This is how badly Intel screwed itself over

LazyGamer said:
If ASML machines aren't putting out, won't TSMC run up against the same issues?
Click to expand...
I think it's not just about the machines. It's more about passionate people wanting to do their best. I don't know what kind of culture Intel fabs have but I have read that at TSMC, almost everyone on the fab floor is a PhD with expertise in some area that helps TSMC be the best. Unlike Intel, they have never rested on their laurels. They take their commitment to their customers more seriously than death itself.
 
My observations suggest that AMD and Intel are following two different ways of designing chips.

AMD: Going about their business in a calm but determined manner. Iteratively refining their designs. They are focusing on smaller but deliverable IPC improvements.

Intel: They have been working hysterically in panic mode since Rocket Lake. All their engineers can think about is how to leapfrog AMD's designs in one fell swoop. This kind of reckless thinking is causing them to take bigger risks and they keep falling flat on their faces.

I would LOVE to knock on Pat's coconut and ask him, why in the hell did you think it was a good idea to:

Not have 12 P-cores only Alder Lake/Raptor Lake halo part?

Not have a monolithic version of Meteor Lake/Arrow Lake in case things didn't pan out the way they were hoping?

Have separate core designs for consumer and server chips when everyone had seen how successful AMD had been with their server first design approach and then trickling fundamentally the same design down to consumer chips since 2018?

Not have SMT enabled in their E-cores when AMD did enable that in their c cores? (I do know the answer to this. The E-core chip architect Stephen Robinson thinks SMT makes zero sense for consumers. Frickin' Nazi!)

Not instruct engineers strictly to start work on enabling AVX-512 in post Raptor Lake consumer chips when Rocket Lake had it and the competition has it?

Not release quad channel HEDT platform featuring slightly more expensive Raptor/Arrow Lake chips to get a leg up on the competition?
 
@igor_kavinski your post is interesting. I recall when the amd and Intel positions had been swapped. When AMD was always scrambling until Lisa Sue came along and took some steady state deep risks and gave the cpu teams time to mature their tech fully.
 

Page 5 - Discussion - Cinebench 2026

Page 5 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.
forums.anandtech.com forums.anandtech.com

My calculations seem to suggest:

Modest 13% IPC increase with at least 20% higher frequency boost.
At least 6.5 GHz frequency AND/OR a much more efficient SMT implementation.

This can only be Zen 6.

Intel can kiss its goose goodbye because 6.5 GHz is going to cook it REALLY good!
 
I'm honestly not sure if Intel even cares about the traditional x86 space anymore. To me, it seems like they're all in on mobile offerings now, at least when it comes to consumer parts. I've yet to have the privilege of a high-end AMD laptop but I can say that I've been happy with every Intel-based gaming laptop purchased, about a half dozen since 2010, although I did try to order some AMD ones only for them to be cancelled.
 
I spent quality time with a Ryzen 5825U HP Aero laptop. Tortured its iGPU with 4K benchmarks. The fan noise was minimal and not once did the laptop lock up. And it was so light that it felt like a cheap plastic toy laptop. The person I bought it for has yet to complain. I was very impressed, at least with that particular chip.
 
The problem with the 'desktop' space is that... most of that is ordered by the big three OEMs, and desktop CPUs are basically a commodity. As in, it doesn't even matter which one is 'better', it's which one fits their requirements and they can get the best deal on.

To that end, for corporate desktops, Intel is plenty experienced in 'checking the boxes'.

But for enthusiasts, especially gamers... they've been missing the mark, and I think they know it.
 
LazyGamer said:
The problem with the 'desktop' space is that... most of that is ordered by the big three OEMs, and desktop CPUs are basically a commodity. As in, it doesn't even matter which one is 'better', it's which one fits their requirements and they can get the best deal on.

To that end, for corporate desktops, Intel is plenty experienced in 'checking the boxes'.

But for enthusiasts, especially gamers... they've been missing the mark, and I think they know it.
Click to expand...
Right now Intel is going to the government trough to sustain itself. And that trough is filled with TAA compliance, and security requirements, and NOT gaming performance... they will sustain themselves and get funding for more production from the US government, even if we have to invade another country for the raw materials... looking at you Iceland/Greenland. (And yea I don't want to get into the politics of it... but it is what it is... disgusting or not.)
 
Yep, even with its issues, the 14900K was the end of the road for gamers using Intel but I will once again say that the 275HK in my laptop would probably do well in a desktop and reviews for the 285 haven't been horrible since all the updates rolled out. I think the 52-core beast will be impressive as well but there's still no denying AMD's dominance.

There were reasons I jumped ship when I retired my 4930K. Gen after gen I saw power/heat increase for Intel chips that while seemingly impressive didn't really justify to me all the costs involved in using. It's been a great ride since on AMD and I'm looking forward to CCDs with more cores/threads combined with 3D Cache.
 
It's literally just the cache (and/or the resulting frametime consistency).

Intel fixes that, and they're back on top.
 
Intel has APX, AVX10.2, rumored code injection in game executables through their APO app and much higher RAM speeds than AMD.

If they STILL falter with Nova Lake, gamers everywhere will be like

1767819592189.png
 
igor_kavinski said:
This is how Intel could potentially un-screw itself:
Click to expand...
They absolutely have the opportunity to drive the market.

Having the hardware available is as important as having the software available (see: Radeon, and lack thereof), but market penetration perhaps matters more than either, because that's what developers will target.

And realistically, for any 'regular' user, having more effective local agentic (or hybrid local and cloud) AI will make more of a difference than faster CPU or GPU cores. Intel has the ability to tailor their products toward end-user demand here so long as they keep their product lines evolving.
 
What's going to matter for your LLM's is the tokens and how much content you can generate with access to your resource data in x time frame.

How many tokens you need to be able to generate per minute/second for natural communications or whatever the use is of the AI.

TOPS matters but Tokens are the real payday.

License your LLM sure... I'm not even sure if LLM size equates to token generation or not. Need to learn more about that.

I have a feeling that businesses are going to over build their AI hosts and buy far more compute than they actually need to get the job done with an AI for the use cases in question.

A balancing act needs to take place as well.

LLM A, lets call it 200 gig in size running on a cluster... that can host the 200 gig LLM in a dedicated memory space. So like.. 5 of the new mac pro mini's with 512gb ram each in a cluster.

How many tokens can that generate per second of a quality needed and how many tokens per second do you need for a phone conversation? (For replacing phone agents of offsetting phone agent manpower needs.)

Now that same farm... same node count... but with an LLM that's more refined.. say only 50 gig in size, still running across all 5 nodes of your mac mini cluster. Do you get better performance for the compute, meaning is the performance to capability offset worth it for the smaller LLM size that could ostensibly process and handle more calls at the same time?

And that's just lab type stuff... you could rent some LLM hosts at AMD or Nvidia and do some testing... find out what your load/need is... build out your cluster to handle that +/- 30% of load. And run with that well... in the event you have a peak load requirement then you go to your provider to add additional compute for the burst time needed of your AI load.

Then you have to do the penny pinching balancing act to say... this is most economical to meet our need, this makes the most sense for data security of our customers and ourselves, and this allows us to have the flexibility of the cloud on demand and the security of on prem the rest of the time.

sorry for the thought exercise. These are the kind of questions we're spitballing.
 
You must log in or register to reply here.
Become a Patron!
Back
Top