Why I Like Talking Chips: A Hot Chips 34 Recap
ฝัง
- เผยแพร่เมื่อ 25 มิ.ย. 2024
- The long and short of the chip industry, as per the recent IEEE conference Hot Chips 34. If you missed the event, here are the key highlights.
[00:00] Intro
[00:26] Sponsor: linode.com/ttp
[00:57] Tesla Keynote Recap
[03:14] Intel Keynote Recap
[04:54] Lightmatter Passage
[06:48] Biren BR100
[08:17] NODAR Stereo Vision
[10:42] Intel Meteor Lake
[12:09] Cerebras WSE uArch
[12:52] NVIDIA Grace
[13:42] Intel Ponte Vecchio
[14:09] CXL (Compute Express Link)
[14:28] Arm Morello
[15:20] AMD's Disappointing Showing
[16:31] Other good talks
[16:46] Summary
[17:24] Hot Chips 35 ?
[18:52] Cat Tax
-----------------------
more-moore.com : Sign up to the More Than Moore Newsletter
/ techtechpotato : Patreon gets you access to the TTP Discord server!
Follow Ian on Twitter at / iancutress
Follow TechTechPotato on Twitter at / techtechpotato
If you're in the market for something from Amazon, please use the following links. TTP may receive a commission if you purchase anything through these links.
Amazon USA : geni.us/AmazonUS-TTP
Amazon UK : geni.us/AmazonUK-TTP
Amazon CAN : geni.us/AmazonCAN-TTP
Amazon GER : geni.us/AmazonDE-TTP
Amazon Other : geni.us/TTPAmazonOther
-----------------------
Welcome to the TechTechPotato (c) Dr. Ian Cutress
Ramblings about things related to Technology from an analyst for More Than Moore
#techtechpotato #hotchips #semiconductor
------------
More Than Moore, as with other research and analyst firms, provides or has provided paid research, analysis, advising, or consulting to many high-tech companies in the industry, which may include advertising on TTP. The companies that fall under this banner include Armari, Facebook, IBM, Linode, MediaTek, NordPass, Qualcomm. - วิทยาศาสตร์และเทคโนโลยี
Amazing summary 👍
The 10 hour streams were a bit much for me to watch. I would love to see several other videos highlighting the best of the event. Lightmatter seemed most interesting to me, CXL was second. But any video you put out would be awesome, whichever topics interest you the most. Thanks for the coverage.
Definitely lightmatter, it would be amazing if we could go back to analog computers in the new style.
Dr Cutress, I can't think of anyone I'd rather have explain things I can't understand. The video on Piccachus per bit transfer sounds like a winner
Yes for the deep dive videos. Looking forward to nuvia cores presentations next year!
Second that re nuvia.
Thanks for covering Hot Chips Ian! Your summary is a huge value add for those of us that don’t have the time to watch the entire event.
I suspect the days of Intel, AMD, and nVidia disclosing anything interesting at these events are over. Their audience is so big now that they’ll reserve anything ‘interesting’ for their own hosted events like what AMD is doing this week. They can also control the messaging better in their own events by setting their own group rules (for example no Q&A). Hot Chips is sadly just a marketing event for them these days…
To be fair, they said the same thing 5, 10, 15 years ago, and every so often it happens that Hot Chips gets a good disclosure.
I would think is more about not give any information to competitors than anything else, they are close in race and any info can give competitors better plan to counter they solution. About AMD I looking for today event, but it would be only Zen4.
YES, absolutely! Please go more in Detail of the DOJO😸(I had to type it for my cat)
Would love a Dojo in depth video 👍
Thank you for not saying 'this begs the question,' as it's my biggest pet peeve.
Very much to hear more details on the Dojo implementation.
Thumbs up for the cat tax at the end!
3:03 yes more details please
6:45 PLEASE PLEASE PLEASE do a vid on the silicon photonics
14:50 Very intrigued about ARM Morello
> a vid on the silicon photonics
Asianometry had one.
@@erkinalp yep i saw that, but it might be more interesting to see Dr Cuttress cover it from his viewpoint using his technical experience
Thank you Dr. Ian Cutress for this useful recap.
btw .:
V.nice flow and presentation of information.
Lots of brands mentioned back to back. Superimposing a table with brand logo on top and product class down the side would make more sense to my brain by hearing the speech and seeing something graphical for each list of variations of things mentioned.
A video detailing the Tesla Dojo Architecture and Microarchitecture is greatly appreciated.
Yes please, do a full video on that thing. also leave room for people who are less well versed in that sort of data center tech related advancements, but want you to out into perspective, the kind of performance increase we are looking at with thaT lightmatter interposer thing.
