Christmas Season CPU Best Sellers Rank on Amazon: AMD vs. Intel

[dkr1986]

I don't really understand what you mean by this statistics.

I think the point is DIY market is very very small (as evidenced by the fact AMD is more dominant in DIY and AMD's revenue numbers).

Intel made $10.22B in revenue selling desktop CPUs in last twelve months (LTM).
Intel made $19.23B in revenue selling notebook CPUs in LTM.
Intel's Operating margin in Client Computing group (Desktop+notebook) is 36.7% LTM
Note that the above numbers are reported by Intel without the impact of Intel Foundry (reported separately - as if Intel operated as a fabless semi company like AMD).

AMD doesn't breakdown its desktop vs notebook revenue split in earnings.
AMD made $6.22B in revenue selling both notebook and desktop in LTM.
AMD' operating margin in Client group is 10% LTM (it was 14% last 2 qtr, so accelerating a bit).

So yes Intel sells more in desktop alone than AMD sells together in client group showing DIY is a very small market. This because DIY market has been skewed towards AMD for at least 18 months (I think 12th gen was last good one for Intel in DIY).
Intel managed to this with a CPU that was considerably power hungry (worse perf/watt than AMD) and ran hotter (cooling requirement is even worse with MB vendors not following Intel power guidelines) on an inferior node process (Intel 7) for the last 3 gens.

Now with LNL and ARL-S\H\HX\U (actually MTL-U Refresh), both of these problems are solved. I think Intel will keep its current market share in client group this gen except DIY which is gaming focused and will be ruled by AMD's x3D chips. This segment as shown above is very small but very vocal (can cause mindshare loss in the long run). But since these products are mostly costly to manufacture (external foundry and advanced packaging), Intel's client group (thereby Intel Products) margin will reduce in 2025.

 

[dkr1986]

... and there is the real issue for Intel (not Amazon sales of desktop processors).

AMD got out ahead of Intel with designs that were architected to scale from thin and light to high core count DC products. They spent time getting chiplets working very well, and designed a very good working IOD with plenty of bandwidth to feed tons of cores.

For the past several years, AMD has totally dominated Intel in DC. Even the new Granite Rapids release with Intel pulling out all the stops still lags AMD's Turin by an average of 40% and is priced higher. At least Intel is no longer being trounced by triple digit performance deficits!

I also worry a little about Clearwater Forest. The Darkmont core is supposed to be a "slightly modified Skymont" core. From what I can see, that means it will not have SMT as I believe that this would require a pretty significant overhaul of the entire core to implement.

At this time, it looks like Skymont is about on-par with Zen 4 in terms of IPC.... as long as you don't throw MT at it, or AVX512. Since it is silly to talk about anything other than MT with DC products, Zen 5c will have a roughly 40% performance per clock (vs instructions per clock) advantage over the current Skymont core due to SMT.

Between SMT and AVX512 .... and Zen 5c's current performance on N3E for Turin, I do wonder if an entire product based on Darkmont (E cores) can hang with Turin Dense in DC workloads.

Is Clearwater Forest supposed to also compete with Turin (non D) and its full Zen 5 cores?

Anyway, this is where I think the real fight for dominance in x86 from a profit standpoint will be.

I believe GNR lags Turin classic by roughly 20% in terms of Perf/watt if we compared 1P congratulations in Phoronix review. (it used to be nearly 40% for SPR vs Genoa generation) That 40% lag for GNR is based on Phoronix review of 2P configuration where GNR was not scaling well for 2P. Michael also said in Reddit and Twitter that there was something weird thing going on in some benchmarks for GNR in 2P configuration. He said Intel acknowledged and reproduced the issues but didn't get back to him in time with a fix when Turin benchmark launched. I don't know if it was fixed now but Phoronix didn't do a re-review yet.

So Intel actually closed the gap in this gen in DC CPU market (imo). SRF is actually meant for different type of workload compared to Turin Dense and is not an apple to apple comparison based on what I learned ( I believe both L1tech and STH said so in their review of SRF too, that it's not meant for HPC market).

 

[siliconbruh999]

I believe GNR lags Turin classic by roughly 20% in terms of Perf/watt if we compared 1P congratulations in Phoronix review. (it used to be nearly 40% for SPR vs Genoa generation) That 40% lag for GNR is based on Phoronix review of 2P configuration where GNR was not scaling well for 2P. Michael also said in Reddit and Twitter that there was something weird thing going on in some benchmarks for GNR in 2P configuration. He said Intel acknowledged and reproduced the issues but didn't get back to him in time with a fix when Turin benchmark launched. I don't know if it was fixed now but Phoronix didn't do a re-review yet.

