News Intel GPUs - Falcon Shores cancelled

[ajsdkflsdjfio]

2% if you're lucky lmao.

pessimism

18% transistor desnity is not optimism.

N3 and N3B are the same process and it is going away. TSMC will continue to run it for Intel/Apple in smaller numbers as they need to support already existing products but no one's going to use it for new products. N7 is pretty much gone as well, TSMC has pushed all new designs to use N6 because it is cheaper due to it being a EUV process that needs fewer masks and passes and there's really no reason to continue to run N7 when N6 is a win-win for customers and foundry.

Similar things can be said about the N4 family compared to the N5 family(refer to ss below). Regardless both exist and are options for customers. N3 vs N3B is different since N3B in general was too ambitious and resulted in too high of costs even with better performance, thus being only suited for applications like apple products and small Intel compute chips. N6 might be a win-win for customers and foundry based on cost per transistor but at the same time costing more per wafer.

My point is that you can't use N7 wafer as N6 wafer costs, while using N4 wafer costs for N5 wafer costs like win2012 did.

 

[Win2012R2]

still comparable to N4 vs N5

If it isn't then you should consider deleting your account...

 

[ajsdkflsdjfio]

If it isn't then you should consider deleting your account...

So you are deleting your account?

 

[Win2012R2]

So you are deleting your account?

Nope

 

[Hans Gruber]

Nope

But you just created your account. You can always start a new account.

 

[ajsdkflsdjfio]

But you just created your account. You can always start a new account.

Maybe that's why his acc is so new, he had to delete his other one for saying stupid shit as well

 

[Win2012R2]

But you just created your account. You can always start a new account.

I've already got a "wager" running on it and it should take 18-24 months to settle it, would be very unethical to even risk it in an unrelated wager. However if the opponent does it then it's on them...

 

[Win2012R2]

Maybe that's why his acc is so new

ajsdkflsdjfio
Member
Nov 20, 2024

 

[adroc_thurston]

18% transistor desnity is not optimism.

Nominal metrics are not real world™ designs. Get real.

 

[ajsdkflsdjfio]

Nominal metrics are not real world™ designs. Get real.

I was comparing nominal metrics to nominal metrics it is absolutely useful in that scenario.

 

[ajsdkflsdjfio]

To quote another one of your posts:

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

Even accounting for density increase going from N6 to N5, if you compare just the transistor counts the b580 performs +25% vs a770 with -10% transistors or about a 38% increase of performance transistor for transistor. Who exactly is smoking crack? You?

And by the way if you are including real world density in your comparison of BMG to ACM, a770 had 21.7 transistors in 400 mm of die area, b580 with 19.6 in 270mm. Even with tsmc's nominal metrics, N6 to N5 should only offer up to a 20% density increase, while ACM-BMG offered a 33% density increase. So not only does the b580 offer +38% performance transistor for transistor (partly due to node but still a significant part due to design), but it also is 10-15% denser by design regardless of node. Combine the two you get 40-50% better ppa. Get real.

 

[gaav87]

"Negotiations with AI and HPC customers, such as Nvidia, suggest these clients can tolerate approximately 10% price hikes for 4nm-class wafers from around $18,000 per wafer to around $20,000 per wafer. As a result, the 4nm and 5nm nodes, primarily used by companies like AMD and Nvidia, are expected to see an 11% blended average selling price (ASP) hike." Source:

TSMC may increase wafer pricing by 10% for 2025: Report

Chips are getting more expensive.

www.tomshardware.com

If N5 is indeed 16k, then price goes up by 11% in the new year and becomes $18k - happy now? Does it change things materially? No.



False for obvious reasons - it simply can't be true: rx 7600 is 204 sq mm die (techpowerup data) and it's on N6 which is priced at $9500 as Google search above shows, Intel uses 50% more silicon for DOUBLE the price, yields will also be worse - they are 3-3.5 times more expensive to make than rx 7600, that can not be +40 bucks unless AMD pays 15 bucks for their dies.



Well assumption does not contract that they lose money on each sale, does it? Price is the only thing going for it, who'd recommend it if it was priced at $400 like it should have been for 300 sq mm N5-class silicon. PS5 Pro is on N4, I would find it very surprising if this chip isn't using it too. N4 to N5 is what N6 to N7.

Yelds are better g21 is less dense...

 

[Hitman928]

pessimism
View attachment 113244
18% transistor desnity is not optimism.


