News Intel 4Q25 Earnings

[mikegg]

Big dawg those have DRAM requirements in 10s of gigabytes.
Get real.

Kimi K2 is 1 trillion params.

Get real. Stop living under a rock. This guy still thinks local LLMs are 7b. 😅

https://twitter.com/x/status/2001817749744476256

 

[poke01]

Kimi K2 is 1 trillion params.

Get real. Stop living under a rock.

https://twitter.com/x/status/2001817749744476256

cool but you know that this is out of topic.

 

[adroc_thurston]

Kimi K2 is 1 trillion params.

Get real. Stop living under a rock. This guy still thinks local LLMs are 7b. 😅

this wank is in no way related to normal PCs with 16G of DRAM.

 

[mikegg]

this wank is in no way related to normal PCs with 16G of DRAM.

Hence, I said Intel is missing the boat on unified memory with high capacity and high bandwidth.

It was you who started by saying local llms = 7b models.

 

[mikegg]

cool but you know that this is out of topic.

No it isn't. Trace the replies back. You'll find that it is on topic.

 

[itsmydamnation]

Hence, I said Intel is missing the boat on unified memory with high capacity and high bandwidth.

It was you who started by saying local llms = 7b models.

really......

like do people choose to just be contrary because thewy have nothing else of value to add.

i am microslop , i can develop a local LLM to do X/Y/Z .

i am picking my hardware target for mass adoption. am i
1. going to pick a 7/14/etc B 8bit model
2 going to pick a 1T *bit model


sigh

 

[DrMrLordX]

https://twitter.com/x/status/2001817749744476256

How expensive is a setup like that? Could you realistically get a server board + CPU + 2TB RAM and reach the same performance levels in the same budget? Would be interesting to see a comparison (since this is an Intel earnings thread, let's say you try it with Granite Rapids).

 

[511]

How expensive is a setup like that? Could you realistically get a server board + CPU + 2TB RAM and reach the same performance levels in the same budget? Would be interesting to see a comparison (since this is an Intel earnings thread, let's say you try it with Granite Rapids).

Entirely depends on ram cost you can get a 128C GNR CPU for like $6K USD the RAM is the only pain point

 

[DrMrLordX]

Entirely depends on ram cost you can get a 128C GNR CPU for like $6K USD the RAM is the only pain point

Well you're paying for RAM whether you buy the Mac Minis or the server board.

 

[511]

Well you're paying for RAM whether you buy the Mac Minis or the server board.

well than 1 maxed out ultra is $14099 with 512GB RAM and 32 TB Storage

https://www.apple.com/shop/buy-mac/mac-studio/apple-m3-ultra-with-28-core-cpu-60-core-gpu-32-core-neural-engine-96gb-memory-1tb

Here is a GNR Server with similarish configuration
https://store.supermicro.com/us_en/configuration/view/?cid=1000429554&5554 fwiw RAM Price was more than half of server price

 

[mikegg]

really......

like do people choose to just be contrary because thewy have nothing else of value to add.

i am microslop , i can develop a local LLM to do X/Y/Z .

i am picking my hardware target for mass adoption. am i
1. going to pick a 7/14/etc B 8bit model
2 going to pick a 1T *bit model


sigh

There exists models in the range of 64GB - 100GB large that machines like Strix Halo and M4 Max are great for. 1T is an example of what's possible now locally using consumer/prosumer hardware.

 

[mikegg]

well than 1 maxed out ultra is $14099 with 512GB RAM and 32 TB Storage

https://www.apple.com/shop/buy-mac/mac-studio/apple-m3-ultra-with-28-core-cpu-60-core-gpu-32-core-neural-engine-96gb-memory-1tb

Here is a GNR Server with similarish configuration
https://store.supermicro.com/us_en/configuration/view/?cid=1000429554&5554 fwiw RAM Price was more than half of server price
View attachment 137323

Take out the 32TB SSD. That adds $5k to the final cost. Just use an external drive you need more storage. Much cheaper.

The most important thing here is the 512GB of unified memory.