The perf/Watt gap was 2X with SPR vs Genoa with complete domination outside of AI now there is no absolute domination in many things there are genuine platform differentiation between both Turin and GNR so it us definitely a good step but need to do a huge jump DMR is supposed to be that they would have moved forward by 3 architecture from redwood cove
Lion Cove
Cougar Cove
Panther Cove ( supposed to be in DMR)

 

[dkr1986]

It doesn't need bailout. At worst case, sellout Altera and Mobileye holdings to repay debt and further delaying Ohio fabs. It also doesn't need a home run. It needs to keep up with competition and recover its financial condition. If just value Intel products alone, it should be more valuable than AMD. Its valuation is shaded by the manufacturing business.

Intel has $24B in cash and short term investments in their balance sheet per Q3'24 10Q.

The CHIPS grant of $7.86B + $3B from Secure Enclave grant probably cash in Intel's balance sheet in the next 2 years.

CHIPs act also provides 25% refundable investment tax credit. This is applicable for all qualifying capex spent on facilities and equipment that can be depreciated and amortized from August of 2022 and that will come to service after Dec 2022.

Intel spent about $25B each year in capex for 2022 to 2024.
Intel's depreciation cost is roughly $9 billion a year (based on LTM). Assuming that is the maintenance capex roughly per year, we have around $36.25B of growth capex invested for this period. Assuming 60% of that is equipment and facilities, Intel may receive 0.25 x 0.6 x 36.25 = $5.5B into Intel balance sheet.

This tax credit will also be applicable for Intel's investment in 2025 too. I don't think Trump admin will be adverse to Intel (IMHO)

Also $10B cost reduction (opex & capex reduction , dividend cut) as announced by Intel recently to be achieved by end of 2025.

They have only received roughly $3B of cash through Arizona SCIP for $30 billion planned investment for 2 fabs. So may be expect another $12 billion from Brookfield over the next couple of years (likely soon as Intel 18A is supposed to be HVM there).

Intel also has a revolver credit of $15 billion that remains unused per recent 10Q. They can draw on that to bridge any gap in funding if required for next couple of years.

Intel's cash flow from operation is still positive for the last twelve months, even with foundry division sucking all the GAAP profit of products team.

Based on this napkin math, I think bankruptcy concerns are overblown at least for next couple of years. All the while their products are getting better in their core market (CPUs) and with products coming back to internal foundry, foundry utilization improves which leads to margin improvement overall.

If consumer doesn't weaken in 2025 & 2026 due to macro economic conditions, I think Intel will have a good next couple of years if they just manage to hold the fort in market share.

 

[OneEng]

I believe GNR lags Turin classic by roughly 20% in terms of Perf/watt if we compared 1P congratulations in Phoronix review. (it used to be nearly 40% for SPR vs Genoa generation) That 40% lag for GNR is based on Phoronix review of 2P configuration where GNR was not scaling well for 2P. Michael also said in Reddit and Twitter that there was something weird thing going on in some benchmarks for GNR in 2P configuration. He said Intel acknowledged and reproduced the issues but didn't get back to him in time with a fix when Turin benchmark launched. I don't know if it was fixed now but Phoronix didn't do a re-review yet.

So Intel actually closed the gap in this gen in DC CPU market (imo). SRF is actually meant for different type of workload compared to Turin Dense and is not an apple to apple comparison based on what I learned ( I believe both L1tech and STH said so in their review of SRF too, that it's not meant for HPC market).

The entire article is here: https://www.phoronix.com/review/amd-epyc-9965-9755-benchmarks#google_vignette

You are correct that something looks wrong with the GNR scaling to 2P, and others have commented on this, but I haven't seen any update on any site saying there is a fix even planned for it, rather on Intel acknowledging it is a problem.

If you look through the article, there are plenty of examples of GNR scaling on 2P though.

Still, if you only look at the 1P numbers AMD has > 18% advantage and that is taking into account the use of more expensive RAM with the Intel offering ... and the GNR processor being more expensive than Turin. There are also an awful lot of benchmarks where Turin is way out front of GNR. I don't know. While Intel is in a much better position with GNR, it still looks pretty bleak until Clearwater Forest and Diamond Rapids.... and that won't likely have any market effect until 2026. That is an awful long time for Intel to continue to lose market share in DC to AMD.