Similar things can be said about the N4 family compared to the N5 family(refer to ss below). Regardless both exist and are options for customers. N3 vs N3B is different since N3B in general was too ambitious and resulted in too high of costs even with better performance, thus being only suited for applications like apple products and small Intel compute chips. N6 might be a win-win for customers and foundry based on cost per transistor but at the same time costing more per wafer.

My point is that you can't use N7 wafer as N6 wafer costs, while using N4 wafer costs for N5 wafer costs like win2012 did.
View attachment 113243

N3 is N3B, they are the same node, do you mean N3E?

By win-win I meant N6 costs less for the customers and TSMC and TSMC could push more wafers through faster with N6, so everyone wins. The foundry saves a bunch of time and money when they can use EUV rather than lots of masks with multiple exposures.

The N5 and N3 families didn’t have this significant of a difference between the original process and its variants.

As far as the pricing goes, the numbers are pretty far off but I don’t really care about what you two are arguing about, just pointing out the corrections above.

 

[DrMrLordX]

Like I said, they could have priced it at 300$ if they wished to making up for the +40 dollars die +10 ram. Regardless of the budget of the BOM, Pat is no longer at Intel and they could have chosen to price it 300, they don't need a year in advance to change the price. They priced it at 250$ because their margins can likely take it and they aren't focused on maximizing profit right now anyways. If they were losing money on each battlemage card at 250$ I guarantee they would price it higher because it would still sell even at 300 dollars.

Hope you're right. Either that or they're underpricing it to move limited volume out of the way and wash their hands of it. Let's see if they keep supplying the channel.

What I don't get is why not have these on "Intel 3".

Sierra Forest, Granite Rapids, and Arrow Lake-U may be taking up all the available Intel 3 wafers. Plus Intel is still in a period of limited wafer volume for anything more sophisticated than Intel 7.

 

[ajsdkflsdjfio]

N3 is N3B, they are the same node, do you mean N3E?

By win-win I meant N6 costs less for the customers and TSMC and TSMC could push more wafers through faster with N6, so everyone wins. The foundry saves a bunch of time and money when they can use EUV rather than lots of masks with multiple exposures.

The N5 and N3 families didn’t have this significant of a difference between the original process and its variants.

As far as the pricing goes, the numbers are pretty far off but I don’t really care about what you two are arguing about, just pointing out the corrections above.

I mean n3 in general vs n3b, n3b wasn't used that much because it was too expensive/too ambitious and TSMC ended up nerfing it and creating a bunch of derivatives at least from what I understand.

You are right about the large differences from N6 vs N7 but also that doesn't necessarily mean that the prices for N6 and N7 are same. IMO its likely that N6 is still more expensive than N7 because in it's also a lot denser and offers higher performance than N7, even with the EUV implementation it still could be more expensive overall to produce and TSMC has an incentive to price it higher anyways. But in general the N6 offering could offer much lower price/transistor than N7 while also being more expensive to purchase, they aren't exclusive.

What numbers do you think are more accurate for N5 vs N6? I'm just curious. I just thought that win2012 comparing N7 vs N4 prices for rdna3 vs battlemage was stupid since it's actually N6 vs N5 and the prices he used were pretty random too. He quotes something like 20k vs N4 and 9500 for N7 meaning that he thought BMG dies were using a 2.1x more expensive node which to me doesn't pass the sniff test.

 

[Joe NYC]

Must be cost (of opportunity as well as ~300 sq mm of Xeon is worth a lot more than 250 bucks retail), plus that node probably would not live long enough anyway.

First, Intel does not have a shortage of server parts, so there is no opportunity cost.

And second, "Intel 3" based Xeon parts are not selling in high volume. These are Sierra Forrest and Granite Rapids parts, launched earlier this year, and they don't have a lot of traction. The main volume Intel server parts are now Sapphire and Emerald Rapids - used in NVidia DGX boxes.

So this would have been a perfect opportunity for Intel to avoid at the same time, its own fab underutilization charges, and also paying to TSMC, for a node that is in high demand (N5)

 

[Thunder 57]

I'm way late to the party but it looks like Battlemage did well. I did screen the reviews launch day but now that I have an actual computer again I'd say it has a chance. Especially if Nvidia and AMD screw up and give us 8GB on the low end parts again.

I've read rumors of the 5060 being 8GB and 12GB. Is it possible the vanilla 5060 has 8GB and the Ti 12GB? With AMD, who knows. Someone suggested if it is a 128 bit card to leave 8GB to OEM's and 16GB only in retail. I don't think they go that way. Maybe more like a 6700/6700 XT approach with 10GB/12GB?

