You are currently viewing SemiWiki as a guest which gives you limited access to the site. To view blog comments and experience other SemiWiki features you must be a registered member. Registration is fast, simple, and absolutely free so please, join our community today!
I wanted to make this post since all three members of the RGB Team are making GPU using N5/N4 TSMC Process.
Intel Arc B580:
Process: N5
Die Size: 272mm2 N5(TSMC says N4 is a 6% optical shrink vs N5 so it would be roughly 256mm2 N4)
Xtor Count: 19.6B
Density: 72.1 MTr/mm2
Relative Performance: 100% (1080P games)
Perf Per xtor: 5.10^10-9
The AI performance is commendable for Arc but yes their xtor density sucks vs competition if they would have been able to take it to Nvidia level density it would be pretty good but maybe next generation
Much has been said about Arc and Xe transistor density compared to AMD and Nvidia, but I never got a really good sense of if the chip size truly is so massively detrimental to product margins.
For instance, is a Xe3 Arc Celestial GPU which has, let’s say 20% better density than Xe2 Battlemage, going to hit Tan’s new mandate of 50% gross margin? Would TSMC N4 vs Intel 18A make a big difference?
Let's get details.
Perf Per xtor benchmark is not important.
What's important is Perf/area.
I have a table as following,
I assume N4p and N4E cost 5% higher than N5.(sorry, it should Perf/cost)
Let's get details.
Perf Per xtor benchmark is not important.
What's important is Perf/area.
I have a table as following,
I assume N4p and N4E cost 5% higher than N5
Let's get details.
Perf Per xtor benchmark is not important.
What's important is Perf/area.
I have a table as following,
I assume N4p and N4E cost 5% higher than N5.(sorry, it should Perf/cost)
well N5 -> N4 is a 6% shrink so the area would be around 256mm2 roughy on scaling it by that factor but i pretty much agree with you for BattleMage for cost i think they would charge around 10-15% more for N4P as it is clear 11% better in PPW and 6% denser than N5.
Much has been said about Arc and Xe transistor density compared to AMD and Nvidia, but I never got a really good sense of if the chip size truly is so massively detrimental to product margins.
For instance, is a Xe3 Arc Celestial GPU which has, let’s say 20% better density than Xe2 Battlemage, going to hit Tan’s new mandate of 50% gross margin? Would TSMC N4 vs Intel 18A make a big difference?
Intel 18AP will help foundry with volume that's it its a product issue not fab issue and a classic issue of Intel design preferring High performance as for celestial it's not out yet to give a remark and Xe2 vs Xe3 is a big architectural change tbf they may be adding more features to the arch.
well N5 -> N4 is a 6% shrink so the area would be around 256mm2 roughy on scaling it by that factor but i pretty much agree with you for BattleMage for cost i think they would charge around 10-15% more for N4P as it is clear 11% better in PPW and 6% denser than N5.
Intel 18AP will help foundry with volume that's it its a product issue not fab issue and a classic issue of Intel design preferring High performance as for celestial it's not out yet to give a remark and Xe2 vs Xe3 is a big architectural change tbf they may be adding more features to the arch.
Given Celestial has been in development for many years, and that it appears pre-Tan that Holthaus re-greenlit a bigger Battlemage product, will Tan actually apply this 50% gross margin approach to a future Celestial release and thus walk back Holthaus leaning in more heavily to Arc? I guess we won’t know until the end of this year at the earliest…
I presume rather than for consumer that all hands are on deck to get their *Shores line on the rails after whatever (again…) strategic misprioritization they had on the DCAI GPU side of the business causing them to cancel and redo the lineup…
Intel products truly has so many issues in their product strategy and execution it’s hard to understand how Tan’s 50% GM comment is going to work… that’s all good when you’re in the lead pack of a product segment, but when you’re trying to come from behind? How does that work?
Given Celestial has been in development for many years, and that it appears pre-Tan that Holthaus re-greenlit a bigger Battlemage product, will Tan actually apply this 50% gross margin approach to a future Celestial release and thus walk back Holthaus leaning in more heavily to Arc? I guess we won’t know until the end of this year at the earliest…
I presume rather than for consumer that all hands are on deck to get their *Shores line on the rails after whatever (again…) strategic misprioritization they had on the DCAI GPU side of the business causing them to cancel and redo the lineup…
Intel products truly has so many issues in their product strategy and execution it’s hard to understand how Tan’s 50% GM comment is going to work… that’s all good when you’re in the lead pack of a product segment, but when you’re trying to come from behind? How does that work?
My view is that an assessment of GPUs for AI at the chip level is very misplaced, because the real magic is happening at the systems (rack-level) and associated software level. That perspective is heavily reinforced by the key focuses of the upcoming Hot Chips 2025 conference in August. First off, the conference is 90% about AI hardware this year (very little on CPUs). The main tutorials are on rack-level hardware and GPU kernel programming.
My view is that an assessment of GPUs for AI at the chip level is very misplaced, because the real magic is happening at the systems (rack-level) and associated software level. That perspective is heavily reinforced by the key focuses of the upcoming Hot Chips 2025 conference in August. First off, the conference is 90% about AI hardware this year (very little on CPUs). The main tutorials are on rack-level hardware and GPU kernel programming.
www.techpowerup.com