IMO, GNR may help reduce the rate that companies are switching DC to AMD, but AMD will continue to gain share with the current lineup.

 

[OneEng]

Intel's cash flow from operation is still positive for the last twelve months, even with foundry division sucking all the GAAP profit of products team.

Funny math going on over at Intel.

Computing your margin on products without charging the products for the production overhead. Cute.

In my neck of the woods they call that "fooling yourself".

"If only it weren't for that horrible foundry!". IMO, Intel's foundry has been the key to all of their products past successes. Intel foundry has created fantastic value for the company over the decades, and only with a few bad (very bad) decisions (like not moving to EUV in time) did they present problems for product.

Even then, Intel managed to move the bar (at least in desktop) by saying "who cares how much power it uses"? Intel foundry found a way (ok, well almost anyway) to keep those transistors running at the insane power the product needed to be competitive. Of course, this trick doesn't work in server.

Product, on the other hand, has stepped in it time and time again over the decades. Itanium and IA64? Refusal to embrace x64 in the consumer market (to protect Itanic I suspect), and Netburst and the promise of the 10Ghz Tejas.

Now Lion Cove that is larger than Zen 5 with a process advantage, and a pretty bad latency issue throughout Arrow Lake.

I have always thought Intel Foundry was just about as amazing as Willy Wonka's Chocolate factory .

I have sometimes though the processor design strategy team needed to find something else to do for a living .

 

[dkr1986]

The entire article is here: https://www.phoronix.com/review/amd-epyc-9965-9755-benchmarks#google_vignette

You are correct that something looks wrong with the GNR scaling to 2P, and others have commented on this, but I haven't seen any update on any site saying there is a fix even planned for it, rather on Intel acknowledging it is a problem.

If you look through the article, there are plenty of examples of GNR scaling on 2P though.

Still, if you only look at the 1P numbers AMD has > 18% advantage and that is taking into account the use of more expensive RAM with the Intel offering ... and the GNR processor being more expensive than Turin. There are also an awful lot of benchmarks where Turin is way out front of GNR. I don't know. While Intel is in a much better position with GNR, it still looks pretty bleak until Clearwater Forest and Diamond Rapids.... and that won't likely have any market effect until 2026. That is an awful long time for Intel to continue to lose market share in DC to AMD.

IMO, GNR may help reduce the rate that companies are switching DC to AMD, but AMD will continue to gain share with the current lineup.

I understand the point you are making about any verified fix for GNR 2P issue not available yet or planned. All I noticed was SPR and EMR scaled roughly 1.5x for 1P vs 2P (AMD's CPUs too) whereas GNR only did 1.2x on geo mean in Phoronix review (iirc). So I assumed Intel didn't drop the ball with anything major here and it's only a software issue or teething pain in a new platform (hopefully that is the case for Intel's sake)

Also something to note is Geo mean score for GNR 1P dropped as well from Phoronix review of GNR when it launched vs on the review of Turin launched because of certain problematic GNR benchmarks added in latter. Just highlighting some issues here that can affect that %18 advantage.

I have no idea if ~18% perf/watt improvement is enough to break Intel's hold on enterprise server customers (as opposed to hyperscaler/cloud where AMD now holds 50% market share now) . We shall see soon.

 

[TICin852]

My two cents here: The flagship DC products are just like flagship products - for show, and low volume. The arguments about Diamond Rapids vs whatever massive 384 core single socket AMD CPU are a bit overblown (I feel) - it doesn't matter as the price of memory is going to be insane for those servers - Even if you have 12/16 channels of memory, each stick would have to be at least 128gb in order to provide enough memory to keep those cores fed. The actual cost is absolutely insane so I doubt most server owners would be using that.

SRF and GNR only needed to be competitive in the lower core counts to keep the market share. The fact that Intel's little cores are able to match AMD's IPC is a key differentiator in the types of workloads they are going after. GNR will have an impact on margins because as we go into 2025 the platform adoption will stop Intel bleeding and stabilize the DC market - margins might take a hit but will at least stabilize, allowing us to see if Intel will really go bankrupt or if they have enough money to keep going for several more years.

 

[XYang2023]

By the way, I just went to the shop to collect the ARC B580 Limited Edition (produced in Vietnam). The store attendant showed me their order book, and all B580 graphic cards were sold out before they even arrived at the store. That store alone sold around 100 units. I assume there are about 50–100 similar stores across Australia. This suggests that the total units sold within Australia could be around 5000-10000. Additionally, there are other regions like China, Southeast Asia, Japan, Korea, India, North America, etc.