So we're looking at $9.5k for a 512GB Mac Studio. You only need 1 Mac Studio to run Kimi K2 1T if you use the Q3 version.

 

[511]

Take out the 32TB SSD. That adds $5k to the final cost. Just use an external drive you need more storage. Much cheaper.

The most important thing here is the 512GB of unified memory.

So we're looking at $9.5k for a 512GB Mac Studio. You only need 1 Mac Studio to run Kimi K2 1T if you use the Q3 version.

Uhh the CPU can do that as well? also if you only want 512 GB Ram than the cost comes way down with storage

 

[mikegg]

Uhh the CPU can do that as well? also if you only want 512 GB Ram than the cost comes way down with storage

A CPU can’t have as much TFLOPs relative to die size.

Try running a large model like DeepSeek on Epyc or Xeon. It’s like watching paint dry.

 

[511]

A CPU can’t have as much TFLOPs relative to die size.

AMX is purpose built for AI it's literally Matmul for CPU

Try running a large model like DeepSeek on Epyc or Xeon. It’s like watching paint dry.

Like i said use a Xeon6 with AMX

Cost Effective Deployment of DeepSeek R1 with Intel® Xeon® 6 CPU on SGLang | LMSYS Org

<p>The impressive performance of DeepSeek R1 marked a rise of giant Mixture of Experts (MoE) models in Large Language Models (LLM). However, its massive mode...

lmsys.org

than we have this

 

[mikegg]

AMX is purpose built for AI it's literally Matmul for CPU

Like i said use a Xeon6 with AMX

Cost Effective Deployment of DeepSeek R1 with Intel® Xeon® 6 CPU on SGLang | LMSYS Org

<p>The impressive performance of DeepSeek R1 marked a rise of giant Mixture of Experts (MoE) models in Large Language Models (LLM). However, its massive mode...

lmsys.org

than we have this

A CPU will never have the combination of TFLOPs and memory bandwidth that can beat a GPU in inference efficiency.

It's ok if a CPU is all you have.

But for local LLMs, we already have something much faster and more cost effective than CPUs: Apple Silicon and Strix Halo.

 

[DrMrLordX]

well than 1 maxed out ultra is $14099 with 512GB RAM and 32 TB Storage

So the Mac solution in this case winds up costing $10k more. Not sure how the performance stacks up, but if what you really need is 2TB of RAM for your local LLM . . .

 

[NTMBK]

To get this thread back on track, Charlie has a write up of his thoughts:

...So Intel is trapped. They underinvested for a decade and are now literally caught up by it. Capacity is finite and those limits have been reached, expansion is years away if started now and the starts that were in progress were scaled back. Were they scaled back correctly? Possibly, this depends on your views about spending like a drunken sailor, had that been done five years ago, what would the payout be now? And how would a potential foundry customer view this?

So all of the pieces come down to the board and their lack of competence or worse. The multi-billion dollar skeletons in the closet that were papered over, buried, and never even acknowledged publicly meant no one was ever held accountable. The rot continued and money wasn’t spent on things it should have been. Every FPGA propping up Ericsson et al for 5G base stations was a brick that wasn’t put in a new fab, and so on. As things stand now, Intel has a new CEO in the hot seat but the problem remains on high as the company suffers. Don’t look for real solutions any time soon, just more denials and lack of accountability.S|A

Intel Is In A Serious Bind With Few Options

Last week’s Intel Q4/2025 analyst call was a disaster that has a single root cause but no solution.

www.semiaccurate.com

 

[eek2121]

I feel like we need a thread dedicated to local AI. The space is a lot more interesting than most here realize. 128gb RAM and a 5090 can do some interesting stuff. Even my measly 64gb/4090 can run some playing models (definitely ones with more than 7 billion params)

There are still a lot of innovations waiting to happen in the local model space as well. One example of a recent one is clawed/molt bot using a sqlite3 database of markdown files to keep long term memory of conversations. (I am not a fan of that project due to security reasons, however that feature stood out)