News Intel GPUs - Falcon Shores cancelled

[Win2012R2]

It's next to impossible to find sources on real wafer prices. The people who know don't go "on the record" blabbing about it.

Providing a source of whoever was blabbing about it would be better than nothing.

We can get a generally good idea how much on average wafers cost - TSMC is a public company and give reasonably good breakdown of revenues by process type, plus we know max wafer rates per fab and utilization rate also published by TSMC, it gets easier with cutting edge since they usually get 100% used.

What we don't know is discounts given to some volume customers, but given TSMC recent attitudes I don't think they get much of a discount - demand is super high and where else would they go? Capacity is very limited too, Apple prepays a lot, so they get preferential access to initial volumes.

$12k for N5 class seems to be rather hard to believe, it's too low given where N7/6 is now, N3 is higher, and N2 will push even more, transistors don't get cheaper.

 

[coercitiv]

That's for N5 class wafer? Source please.

Have you ever provided a source for N5 pricing other than linking to sources talking about other nodes?

All of this started from your claim that Intel is losing money with every BM card because the chips are too big. The burden of proof was on you, yet most folks around here were courteous about it and engaged in honest debate, since we all know wafer prices are very hard to pinpoint.

 

[GTracing]

Providing a source of whoever was blabbing about it would be better than nothing.

Huh? I said people don't go blabbing about it. Sources about actual wafer prices are scarce.

We can get a generally good idea how much on average wafers cost - TSMC is a public company and give reasonably good breakdown of revenues by process type, plus we know max wafer rates per fab and utilization rate also published by TSMC, it gets easier with cutting edge since they usually get 100% used.

What we don't know is discounts given to some volume customers, but given TSMC recent attitudes I don't think they get much of a discount - demand is super high and where else would they go? Capacity is very limited too, Apple prepays a lot, so they get preferential access to initial volumes.

From what I've heard (on this forum and elsewhere on the interwebs), virtually all major customers get discounts. The price numbers in the Tom's Article are like the MSRP at Kohl's, nobody actually pays that much.

$12k for N5 class seems to be rather hard to believe, it's too low given where N7/6 is now, N3 is higher, and N2 will push even more, transistors don't get cheaper.

Do you find it easier to believe that N5 costs as much as N3B?

 

[Win2012R2]

Have you ever provided a source for N5 pricing other than linking to sources talking about other nodes?

I linked to what I found via Google, from a reasonably trusted publishing source (TomsHardware), what do you want me to give you - NDAed copy of wafer purchase agreement? At least I provided a source to my information, which does not contract general knowledge about the pricing, yet here we have claim of $12k per N5, which has no source whatsoever and contradicts what we know about how much TSMC generally charges and current pricing trends (upwards).

The burden of proof was on you

I've provided my calculations using two models - one what we know about chip/wafer pricing and another comparative with PS5 Pro which happens to have about the same sized chip built on about the same tech (N5-class), both calculations suggest the same conclusion, which is approximately infinitely times more credible in my view than unjustified in reality claims that Intel can break even on selling for $250 something that others sell for lot more.

Do you find it easier to believe that N5 costs as much as N3B?

N3B is a dead bad node, it's not for sale now - only (known to me) volume users were Apple, Intel.

N3E certainly costs more than N5-class, so much more that even NVIDIA won't use it for 5000 series, and neither cheapskate AMD for 8000 series. That does not mean that N5 suddenly costs $12k - absolutely no chance, it is very likely to never ever cost that little.

virtually all major customers get discounts

You do understand that only serious major customers have the capacity to design and pay for pretty modern nodes like N5 and N3? These nodes are near 100% utilised, there is serious competition for wafers - TSMC model is effectively auctioning wafers to clients on a regular basis, so if there is high demand they pay more, demand is red hot right now.

 

[gdansk]

from a reasonably trusted publishing source (TomsHardware)

They may be trusted in reviews (that depends on the author and is a subject for another thread) but their news is repeating what they heard elsewhere and has no merit.

 

[Win2012R2]

They may be trusted in reviews (that depends on the author and is a subject for another thread) but their news is repeating what they heard elsewhere and has no merit

Ok, and who has got more merit - guy on this forum who claims that N5 class wafer is $12k?

Here is more: "TSMC minced no words when talking about per wafer costs. SemiAnalysis often shocked people years ago when we discussed N7 wafers being roughly $10,000 each. N5 is even more eye popping at around $16,500, meaning the cost per transistor is up significantly. We are now hearing that N3 is above $20,000"

This is from Oct 2021- source: https://semianalysis.com/2021/10/14/tsmc-3nm-wafer-shipments-pushed-into/

Is that serious enough for you? N5-fabs equipment has not depreciated yet, will take at least a few more years, so prices won't come down anytime soon - it can not be $12k with current level of demand.