 

[DaRAGingLunatic]

Intel has $24B in cash and short term investments in their balance sheet per Q3'24 10Q.

The CHIPS grant of $7.86B + $3B from Secure Enclave grant probably cash in Intel's balance sheet in the next 2 years.

CHIPs act also provides 25% refundable investment tax credit. This is applicable for all qualifying capex spent on facilities and equipment that can be depreciated and amortized from August of 2022 and that will come to service after Dec 2022.

Intel spent about $25B each year in capex for 2022 to 2024.
Intel's depreciation cost is roughly $9 billion a year (based on LTM). Assuming that is the maintenance capex roughly per year, we have around $36.25B of growth capex invested for this period. Assuming 60% of that is equipment and facilities, Intel may receive 0.25 x 0.6 x 36.25 = $5.5B into Intel balance sheet.

This tax credit will also be applicable for Intel's investment in 2025 too. I don't think Trump admin will be adverse to Intel (IMHO)

Also $10B cost reduction (opex & capex reduction , dividend cut) as announced by Intel recently to be achieved by end of 2025.

They have only received roughly $3B of cash through Arizona SCIP for $30 billion planned investment for 2 fabs. So may be expect another $12 billion from Brookfield over the next couple of years (likely soon as Intel 18A is supposed to be HVM there).

Intel also has a revolver credit of $15 billion that remains unused per recent 10Q. They can draw on that to bridge any gap in funding if required for next couple of years.

Intel's cash flow from operation is still positive for the last twelve months, even with foundry division sucking all the GAAP profit of products team.

Based on this napkin math, I think bankruptcy concerns are overblown at least for next couple of years. All the while their products are getting better in their core market (CPUs) and with products coming back to internal foundry, foundry utilization improves which leads to margin improvement overall.

If consumer doesn't weaken in 2025 & 2026 due to macro economic conditions, I think Intel will have a good next couple of years if they just manage to hold the fort in market share.

From what I have been hearing from businesses here in Bendigo, Victoria, Australia, is that the higher interest rates on mortgages have really dealt a blow to consumer spending by about 30% compared to 2023.


So we the consumers have Arrow Lake for desktop until Nova Lake launches on 14A? When is that? End of 2027?

 

[siliconbruh999]

From what I have been hearing from businesses here in Bendigo, Victoria, Australia, is that the higher interest rates on mortgages have really dealt a blow to consumer spending by about 30% compared to 2023.


So we the consumers have Arrow Lake for desktop until Nova Lake launches on 14A? When is that? End of 2027?

NVL is 18AP in H2 26 if the rumours are true

 

[XYang2023]

Running some code on Intel B580 GPU

x.com

x.com

 

[XYang2023]

Got PyTorch running on Intel B580 GPU. The overall experience is relatively smooth. Will run some tests later.

 

[MKWVentures]

From what I can see, Both Sierra Forest and Granite Rapids are significantly behind in DC workload testing:

https://phoronix.com/benchmark/result/amd-5th-gen-epyc-9005-turin-performance-benchmarks/geometric-mean-of-all-test-results-result-composite-a5ge9tpb.svgz

This is the deficit that Clearwater forest has to erase. As I said earlier, at least Intel is no longer losing by tripple digit numbers in these tasks as it was; however, Intel prices are significantly higher than Turin the last time I looked. Not sure I would want to be an Intel DC salesman at this time. How to you pitch a higher cost, lower performance, lower efficiency product?

Clearwater forest may well help clean this up; however, as pointed out, I don't see this helping Intel financially in 2025.

And then there is the AMD Zen 6 which looks to be released around Early to mid 2026 with its reported 32 core CCD's (Desktop I think is rumored to be only 12 core CCD's). 384 core single socket anyone?

Intel needs a home run with Clearwater forest IMO.... or a bailout by the US government (or both).

Again, remember that DC CPUs ramp slowly .... a peak 2 years after launch is quite typical. Plus granite and Sierra forest just launched this year. nothing will get better in 2025, I imagine thats what the board was told. In 2026 it will start..... But Intel has to spend tons of money it doesn't have to make that happen.... and it owes a lot to Brookfield/apollo.

even If 18A works perfectly and all products launch on time, it is still a difficult financial challenge to become successful. Do a cash flow worksheet and see if you can make the numbers add up. I am guessing Stacy has a plan. Intel should not and will not get a bailout. We covered why this would happen back in 2021.

