merged AXG into DCAI group....

View attachment 113136

That business unit isn't just consumer graphics cards. It included HBM server CPUs and server GPUs. The bullet points are clearly not talking about Alchemist since Intel had no dedicated GPUs in Q4 2021. From the Q4 2022 financial report:

AXG includes CPUs for high performance computing (HPC) and graphic process units (GPUs) targeted for a range of workloads and platforms from gaming and content creation to HPC and artificial intelligence (AI) in the data center.

I should clarify that when I say I don't think Battlemage is selling at a loss, I mean that the cost to manufacture and distribute the graphics card is less than MSRP. I'm not including R&D or other fixed costs. Those fixed costs are huge so I'm sure Intel's GPU division is losing money, but imho they lose less money when they sell more graphics cards.

 

[Win2012R2]

It included HBM server CPUs and server GPUs.

None of these sold well, HBM Xeons now discontinued and server GPUs are dead it seems - Aurora supercomputer used both and that's probably it.

I mean that the cost to manufacture and distribute the graphics card is less than MSRP.

You might be right, but very marginally (pun intended)

1) 4nm wafer is 18k, will be 20k next year (source: https://www.tomshardware.com/tech-industry/tsmc-may-increase-wafer-pricing-by-10-for-2025-report )
For a chip near 300 sq mm (16 x 18 mm) with perfect 100% yield we get around 200 chips per 12'' wafer (source: https://anysilicon.com/die-per-wafer-formula-free-calculators/ )
So we are at $100 per chip, minimum - at perfect yield, however if we use defect density of 0.07 as supposedly on N7 ( source

https://twitter.com/x/status/1536710027943284737

) then factoring this in we get 155 chips (source: https://isine.com/resources/die-yield-calculator/ ) or more realistic price of $130 (using next years wafer pricing).
2) 12GB of GDDR6 using average session spot price is $27.60. (source: https://www.dramexchange.com/ )
3) PCB - don't know well that, should be cheap but must be at least $10?
4) Packaging all the above together - can't be free, at least $5?
5) Packaging (as in box etc) - not cheap these days, at least $5?

So far - 130+27.60+10+5+5= $ 177.60

Add to this which isn't not BOM per se, but needs to be built-into sell price -

6) Retailer margin - 10% (one would think it's 5% but card transactions these days take easily 2-3% off) - $25
7) Transport - ???

That's already over $200 and I think the above is very conservative calculation - the chip is so strangely big that it's real yields (binning) might be worse than that, no margin for error!

In all this the only saving grace is massive drop in GDDR6 pricing - zero excuse at current levels to give anything less than 12-16 GB, but who knows at that levels Intel pre-booked guaranteed supply, it could be 2022-s levels, but this is total speculation.

All good so far... but it isn't - 5060/8800 stuff lands and this product will have to drop at least 50 bucks more.

 

[igor_kavinski]

AIB cards have excellent cooling.

Sparkle Arc B580 Titan OC Review

The Sparkle Arc B580 Titan OC is a custom-design version of Intel's newest card. It comes with a triple-fan cooling solution that still fits into two slots. Data in our review confirms that the cooler is considerably improved over Intel's reference card, reaching much lower temperatures.

www.techpowerup.com