News Intel GPUs - Falcon Shores cancelled

[Win2012R2]

Please explain.

💀💀💀

^^^^^^^^^^^

This - for GPUs.

oneAPI might still have some legs left in HPC (CPUs) though. Market can't handle 3 different APIs - CUDA is top and the closest distant second option is ROCm.

Frankly until EU legislates directly or ECJ rules clearly that it is legal in EU to have compatibility translation layer from CUDA API to say ROCm or even oneAPI capable hardware nothing will happen with CUDA moat, it needs to be breached that way, NVIDIA banned this in their ToS, but that won't hold too long.

"Source: I made it up"

Out of professional interest I keep an eye on this field, fairly easy to verify what I said by spending little time Googling.

 

[ajsdkflsdjfio]

Out of professional interest I keep an eye on this field, fairly easy to verify what I said by spending little time Googling.

Your "Googling" is coming up with 20k$ for an N5 wafer so....

"I provided sources to the costs, they are also easily obtained by Googling
1) tsmc n7 wafer cost
2) tsmc n4 wafer price"

 

[Win2012R2]

And what's your Googling coming up with, have you got some alternative sources contradicting what I have offered?

So far it appears you've got squat whereas I've provided the best currently freely available.

 

[DavidC1]

@adroc_thurston Battlemage indeed does have noticeably better perf/W and perf/mm2 improvements.

The performance is about equal to Alchemist at the same clock. So if it was at same clocks, your point would be somewhat valid. However, a new node brings ~20% improvements, and BMG clocks more than 20% higher, so based on that all the perf/W improvements on a new node should be spent on clocks, yet it does that at lower power.

27% improvement in perf/W ISO-node according to TPU results. There's no argument here, it's settled. If it sucked like ACM, then there's no way you can get 20% clock boost plus 27% difference in power use, and this is at a card-level, not chip level.

For mm2, N5 brings ~50% density gain over N7, but ACM uses N6. N6 is 18% increased density over N7. So the density advantage of N5 over N6 is 27%, which is not enough to make up for difference in die size. ACM is 48% larger. A straight shrink of A770 would be 320mm2, or 18% larger than BMG.

Combined with the clock speed increase, a 14-20% improvement over A770, meaning 34% to 40% increased perf/mm2 ISO node. The worst case scenario if you assume N5 is 56% denser than N6(and it's not) you are still talking about 10-15% advantage in perf/mm2. And again N5 having 84% increased density over N7 doesn't happen with actual chips, maybe on Cortex A710.

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 was expecting Battlemage to do better in UE5 games, but apparently it does not despite having Execute Indirect support. It's not the only engine that uses it, but it was supposed to have outsized impact.

Also Battlemage does better in 1440p and 4K than Alchemist, and in some games noticeably so which is a surprise, as it's also a bit contrary to expectations.

This tells me there's some possibility that they could intro a driver to bring maybe ~10% improvement in 1080p. We still yet to get a driver that improves DX11 overhead either. We have it on a game by game basis, but we were supposed to have one that doesn't need whitestling like it does now.

The driver alone may tip the scales enough to counter RDNA4 and RTX 5000 series competition, and maybe they can aim it for when the competitor card launches. Also, they may need the driver because the B770 is going to be much faster and thus have more bottlenecks.

As of right now, the card is much, much better than Alchemist. The games seem stable, and the claims of improved compatibility is true. Also past is not a guarantee of future. RDNA4 and RTX 5000 still needs to execute, and even the best can falter at times.

@Win2012R2 I highly doubt N5 is $16K as you claim. 2021 is a while ago, and it does get cheaper over time, for obvious reasons. It's not a new node anymore, and initial costs would have been paid off long ago, yields improve, and the whole point of mass production lowers costs overall.

 

[Hans Gruber]

@adroc_thurston Battlemage indeed does have noticeably better perf/W and perf/mm2 improvements.