 

[MKWVentures]

It doesn't need bailout. At worst case, sellout Altera and Mobileye holdings to repay debt and further delaying Ohio fabs. It also doesn't need a home run. It needs to keep up with competition and recover its financial condition. If just value Intel products alone, it should be more valuable than AMD. Its valuation is shaded by the manufacturing business.

100% agree on valuation. If IFS didnt exist and Intel outsourced all manufacturing, Intel should double in value as a minimum

I do not show that Intel needs Ohio Fabs anytime soon. Where is the demand for Ohio Fabs before 2029?

 

[siliconbruh999]

Again, remember that DC CPUs ramp slowly .... a peak 2 years after launch is quite typical. Plus granite and Sierra forest just launched this year. nothing will get better in 2025, I imagine thats what the board was told. In 2026 it will start..... But Intel has to spend tons of money it doesn't have to make that happen.... and it owes a lot to Brookfield/apollo.

I think they will ramp in 25 just fine case The Low core count variants are launching as well and ARL-U on Intel 3 as well Intel 3 should ramp nicely they already have the dd under control for it

even If 18A works perfectly and all products launch on time, it is still a difficult financial challenge to become successful. Do a cash flow worksheet and see if you can make the numbers add up. I am guessing Stacy has a plan. Intel should not and will not get a bailout. We covered why this would happen back in 2021.

Stacy the same guy who was at Boeing during yhe debacle i am having deja Vu with him

 

[dkr1986]

Funny math going on over at Intel.

Computing your margin on products without charging the products for the production overhead. Cute.

In my neck of the woods they call that "fooling yourself".

I don't think there is anything wrong with that margin calculation.

I believe Intel Product group margin is being calculated with market wafer pricing for similar node process in market used for making those products. This is like Intel Product team buying wafers from TSMC prices from Intel foundry. (I think that is what they are supposed to be doing ideally but in reality, there may be some bias or massaging numbers stuff going on).
This way, Intel foundry being not a sound business financially (high cost, inefficiency, low utilization, idle fabs, huge depreciation cost etc) is not Intel products team's concern and it gives an opportunity for Intel management to identify inefficiencies and fix that.

Now I fully acknowledge that Intel Foundry was & is a strong moat for Intel that made them king of compute for the last couple of decades. But that does not mean they were run efficiently or profitably (in an IDM structure, it was not even a concern). Even in 2021, when Intel was making record revenue of $79b and profits, Intel foundry was making an operating loss of $5b!. Their operating margin at that time was 25% for the entire business, so Intel products must have had much higher margin to offset this foundry loss.
https://www.intc.com/filings-report...0000050863-24-000068/0000050863-24-000068.pdf

So I don't believe Intel Foundry division was ever profitable on their own and being an IDM, nobody at Intel cared. Pat and Dave spoke about it many times about too many steppings, hot lots etc.

The real problem I have with Intel margin calculation there is another operating expense bucket called "Unallocated Corporate Expenses" that include Stock Based Compensation but is not included in segment operating profit calculation. I believe that is supposed to be included in COGS, R&D and SGA depending on whoever gets these stocks. So, I do think Intel product margin is overstated a little due to this. I am not a finance person to truly know if what they are doing this way is acceptable per GAAP but these results were audited and approved.

Now Lion Cove that is larger than Zen 5 with a process advantage, and a pretty bad latency issue throughout Arrow Lake.

I think(imho) Lion Cove is actually not that bad. There is definitely a latency issue due to the tiles approach vs monolithic but Robert Hallock (Intel) said in recent interview its firmware/microcode issues that is overblowing this issue. Also recent patches/updates by game developers had improved gaming performance of ARL-S, for example Cyberpunk 2077 had a huge uplift that puts the 285K in line or slightly above 14900k while consume less power. So with time, some of these issues will be ironed out. I don't expect it to beat 9940x or 14900K across the benchmarks anytime.

Also with regard to process nodes, Zen 5 is on N4P (if I am not mistaken), so N3 (which N3B now) is 3 to 8% better in power and 0-4% better in performance per following graphic. And I am pretty sure, N3 is very costly too.
Considering this is one of the first time, Intel engineers are working with an external foundry to make their products and first tiles approach on desktop sort of Zen 2 for AMD (based on what I see, Zen 2 was not received well for gaming too), I think we can cut them some slack this one time.

