Excited to share the news: Lunar Lake is coming in Q3!
>60 TOPS from GPU w/ XMX systolic arrays
>45 TOPS from new NPU
1.5X faster raster gfx gen-on-gen
2x compute in low power island
Profound changes in low power architecture
Fun summer ahead. :)
~1 hour until the PC industry snaps back to reality. Lunar Lake is the one and only chip for low power + graphics + AI + core performance. Devices shouldn't be a "choose two" situation.
Oh, also, Lunar is more than laptops. See y'all on the other side. ✌️
LLaMA3-8b running on Core Ultra? Yeah, we can do that. :) Short gif of it up and running on a Core Ultra 7 155H Arc GPU with DP4a. You can see the GPU activity.
Intel is the only CPU vendor demonstrating functional enablement on day 0 of the announcement.
Chuckling at this, as it was only 7 years ago when people were saying to my face that 8 cores in a general desktop was excessive and pointless. And now here we are: even game installers use it.
"Look at what they need to mimic a fraction of our power."
Interesting conversations about notebook graphics performance today. Just gonna scooch this one into the public record. 👇
Just carrying around a Meteor Lake in my pocket this week, nbd. Love this!
If you look reeeeeal close, you can see the seam between the IO Tile and the Intel 4 Compute Tile.
501 models, anyone? 😊 Microsoft Phi-3 up and running on
@intel
Core Ultra. Day0 support. Intel once again on the bleeding edge of new AI model enablement on the PC.
@carygolomb
There are many architectural changes we want people to understand before we get to the Q3 launch. We have more firmware and driver work to do. The world will get to review LNL later this summer, without restriction or reservation.
ICYMI:
SLM/LLM benchmarking is rapidly evolving and has unique gotchas not found in other categories of testing. What explains the differences in this data, and how can you avoid errors? Read on:
While I took a sabbatical from tech, I started a small biz that helps design high-performance car parts. Very excited to announce my first co-designed project is on sale with my first customer:
Designed for Mk5 Supra, but works on any BMW B58 Gen 1 or 2.
SLM/LLMs are a brand new field of benchmarking. There are a few novel requirements to look out for, and even seasoned benchmarkers are getting it wrong.
Have been mulling this one over for a while, and decided it's worthwhile to do a writeup 👉
We also have two special features of this LLM running:
1) Resource Augmented Generation (RAG), which can cite your docs or data for hyper-specific and detailed answers.
2) React Agents, which can tap external websites for up-to-the-minute post-training data. Demo @ MS Build!
Fiiiiinally started Starting Trek: Strange New Worlds. This is rapidly becoming one of my all-time favorite Trek series. The writing is fantastic, the continuity touches are brilliant, the characters are great. I'm mad at me for sleeping on this for so long.
Pretty amazing how quickly DDR5 prices dropped with scale. Was perusing some D5 kits today, and you can get super solid 32GB kits for under $100. Another W for the PC industry.
Made good use of the long weekend.
1. New turbo installed; spools like stock but flows ~750WHP.
2. Built trans swapped, rated for 1200WHP/1000WTQ/190ms shift time.
3. Walbro 525 auxiliary fuel pump installed.
4. CAN bus expander to integrate pump + port injection.
@Tobias_Writes
You're right. Any CPU can execute an LLM, and it's a low bar to clear. I run them with Jan on my TGL laptop. But perf/W is another story, especially as the models get more complex and/or become background resident in computing. Moving to accelerators is a must & harder. 🖖
@IanCutress
@draginol
"wow it's so cool you've been to all these places for work!"
Me: "yeah, the hotel bar and I'm very familiar with Westin Taipei meeting rooms."
Codex Entry: The Noodle
Subject: The Last Lasagna in Hell
Amidst the scorched plains of the Netherrealm, where the air crackles with the wails of the damned echo into eternity, there existed an artifact of unexpected power: the Last Lasagna.
AI on the PC is a lot like gaming.