The performance is about equal to Alchemist at the same clock. So if it was at same clocks, your point would be somewhat valid. However, a new node brings ~20% improvements, and BMG clocks more than 20% higher, so based on that all the perf/W improvements on a new node should be spent on clocks, yet it does that at lower power.

27% improvement in perf/W ISO-node according to TPU results. There's no argument here, it's settled. If it sucked like ACM, then there's no way you can get 20% clock boost plus 27% difference in power use, and this is at a card-level, not chip level.

For mm2, N5 brings ~50% density gain over N7, but ACM uses N6. N6 is 18% increased density over N7. So the density advantage of N5 over N6 is 27%, which is not enough to make up for difference in die size. ACM is 48% larger. A straight shrink of A770 would be 320mm2, or 18% larger than BMG.

Combined with the clock speed increase, a 14-20% improvement over A770, meaning 34% to 40% increased perf/mm2 ISO node. The worst case scenario if you assume N5 is 56% denser than N6(and it's not) you are still talking about 10-15% advantage in perf/mm2. And again N5 having 84% increased density over N7 doesn't happen with actual chips, maybe on Cortex A710.

I was expecting Battlemage to do better in UE5 games, but apparently it does not despite having Execute Indirect support. It's not the only engine that uses it, but it was supposed to have outsized impact.

Also Battlemage does better in 1440p and 4K than Alchemist, and in some games noticeably so which is a surprise, as it's also a bit contrary to expectations.

This tells me there's some possibility that they could intro a driver to bring maybe ~10% improvement in 1080p. We still yet to get a driver that improves DX11 overhead either. We have it on a game by game basis, but we were supposed to have one that doesn't need whitestling like it does now.

The driver alone may tip the scales enough to counter RDNA4 and RTX 5000 series competition, and maybe they can aim it for when the competitor card launches. Also, they may need the driver because the B770 is going to be much faster and thus have more bottlenecks.

As of right now, the card is much, much better than Alchemist. The games seem stable, and the claims of improved compatibility is true. Also past is not a guarantee of future. RDNA4 and RTX 5000 still needs to execute, and even the best can falter at times.

@Win2012R2 I highly doubt N5 is $16K as you claim. 2021 is a while ago, and it does get cheaper over time, for obvious reasons. It's not a new node anymore, and initial costs would have been paid off long ago, yields improve, and the whole point of mass production lowers costs overall.

The efficiency gains going from N6 to N5 are not 20% but are significant. Intel Battlemage has greatly increased efficiency way beyond the silicon efficiency gains. Ignore the fps between the B580 and RX 6700 in this video. Look at the power consumption numbers only. The B580 is nearly half the power consumption of the 7nm RX 6700. There are different youtube video comparisons between cards and Battlemage power usage is insanely low. I do think Intel needs to release their B770 sooner rather than later. The RDNA 4 and 50 series is coming soon.

 

[beginner99]

The driver alone may tip the scales enough to counter RDNA4 and RTX 5000 series competition, and maybe they can aim it for when the competitor card launches.

Yeah they will have a hard time. I myself am waiting to see what AMD and NV has to offer. I will probably go with AMD given that NV will likely offer too little vram once again for the price point and B580 is a tad too slow for my taste but more so the unknown compatibility with old games.

 

[DavidC1]

The efficiency gains going from N6 to N5 are not 20% but are significant.

It is around there if you are using it all for performance gains. You can get more if you focus on performance per watt, but then you won't get extra clocks.

TSMC Details 5 nm

TSMC details its 5-nanometer node for mobile and HPC applications. The process features the industry's highest density transistors with a high-mobility channel and highest-density SRAM cells.

fuse.wikichip.org

From a device power and performance perspective, TSMC says devices offer 15% higher speed at iso-power or, alternatively, 30% lower power at the same speed.

N5 offers 15% higher speed, which Battlemage already does better on. So saying Battlemage gains are entirely process is either misleading or they don't know what they are talking about but pretending they do.