 

[DAPUNISHER]

UK Steve put the Sparkle through its paces in 15 games. He says it's not, but it's CPU limited. He runs a 14900K at 125W and wonders why the Sparkle drops from max boost clock in town in RDR2. Anytime the card is 95% or less it's waiting on the CPU in that game. He should use GPU busy, as I am confident it is the issue. ARC can also have unusually high CPU overhead at times, so comparing it to the 4060 means nothing.

I suspect his raptor degraded as he used to run it on a 380mm or 420mm AIO stock settings. He never shows stats for the 14900K in AB anymore either. That speaks volumes, as he used to always have it OSD.

 

[ajsdkflsdjfio]

I linked to what I found via Google, from a reasonably trusted publishing source (TomsHardware), what do you want me to give you - NDAed copy of wafer purchase agreement? At least I provided a source to my information, which does not contract general knowledge about the pricing, yet here we have claim of $12k per N5, which has no source whatsoever and contradicts what we know about how much TSMC generally charges and current pricing trends (upwards).

The Tomshardware source you gave listed a theoretical price for N4 wafers of being around 18k. You then extrapolated from a speculative sentence from the article to claim that N4 wafer pricing was actually 20k. Then you applied that 20k price to N5 which is a completely separate node, when all other available sources list 15-16k as the price for N5.

Is that serious enough for you? N5-fabs equipment has not depreciated yet, will take at least a few more years, so prices won't come down anytime soon - it can not be $12k with current level of demand.

Why can't it be 12k? If intel were able to have a 40% discount on N3 (higher demand less depreciated equipment) why wouldn't similar discounts be possible on other older nodes? Either way, wafer pricing is clearly more complicated than what is publicly listed on various hardware news sites. I suspect neither tomshardware nor semianalysis knows the actual up-to-date pricing for wafers. But even by the sparse public information, no stretch of imagination should be able to arrive at your 20k vs 9.5k figure.

 

[GTracing]

N3B is a dead bad node, it's not for sale now - only (known to me) volume users were Apple, Intel.

That doesn't make any sense. N3B is still better than N5. Why would TSMC sell it for less than N5?

 

[DAPUNISHER]

This thread is starting to suck. It reads like obfuscation at this point. I am going to start a Battle Mage review thread since some seem to determined to discuss anything but the performance of the GPUs.

 

[Win2012R2]

That doesn't make any sense. N3B is still better than N5. Why would TSMC sell it for less than N5?

TSMC does not want to sell N3B - it's a bad node that cost them too much and gave too little benefit to the end customers, it's not for sale now, and prices you've seen were for initial customers when N5 was quoted to be $16500 - since then (what a shocker) prices gone up across the board! That's why there is nothing surprising in N5 class to cost 18k this year and 20k next - annual 7-10% increase will do it easily, which is what TSMC been doing recently.

And if you had read up about the first (and only) iPhone with N3B then you'd know it wasn't "better", that's why very few takers.

Then you applied that 20k price to N5 which is a completely separate node

No, N5 is not completely separate node, just like N6 isn't "separate" from N7 - it was explained earlier in the thread and instead of Googling to double check you insist on your version of events that have no bearing with reality. Unless you have a better external source that N5 costs any other amount then what Intel is paying this year can be reasonably assumed to be $18k and going up to $20k next year as per Toms article, which is an infinity times better than any sourcing that you've got (ie - none).

 

[ajsdkflsdjfio]

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

B580 +25% perf at 1080p with 272 mm2 vs 406 mm2 A770. 86% increase
BMG +25% perf at 1080p with 19.6 vs 21.7 transistors. 38% increase transistor for transistor
BMG is also designed denser regardless of the node. ACM-BMG offered a 33% transistor density while the N6-N5 improvements based on TSMC's own numbers only has a 20% transistor density increase. This means that BMG performance 38% better transistor for transistor while also being designed 10%+ denser than ACM regardless of node.

 

[ajsdkflsdjfio]

No, N5 is not completely separate node, just like N6 isn't "separate" from N7 - it was explained earlier in the thread and instead of Googling to double check you insist on your version of events that have no bearing with reality. Unless you have a better external source that N5 costs any other amount then what Intel is paying this year can be reasonably assumed to be $18k and going up to $20k next year as per Toms article, which is an infinity times better than any sourcing that you've got (ie - none).