AI models are like game engines
Gfx and AI both have APIs and LL-APIs
AI apps are like games, w/ specific optimization
NPU has monthly drivers, just like GPU
The code hits multiple engines
Performance and reliability is all in the SW. 👇👇
We are putting resources into Intel now. multiGPU training worked out of the box, and haven't seen a lick of driver instability. The A770 card is currently slow though, we need to add XMX support ($400 bounty added)
@sidkshatriya
Lunar Lake will also support Copilot+! Surface Pro 10 has Meteor Lake, too. But Lunar launches in Q3, so you can see it doesn't line up with a Surface in May/June. :) Just timing.
Hope all the attendees had a great time at
#IntelTechTour
. There is a LOT to learn (and I'm not long ahead of y'all 😂). Hope to see many of you myself at
#IntelInnovation
in a few weeks!
@JamesDSneed
Too early to give an exact number in hrs; systems and drivers still going through final designs. But I said the word "power" about 9mil times this week and there's a very good reason. We're aiming at the top.
@DAW00KIE
I can't answer this question, yet, but I wanted to reply so you know I saw this question. We'll talk form factors over the summer. You have my word.
My Pixel 6 Pro often sends incoming calls straight to VM without ever ringing. Others report they can't hear me talking, but I hear them loud and clear. There may be some correlation with VoWIFI, but I can't quite tell.
Anyone have this? Is this a Google issue, a TMobile issue?
@terafoenix
Because about 40% of the software developers we've polled want to do AI on GPU, 30% on NPU, and 30% on AVX256. There's both because that is what the software industry is expecting. The NPU alone covers < ⅓ of SW dev plans.
LPCAMM is super cool, and system verification was completed on an Intel platform!
LPCAMM is notebook LPDDR memory on a removable "stick." 32-128GB per module, up to LPDDR5X-7500 at present. The
@anandtech
crew have a great writeup:
Just wrapped season 1 of Strange New Worlds. It's easily one of the best seasons of Star Trek I have ever watched. Wow wow wow. I'm mad I slept on this for so long.
Unreal Tournament (1999), With all of the Quake remasters available, do you think we should see some Unreal remasters too? Should we start with one of the best competetive FPS games of all time?
#firstpersonshooter
#videogames
#fpsgamer
#gamer
#fps
It's the world's first upgradeable a2w manifold and port injection solution for B58. Can support over 3000cc of port injection across 12 injectors, no injectors, half injectors, and it's designed so users can do fuel first and cooling later if they want. The value is insane.
What's the game soundtrack that instantly transports you back to your childhood? For me, it's the EverQuest soundtrack. I've been PC gaming since I was four y/o, but NOTHING grabbed my attention like EQ did. It changed my life. Some tracks bring back years of memories.
@superbphoenix
@IanCutress
It is fine. This is what modern process technologies support for long-term use. It's not 2006 with CPUs that explode at 65C anymore.
Reintroducing Quake II, the legendary sci-fi action FPS. This enhanced version features new content, online multiplayer/co-op with cross-play, optional deathmatch bots, visual upgrades and now up to 120hz in 4K!
Available now on all platforms and
@XboxGamePass
!
@Krewell
@Grady_Booch
Because they're afraid someone might actually challenge their weak and fragile worldview, which entirely depends on being cruel and uncivilized, or harnessing marks of that nature as profit centers.
@sebastianpeak
It very well might be. Knock on wood, but I killed VoWiFi and things have been going okay. My previous phone was a OnePlus 5, and I'm recalling that VoWiFi was ultra sketch on that, too.
There's a certain magenta that is common, here.
Understand the legalities, but on macro: thie spirit of this effort is what I'm really excited about in AI land. NPC AI, voiced dialog, procedural quests, flavor and quest text, music generation, etc. There's an exciting step function in immersion and replayability looming.
The creator of an eye-catching Grand Theft Auto 5 mod that promised a living, AI-powered story mode has admitted defeat after Take-Two hauled the mod offline.