I do think Intel needs to release their B770 sooner rather than later. The RDNA 4 and 50 series is coming soon.

Right, but according to current rumors, they are nowhere near to launching B770. They aren't doing this on purpose of course. They aren't ready. People say it's sometime roughly around Q2 or even Q3 of this year.

A 32 Xe2 core performing like 4070 Super requires clocks in the 3.3-3.5GHz range, and it'll be a challenge doing that with a bigger die.

I also think we shouldn't automatically assume the B770 will be really good in perf/$ just like the B580 is. It'll face next generation competition and unless they both fall flat on their faces, they probably can't price it beyond $449 USD at the very maximum even if it performs 4070 Super level.

 

[adroc_thurston]

ISO-node according to TPU results.

Cyberpunk is an outlier.

improvement over A770, meaning 34% to 40% increased perf/mm2 ISO node

Where
It's a 270mm^2 die.

 

[Win2012R2]

I highly doubt N5 is $16K as you claim. 2021 is a while ago, and it does get cheaper over time, for obvious reasons

$18-20k pricing for 2024-2025, claimed by source: https://www.tomshardware.com/tech-industry/tsmc-may-increase-wafer-pricing-by-10-for-2025-report

If you got better source showing that price is different then please share, all evidence from TSMC so far points to advanced nodes being charged at higher and higher prices due to high demand, plus N5 is too new for fabs to be fully depreciated (can take longer than usual due to higher fab prices in the first place).

 

[DigDog]

AAAAGH I HAVE NO THERMAL PASTE

 

[igor_kavinski]

AAAAGH I HAVE NO THERMAL PASTE

You can use toothpaste for about 15 days:

https://www.reddit.com/r/AskPCGamers/comments/gp472v/whats_the_best_alternative_for_thermal_paste

Also:

 

[ajsdkflsdjfio]

$18-20k pricing for 2024-2025, claimed by source: https://www.tomshardware.com/tech-industry/tsmc-may-increase-wafer-pricing-by-10-for-2025-report

If you got better source showing that price is different then please share, all evidence from TSMC so far points to advanced nodes being charged at higher and higher prices due to high demand, plus N5 is too new for fabs to be fully depreciated (can take longer than usual due to higher fab prices in the first place).

How many times have I got to say it, that source claims 18k listed price for N4 which CANNOT be used for N5, all other sources say N5 is 16K AT MOST and multiple people have already come forward with evidence showing how N3B was discounted to below 15k$ already with a 40% discount, if you think Intel and other partners didn't have discounts on even older nodes then you are delusional.

You know what's funny too? On the same exact tomshardware article you hold as the single truth, according to their graphs, TSMC N5 and N7 are at the same prices relative to their original 2021 prices with similar increases up until 2024:

If you look at the graph very carefully you (if you are rational) may see that as of 1Q24 (the last actual data point they had) TSMC 5nm and 7nm were both around 5-10% above 2021 prices and this trend was shown to continue even in tomshardware's own projections until at least 2025. Which means that the majority of battlemage cards produced thus far have been on that old pricing. It's only after 2025 which Morgan Stanley expects a 10+% increase in N5 prices, which tbh I don't think they can accurately predict anyways. Nonetheless its irrelevant to our discussions since Intel likely pre-purchased wafers anyways regardless of whether they continue producing battlemage in 2025 or not.

Where
It's a 270mm^2 die.

He literally just showed u where... You give Adroc Thurston a bad name

 

[Win2012R2]

TSMC N5 and N7 are at the same prices

N5 pretty clearly going up in price faster than N7 in 2025 - so what if rate of inflation was close before, that does not mean they had same absolute prices!

And where on that chart nodes N6 and N4 if they are separate

It's only after 2025 which Morgan Stanley expects a 10+% increase in N5 prices

Q1 25 is pretty clearly taking off for N5, I guess if Intel stops making GPUs on it then they will be ok.