N6 is absolutely a different node from N7, it was literally explained how the manufacturing process and design specs are different from node to node. Just because N7 is no longer widely used in favor of N6 doesn't mean you can assume they use the same price. It's like assuming the 5090 will be the same price as the 4090 just because it replaced the 4090.

Keep coping though, N6 = N7 pricing, N5 = N4 = 20k, and intel is losing boatloads of money on each battlemage card. This obviously explains why they are pricing the card so competitively because everyone knows when your products are losing you money the right strategy is to price your products even lower despite being competitive at higher prices.

 

[GTracing]

TSMC does not want to sell N3B - it's a bad node that cost them too much and gave too little benefit to the end customers, it's not for sale now, and prices you've seen were for initial customers when N5 was quoted to be $16500 - since then (what a shocker) prices gone up across the board! That's why there is nothing surprising in N5 class to cost 18k this year and 20k next - annual 7-10% increase will do it easily, which is what TSMC been doing recently.

You have no idea what you're talking about about. Lower utilization would make a node's financial outlook worse, not better.

 

[Win2012R2]

Lower utilization would make a node's financial outlook worse, not better

Yes, it's a game of numbers, utilisation is the key, so what about it? TSMC got near 100% current use for N5-class nodes, which is why prices going up, not down.

 

[Kepler_L2]

N6 is absolutely a different node from N7, it was literally explained how the manufacturing process and design specs are different from node to node. Just because N7 is no longer widely used in favor of N6 doesn't mean you can assume they use the same price. It's like assuming the 5090 will be the same price as the 4090 just because it replaced the 4090.

Keep coping though, N6 = N7 pricing, N5 = N4 = 20k, and intel is losing boatloads of money on each battlemage card. This obviously explains why they are pricing the card so competitively because everyone knows when your products are losing you money the right strategy is to price your products even lower despite being competitive at higher prices.

N6 was literally created to be a cheaper version of N7 for long term use.

 

[ajsdkflsdjfio]

N6 was literally created to be a cheaper version of N7 for long term use.

Cheaper in what metric? Cost per transistor I can definitely see, cost per wafer is another question. You can have an 18 percent denser node (double digit performance/power) that costs only 10% more per wafer, nonetheless the wafer price is increased. Also it's not just easier manufacturing process steps that are included with N6, it's also density performance and power improvements as well which are made through other engineering changes. Even if the resulting N6 wafer is cheaper for TSMC to produce(dubious), I don't see why TSMC would be obligated to price it lower. With high utilization of their EUV machines in the AI boom I also don't see how the cost savings through EUV are as applicable versus regular DUV. Either way it's impossible to definitively say N7=N6 pricing and it's publicly known that N4 is more expensive than N5 so again his original assumption of 20k vs 9.5k was horrible even before factoring in things like volume discounts.

A 20k N5 wafer would imply that the b580 die costs 120$ alone which is economically impossible for the product's pricing if you use elementary reasoning. If this were the case, they wouldn't be so aggressive on pricing and/or wouldn't have released the product with as much volume and planned continued production as they have shown.

 

[Fanatical Meat]

Again why does this matter? Intel is quite obviously aware of cost & margins. They priced it as is for a reason.

 

[ajsdkflsdjfio]

Again why does this matter? Intel is quite obviously aware of cost & margins. They priced it as is for a reason.

Some people here refuse to use basic reasoning. It's obvious they aren't making much off of battlemage but at the same time it's impossible they are losing money. If they were losing let's say 10-20 dollars per card they'd just price it 10-20 dollars more since it'd still be competitive at that pricing. If they were losing 50+ then there simply wouldn't be this large of a launch. Despite rumors of low-ish supply for Battlemage, many people actually were able to buy the card and personally attest to it on both anandtech and twitter, unlike the alchemist launch which was a true paper launch. This simply wouldn't be the case if each card produced were losing intel tons of money.

 

[Fanatical Meat]

Some people here refuse to use basic reasoning. It's obvious they aren't making much off of battlemage but at the same time it's impossible they are losing money. If they were losing let's say 10-20 dollars per card they'd just price it 10-20 dollars more since it'd still be competitive at that pricing. If they were losing 50+ then there simply wouldn't be this large of a launch. Despite rumors of low-ish supply for Battlemage, many people actually were able to buy the card and personally attest to it on both anandtech and twitter, unlike the alchemist launch which was a true paper launch. This simply wouldn't be the case if each card produced were losing intel tons of money.