 

[siliconbruh999]

Got PyTorch running on Intel B580 GPU. The overall experience is relatively smooth. Will run some tests later.

mine as well btw you can't install xpu and cuda at the same time

 

[siliconbruh999]

I don't think there is anything wrong with that margin calculation.

I believe Intel Product group margin is being calculated with market wafer pricing for similar node process in market used for making those products. This is like Intel Product team buying wafers from TSMC prices from Intel foundry. (I think that is what they are supposed to be doing ideally but in reality, there may be some bias or massaging numbers stuff going on).
This way, Intel foundry being not a sound business financially (high cost, inefficiency, low utilization, idle fabs, huge depreciation cost etc) is not Intel products team's concern and it gives an opportunity for Intel management to identify inefficiencies and fix that.

Now I fully acknowledge that Intel Foundry was & is a strong moat for Intel that made them king of compute for the last couple of decades. But that does not mean they were run efficiently or profitably (in an IDM structure, it was not even a concern). Even in 2021, when Intel was making record revenue of $79b and profits, Intel foundry was making an operating loss of $5b!. Their operating margin at that time was 25% for the entire business, so Intel products must have had much higher margin to offset this foundry loss.
https://www.intc.com/filings-report...0000050863-24-000068/0000050863-24-000068.pdf

So I don't believe Intel Foundry division was ever profitable on their own and being an IDM, nobody at Intel cared. Pat and Dave spoke about it many times about too many steppings, hot lots etc.

The real problem I have with Intel margin calculation there is another operating expense bucket called "Unallocated Corporate Expenses" that include Stock Based Compensation but is not included in segment operating profit calculation. I believe that is supposed to be included in COGS, R&D and SGA depending on whoever gets these stocks. So, I do think Intel product margin is overstated a little due to this. I am not a finance person to truly know if what they are doing this way is acceptable per GAAP but these results were audited and approved.


I think(imho) Lion Cove is actually not that bad. There is definitely a latency issue due to the tiles approach vs monolithic but Robert Hallock (Intel) said in recent interview its firmware/microcode issues that is overblowing this issue. Also recent patches/updates by game developers had improved gaming performance of ARL-S, for example Cyberpunk 2077 had a huge uplift that puts the 285K in line or slightly above 14900k while consume less power. So with time, some of these issues will be ironed out. I don't expect it to beat 9940x or 14900K across the benchmarks anytime.

Also with regard to process nodes, Zen 5 is on N4P (if I am not mistaken), so N3 (which N3B now) is 3 to 8% better in power and 0-4% better in performance per following graphic. And I am pretty sure, N3 is very costly too.
Considering this is one of the first time, Intel engineers are working with an external foundry to make their products and first tiles approach on desktop sort of Zen 2 for AMD (based on what I see, Zen 2 was not received well for gaming too), I think we can cut them some slack this one time.

View attachment 2605

i can't cut them some slack for their poor choices especially with such a slow L3 and regression in performance genuine issues need to be criticized

 

[Xebec]

Thanks for giving me the opportunity to do some math homework!

LOL - This math stuff is scary!

Now, let's compare Arrow Lake and Zen 5 using their top models.

The compute tile of the Core Ultra 9 285K (Arrow Lake-S) is 114.5 mm². The 8-core CCD of Zen 5 is 70.6 mm², and the 9900X/9950X models require two CCDs, totaling 141.2 mm².

141.2 / 114.5 = 1.233, meaning Zen 5 uses 23.3% more area for its top model compared to Arrow Lake.

The next question is cost, which is tricky to estimate. It's unclear which N4 node Zen 5 uses—N4, N4P, N4X, or even N4C (a cheaper version of N4P).

Thanks and always appreciate the references. It'll be interesting to see what other die sizes Arrow Lake eventually ships with on desktop (January/February "Ultra 5" and "Ultra 3" non-K models I'm assuming).

Isn't it still a little misleading though to say "Intel is using a more expensive process"? If Intel and AMD had similar die sizes and performance, but AMD was on N4 and Intel on N3(B), that means Intel's design is more expensive because they used more transistors for the equivalent performance. This is oversimplifying of course, but it seems like it's fair to say that Intel is using a "more advanced process" (with cost/area caveats), but the design is what drives the final cost/performance.

That said, I suspect AMD's Zen 5 range is cheaper to produce overall than Arrow Lake on desktop -- but because of the overall design philosophy, for all of the reasons you wrote. Smaller CPU chiplet, packaging looks less complex (no large "Base Tile", "Filler Tiles", and only 2-3 dice vs. 4 dice).