Intel likely pre-purchased wafers

That's not how it works for the likes of Intel - TSMC auctions wafers on regular basis, that's how they put in price increases (or decreases if utilisation is low), obviously they won't allow long term pre-purchases for advanced nodes at cheap price unless it's Apple level who banrolls whole new fabs.

 

[Hans Gruber]

Here is another B580 gaming comparison. This time the RTX 4060. The 4060 is using 4N silicon, which is far superior to N5. The B580 power consumption numbers are lower on the B580 in real world gaming.

 

[ajsdkflsdjfio]

N5 pretty clearly going up in price faster than N7 in 2025 - so what if rate of inflation was close before, that does not mean they had same absolute prices!

No one said they had same absolute prices but at the same time the prices are not significantly different from quoted 2022 N7 vs N5 prices of 10k and 16k respectively, atleast according to this tomshardware post you love so much.

And where on that chart nodes N6 and N4 if they are separate

Keep being ignorant. Firstly these numbers are all market estimations not actually prices. The prices vary from sale to sale and company to company, meaning that prices are in fact different for the exact node you are using like N3E vs N3x or N5 vs N4. Secondly if they were talking about N5 prices they would have said N5 not N4. Them saying N4 means that they are specifically referring to N4 separate from N5 as a whole.

Q1 25 is pretty clearly taking off for N5, I guess if Intel stops making GPUs on it then they will be ok.

"Pretty clearly" according to who? Morgan Stanley's market projections? The story is a lot more complicated than that but sure let's say their projections are the absolute truth. According to their projections N5 class wafers are going up by 10-15% over the next two years right? Guess what other products are also going to be produced on N5 family of nodes? RDNA 4 and Blackwell. We'll see what happens with those generations, but I'd say that blackwell isn't going to provide the low-end with any improvements and RDNA4 will have to bring a +25% price/perf increase to compete at 250$ with BMG.

That's not how it works for the likes of Intel - TSMC auctions wafers on regular basis, that's how they put in price increases (or decreases if utilisation is low), obviously they won't allow long term pre-purchases for advanced nodes at cheap price unless it's Apple level who banrolls whole new fabs.

As for pre-purchasing wafers, it absolutely is how it works. There are already numbers out there with the number of N2 wafers Apple/Intel/AMD have pre-purchased from TSMC . Arrow lake N3B wafer contracts were placed more than 2 years ago at least while the actual products released this year. As for pricing of these pre-purchased wafers, it is a lot more complicated than one singular price figure. You can't just assume just because the average sales price of a wafer goes up in year 20XX that every single company producing chips from TSMC experiences the same exact price change, if at all. The nature of these contracts is a lot more complicated than anybody of us realizes and also is confidential between these companies.

I guess if Intel stops making GPUs on it then they will be ok.

Some random market analyst from Morgan Stanley pulls out of his ass projections for markets +2 years from now -> You apply this 1 to 1 exactly to Intel even though they likely pre-purchased their wafers long before 2025 and most importantly the details of their agreements with TSMC can vary in prices by a ton -> Then you use this to extrapolate that Intel battlemage will be in the negative for each card produced in 2025 from what reasoning?

You don't even know the BOM to begin with or even the cost of a BMG die outside of rough projections. You don't know the margins from which other cards in the price point are making. You don't know the cut that the AIBs are taking. You don't know the parametric yield of the die itself. A random analyst's projections for the next 2 years is hardly from a reputable source. You don't know the specific nature of Intel's agreement with TSMC and the pricing of the wafers that Intel themselves paid for.

Do you see how many assumptions you are making?? Either way there is a pretty easy way to see how likely/unlikely it is that battlemage is losing money for each card Intel produces. If Intel continues to produce battlemage cards and continues to sell them at competitive pricing, it's highly likely they aren't in fact losing money for each card they produce.

 

[Win2012R2]

these numbers are all market estimations

Have you got anything more credible than that?

 

[ajsdkflsdjfio]

Have you got anything more credible than that?