Agreed, intel has a strategy and that’s that.
From the beginning of this and other threads I found it really odd that some people appear to want intel to fail.
I guess they like the idea of nvidia dominating the market for whatever reason.

 

[Win2012R2]

Again why does this matter? Intel is quite obviously aware of cost & margins. They priced it as is for a reason

It matters for the future of Intel consumer discrete GPUs and their buyers - Pat was fired two weeks ago and the company is under the new bean-counting MBA-toting management, so if Arc loses too much money they might shut it all down next year, other than that I guess it does not really matter.

at the same time it's impossible they are losing money

Intel is absolutely losing money on Battlemage, the only question whether it's a negative gross margin or it's slightly positive, other than that it's a huge money loss for certain, I guess it's all academic now because its future would not change even if they are making +20% gross, that's not enough - Intel needs 50%+ margins, stock market demands it, any product that dilutes it is already bad for the whole company, that's how bean counters think. Only way to counter this is to make low volume and creatively account for R&D as shared among iGPUs too, but that only goes so far.

 

[Fanatical Meat]

It matters for the future of Intel consumer discrete GPUs and their buyers - Pat was fired two weeks ago and the company is under the new bean-counting MBA-toting management, so if Arc loses too much money they might shut it all down next year, other than that I guess it does not really matter.

Then they’d price it higher or they will price it higher.
We don’t need to hyper evaluate intels financial strategy, that is intels job.

 

[Win2012R2]

Then they’d price it higher or they will price it higher.

They clearly priced it to sell (which in itself invites some thoughts why).

Do you think reviews would have been as positive if the price was $300 or $350-375? To have sustainable discrete product they need 45-50% gross margins, maybe even higher to justify the aggro.

We don’t need to hyper evaluate intels financial strategy, that is intels job.

Always seems sensible thing to do when buying a product that will depend on future updates from the company.

 

[ajsdkflsdjfio]

It matters for the future of Intel consumer discrete GPUs and their buyers - Pat was fired two weeks ago and the company is under the new bean-counting MBA-toting management, so if Arc loses too much money they might shut it all down next year, other than that I guess it does not really matter.



Intel is absolutely losing money on Battlemage, the only question whether it's a negative gross margin or it's slightly positive, other than that it's a huge money loss for certain, I guess it's all academic now because its future would not change even if they are making +20% gross, that's not enough - Intel needs 50%+ margins, stock market demands it, any product that dilutes it is already bad for the whole company, that's how bean counters think. Only way to counter this is to make low volume and creatively account for R&D as shared among iGPUs too, but that only goes so far.

From what i've seen the new management is MORE focused on products not less. Their bean counting is mostly about foundry not the graphics department. They also seem to be intent on trying to compete with Nvidia and AMD in AI and guess what's crucial to that venture, their graphics department namely xe3 and falcon shores. The R&D put into GPUs was going to happen anyways and is necessary to catch up to Nvidia/AMD even if it doesn't immediately turn a profit. Releasing net-zero profit cards based on R&D you were already doing anyways can only do good for Intel not the other way around, by establishing some semblance of market share and providing a testbed for software development on XE.

They clearly priced it to sell (which in itself invites some thoughts why).

Do you think reviews would have been as positive if the price was $300 or $350-375? To have sustainable discrete product they need 45-50% gross margins, maybe even higher to justify the aggro.

Reviews absolutely still would have been good if the product were 300$ given that it beats the 4060 at 1080p, raytracying, and offers 12GB of vram making the only viable budget 1440p card on the market.

45-50% gross margins may be necessary to pay for R&D and other expenses and profit on top of that. But this is their 2nd generation which means it's not good enough to turn a profit for the entire graphics division yet and IDK why anybody is surprised at that. As with entering any new market segment you are going to be in the red for at least the first few years if not longer. Intel may be struggling financially but graphics is part of their core business going forwards and even the bean counters at intel likely recognize that.

 

[Win2012R2]

the new management is MORE focused on products not less

CPUs, which is the core business where they are losing massive market share especially in Data Center that brings almost no money, plus big risk to Laptops share that can kill them if they lose there too.

GPUs don't even factor in this - they can go to NVIDIA tomorrow and offer to kill their GPUs in exchange for NVIDIAs iGPUs across all Intel line fabbed on Intels 18A, unless regulators block this (which I think should happen, but who knows), or do that with AMD which will be cleared by regulators, either way it's 💀 to Arc

They also seem to be intent on trying to compete with Nvidia and AMD in AI

That boat has sailed, even AMD is barely on it, Intel board understands that bit well.