 

[Kepler_L2]

I'm way late to the party but it looks like Battlemage did well. I did screen the reviews launch day but now that I have an actual computer again I'd say it has a chance. Especially if Nvidia and AMD screw up and give us 8GB on the low end parts again.

I've read rumors of the 5060 being 8GB and 12GB. Is it possible the vanilla 5060 has 8GB and the Ti 12GB? With AMD, who knows. Someone suggested if it is a 128 bit card to leave 8GB to OEM's and 16GB only in retail. I don't think they go that way. Maybe more like a 6700/6700 XT approach with 10GB/12GB?

5060Ti is using cutdown GB205 with 192-bit bus and 5060 uses GB206 with 128-bit bus. 12GB on 5060 is possible if they delay it until 3GB GDDR7 is ready but seems unlikely imo.

 

[Thunder 57]

5060Ti is using cutdown GB205 with 192-bit bus and 5060 uses GB206 with 128-bit bus. 12GB on 5060 is possible if they delay it until 3GB GDDR7 is ready but seems unlikely imo.

I think they may do a "Super" refresh with 3GB GDDR7 a year after or so.

 

[DrMrLordX]

So this would have been a perfect opportunity for Intel to avoid at the same time, its own fab underutilization charges, and also paying to TSMC, for a node that is in high demand (N5)

Do we know for sure that they're idling some of their fabs tooled for Intel 3? That seems ironic, considering how much trouble they had getting enough EUV machines to push Intel 4/3 volume.

 

[Win2012R2]

And second, "Intel 3" based Xeon parts are not selling in high volume

Just recently launched Xeon 6 is on Intel 3 - it is indeed in low volume, I can't find it on sale in the UK (have to say not desperate to buy it though), but most likely Intel 3 is a low volume in the first place - why keep it big if it's a node that will last a year and the whole farm is bet on 18A. Either way it seems not prudent to put 250 MSRP device via node that competes with Xeons, plus when they designed B-mage (few years ago) they could not have counted on Intel 3 being ready on time anyway.

paying to TSMC, for a node that is in high demand (N5)

It was my long standing theory that this was a long standing (since at least 2021 when Pat negotiated deals) plan by Intel to limit TSMC capacity otherwise available to AMD - from this point of view selling one B-mage at loss prevents AMD from selling another 32 core EPYC that would eat into Xeons and cause even bigger losses, perhaps at real catastrophic level.

 

[igor_kavinski]

It was my long standing theory that this was a long standing (since at least 2021 when Pat negotiated deals) plan by Intel to limit TSMC capacity otherwise available to AMD - from this point of view selling one B-mage at loss prevents AMD from selling another 32 core EPYC that would eat into Xeons and cause even bigger losses, perhaps at real catastrophic level.

Knowing Intel, I would not be surprised if they had actual meetings to discuss this strategy. Oh well. At least some people get the GPU they deserve instead of 8GB Geforce crap.

 

[Win2012R2]

Knowing Intel, I would not be surprised if they had actual meetings to discuss this strategy.

It's a sensible strategy (when devised at time of big money they still had in 2021), I would have done it too, heck anybody in their position should have - I don't even think it's illegal - in any case they have plausible deniability in a form of B-mage, only with this such release makes any commercial sense, but the wheels I fear are coming off the cart now.

 

[Hitman928]

I mean n3 in general vs n3b, n3b wasn't used that much because it was too expensive/too ambitious and TSMC ended up nerfing it and creating a bunch of derivatives at least from what I understand.

You are right about the large differences from N6 vs N7 but also that doesn't necessarily mean that the prices for N6 and N7 are same. IMO its likely that N6 is still more expensive than N7 because in it's also a lot denser and offers higher performance than N7, even with the EUV implementation it still could be more expensive overall to produce and TSMC has an incentive to price it higher anyways. But in general the N6 offering could offer much lower price/transistor than N7 while also being more expensive to purchase, they aren't exclusive.

What numbers do you think are more accurate for N5 vs N6? I'm just curious. I just thought that win2012 comparing N7 vs N4 prices for rdna3 vs battlemage was stupid since it's actually N6 vs N5 and the prices he used were pretty random too. He quotes something like 20k vs N4 and 9500 for N7 meaning that he thought BMG dies were using a 2.1x more expensive node which to me doesn't pass the sniff test.

The original N3 became N3B and N3E quickly replaced it as the “base” N3 family process. The rest of the N3 iterations stem from there.

Pricing wise, the gap is significantly smaller than the numbers mentioned.

 

[SolidQ]

A770 vs B580