I go on further to explain how many assumptions you are marking in your analysis but ok. You assume that a x% change in N5 wafer prices overall reflects one to one in Intel's specific contracts, you assume that N5 price increases somehow only affects BMG and not RDNA4/Blackwell.

You assume that BMG margins cannot take a further 15% die price increase which is braindead because even if the die were 100$ that's only a 15$ increase and you cannot suddenly assume BMG is in the red from a 15$ increase without knowing the BOM or AIB margins. Even if that were the case, just price BMG at 265 instead of 250 LOL.

This one point is enough tbh anyways. Why do you take a random Morgan Stanley projection as a magic ball of truth that can predict the prices of TSMC wafers. It's like reading one Bloomberg article saying that Intel stock is a must-buy in 2024 and then betting your entire house on it. Even if the people making these projections were the smartest men alive they still wouldn't be able to predict the future to that accuracy.

Even with Blackwell/RDNA4 coming I don't see how this spells the end to BMG competitiveness. the RX6800xt was an excellent midrange card in 2022 at 500$ before lovelace/RDNA3, and for the most of 2023 and 2024 has still been a excellent choice for a midrange card AT THE SAME EXACT PRICE. New generations don't mean better price/perf. Even if they do bring price/perf increases, we have seen that this mostly applies to upper-range cards NOT budget cards. Why do you think battlemage is received so well in the first place? Because budget cards have not had serious contenders enter the market for many years now. It is entirely possible that even with the releases of Blackwell/RDNA4, BMG will continue to be competitive without having to cut down its prices. It's happened before for countless cards like the rx6700,6800xt etc.

I understand critical thinking is an issue for you, so maybe take a break from making such bold predictions like that BMG is doomed in 2025 or that Intel ARC is going to be canned.

 

[DigDog]

so the situation is, i do have paste, but it's conductonaut, and it absolutely refuses to spread on the wrath stealth cooler base.
the base is not aluminium, fyi.
and i cannot afford to have little droplets of liquid metal rain on my mobo.
i ordered a tube of boring old MX-6 and amazon is sending it thursday.

 

[NTMBK]

And what's your Googling coming up with, have you got some alternative sources contradicting what I have offered?

So far it appears you've got squat whereas I've provided the best currently freely available.

WWYBYWB

 

[Win2012R2]

I go on further to explain how many assumptions you are marking in your analysis but ok. You assume that a x% change in N5 wafer prices overall reflects one to one in Intel's specific contracts, you assume that N5 price increases somehow only affects BMG and not RDNA4/Blackwell

Can you produce better assumptions backed up by credible sources, for example such saying that Intel gets X% discount on average wafer prices, can you? I reckon you can't.

I can though easily dig up something relevant -

October 31, 2024 "A new report from Reuters is shedding light on behind-the-scenes tensions between Intel and TSMC that might have caused a rift. Intel has been using TSMC to fab some of its chips for several years now and is using it exclusively for its current desktop and mobile architectures. Apparently, it's paying full price for those wafers, thanks to comments made by its CEO, Pat Gelsinger.

The Reuters report says it's based on four sources who have knowledge of an arrangement between the two companies, which included a hefty discount on 3nm wafers. Reuters says TSMC was offering Intel a steep 40% discount on the the wafers, which it says cost Intel $23,000 each"

Source: https://www.extremetech.com/computi...eep-discount-at-tsmc-with-dismissive-comments

Now if the above report sourced from Reuters is true, then I find it unlikely that Intel would have kept any other discounts on N5 and other nodes, perhaps the prices even gone up? The report obviously had lag, probably months, so in this case that would have been just around the time Battlemage was manufactured, maybe I should factor in 15-20% uplift on N5 wafers into my BOM calculations...

 

[ajsdkflsdjfio]

Can you produce better assumptions backed up by credible sources, for example such saying that Intel gets X% discount on average wafer prices, can you? I reckon you can't.

I can't which is why I don't assume that Intel is getting a discount or not getting a discount. I don't assume that prices are or aren't going to rise +15% in 2025 either. And you know why I (or you or anybody else) cannot provide credible sources to back up claims like these?? It's because THEY DON'T EXIST !!!!!!!!1!!1!1!! These things aren't publicly known and anybody or any news site claiming to have an inside scoop is likely to be BS 99 to 1.

I can though easily dig up something relevant -

October 31, 2024 "A new report from Reuters is shedding light on behind-the-scenes tensions between Intel and TSMC that might have caused a rift. Intel has been using TSMC to fab some of its chips for several years now and is using it exclusively for its current desktop and mobile architectures. Apparently, it's paying full price for those wafers, thanks to comments made by its CEO, Pat Gelsinger.

The Reuters report says it's based on four sources who have knowledge of an arrangement between the two companies, which included a hefty discount on 3nm wafers. Reuters says TSMC was offering Intel a steep 40% discount on the the wafers, which it says cost Intel $23,000 each"

You take this disagreement from Pat and TSMC resulting in no N3B discounts, apply it to N5 and further apply it to intel's future pricing of N5 when they have a completely different CEO/CEOs. This report alone isn't even enough to prove that Intel lost some imaginary 40% discount either.

Reuters recently reported on a supposed below 10% yield for their 18A node citing "confidential industry sources". Unsurprising to anyone with above room-temp IQ, these claims were largely disproven and shown to be nothing more than FUD with 0 actual evidence to back them up.

Besides even if the 40% thing were true, why do you think its likely that TSMC also removed discounts from N5? Do you know the inner workings of TSMC executives' decision making? Why would they go all out from the get-go from one dispute with Gelsinger? Couldn't it be possible that they only removed the N3B discount as a warning without further removing all the other discounts?? Either could be true, but its impossible for any of us to know yet here you are doubling down on it. You are either a genius or a dumbass, and there are a lot more dumbasses in the world than geniuses so...

 

[Win2012R2]

This report alone isn't even enough to prove that Intel lost some imaginary 40% discount either.

It looks credible, especially if you know the background as to why possibly Intel would even get such a huge discount for such a new process in the first place. Either way it's enough to make some rough assumptions about how TSMC and Intel relationship is going, and it's miles better than any of your sources - which are none.

Besides even if the 40% thing were true, why do you think its likely that TSMC also removed discounts from N5?

I don't think TSMC would have given any discounts for N5. The reason Intel got big discount for N3B is because Pat early on prepaid it (2021 I think) - alongside with Intel, this in my view was desperate measure to deny AMD volume on that node, however he could not do the same for N5 as it was already fully prepaid by Apple.

Reuters recently reported on a supposed below 10% yield for their 18A node

And Pat "bet the farm on 18A" Gelsinger got fired shortly after - yeah, 18A is in "great" shape, can safely assume that too now

 

[maddie]

Back in your cages now, I declare a draw.

 

[ajsdkflsdjfio]

Clearly you aren't capable of reasoning so let's make another bet on top of our celestial bet:

Based on your assumptions here, if what you say is true it'd be impossible for Intel to continue producing and selling battlemage for more than a couple months into 2025, so I will bet the opposite.

If Intel effectively or outright cancels BMG in the first 9 months of 2025 (no more sales) I will never speak again on any Intel-related thread, otherwise you will never speak on any Intel-related thread.
Intel related thread means that the thread title is about intel and its products, if intel comes up in another thread you are free to comment.

I would not normally be willing to bet that BMG endures for 9 months but there is nothing else I can do since I can't reason with you and spamming these threads with thinly veiled hate speculation isn't bannable.

Agreed? or would you rather change the terms.

 

[QueBert]

I saw the Asrock Steel Series B580 and fell in love. For some reason it's 3 slots, and of course my case only supports 2.5 lol. I'm confused why it's 3 slots and has 2 8 pin connectors but it's sure a nice looking card.