I really wanted to watch hotchips, but my new job takes most of my time, thanks for this video sir. Really blessed by your info and coverage. maybe if God wills it I'll work in your industry due to the passion I've developed after learning so much from people like you. much love!
Tesla Dojo Deep Dive Video - Yes Pls. 👍
We definitely need full video about DOJO
Please do separate videos on each of the sub-sections you rightly suspected needed a separate video. I am very interested to know what you have heard is coming in computing both from interconnect, to memory, to transistors & materials advances.
With regards to silicon photonics I'm interested in particular in what kind of impact the difference in propagation delays between it and traditional interconnects can influence the uses for the technology. As I work in the service provider networking space I know there are also some interesting challenges with long distance optical connections with regards to chromatic dispersion and polarizing modal dispersion, and would be interested to learn if they anticipate needing to address any of that for optical compute interconnects.
definitely need a dedicated video for Dojo!
Definitely interested to learn more from you about the dojo v1 and the lightmatter silicon photonics.
I'd love to see a deep dive each of various subjects in this video.
I'd really like to see a full video on the lightmatter photonics stuff, photonics is always fascinating because it sidesteps so many problems if they can get it right.
I watched a few bits of the livestream and appreciated your insights. I would love it if you could somehow get more information on MTL with regards to CXL because it really sounds like they were using it within the chip itself.
Yup just within the chip, I also want to know more
HotChips is always a great snack every year. And look forward to the video about dojo (ttp inside!)
While the number of foundries are/have been decreasing, the number of chip makers/designer seems to be increasing.
I personally am very curious about the analog chips designs for Neural network/machine learning that has previously been announced, perhaps making a video of it would be interesting if possible.
@3:02 - leave comment if you want more detail on Tesla supercomputer system
thanks for the nice summary on DOJO D1. It looks like Tesla is going with a data streaming approach similar to cerebras wafersize chip. Makes sense given Tesla wants to train on video stream, instead of individual images from each camera separately. It would be nice to compare Cerebras to D1
I would like to see consumer submersion systems! With high end PCs tipping over 1000w they would offer a great way to cool them, look amazing plus be super quiet.
I would love for you to do a video on Submer Ian.
Great summary. Thanks. Would love to find out which silicon photonics company AMD is going to buy to integrate their solution into their architecture 👍
I'd love to hear more about Tesla's Dojo architecture.
wow, what a compendium!
For the Dojo specific video maybe you can do one after their AI day presentation since they said they'd share more detail then
I'm crossing my fingers for an invite to attend. Will see
@@TechTechPotato Maybe try contact youtubers which cover Tesla if they could get you contact to someone who have power to give you invite. Because seats for wallstreet analysts are waste of seats ;)
More detail on dojo would be awesome.
Would love any and all deep dive videos as 9 hours livestream i definitely wont be able to or want to attend lol.
Did not go to hot chips and barely know how to use my computer lol let alone such level of knowledge events
I can't get enough silicon photonics, please make that video. 🙂
Do you actually believe that Tesla as any of the things they are talking about, or was their presentation just their wish list made to sound like they’d already achieved it?
Haven't made it to a Hot Chips in person since about HC5 or HC6. I miss hearing that level if detail w/o the marketing hype. I guess IBM is in mid-cycle since Power10 and Telum are hitting the market, and the next generation isn't ready for talks yet...
Dojo was cool, but Lightmatter's talk was my favourite, it made me feel dumb in the best way possible
Pet peeve of mine: Hyperthreading is Intel specific branding for its SMT. So please don't use the term with other architectures.
That's fair, it bothers me as well but for whatever reason it didn't trigger in my mind this time.
@@TechTechPotato it bugs me too and I still say it all the time. Silly human brain lol
long live the original UK inmos, transputer SMT & Ferranti sram/dram/silicon substrate etc, etc 🤣
-transputer gfx chip- svga & all the later IBM non integer derivative's they licenced
Maybe google could talk about their dedicated transcoding hardware for TH-cam.
Ian, I was wondering... you mentioned a conference that Cool Chips? Have you ever covered that?
Not properly. It's not promoted as much, and every other year it's held in Japan. I'd love to go, but not many of my usual companies that I follow are there
@@TechTechPotato
I can understand that, but I will be honest it seems to be that... from the point of the consumer world, the low-end of this is a bit more interesting.
As much as HPC is something to keep an eye on, I feel that the delta between work done and energy consumed has a rather big place in the world these days. Not sure if Cool Chips touches on that at all.
I'm hoping for Intel to really show off some silicon photonics stuff in processor packaging. Intel's 'secret' maneuver was to get a cross license for IBM's photonics technology as part of their recent agreement. I really do expect Intel's next big server platform post Granite Rapids to adopt photonics for processor-to-processor interconnect and scale up stupidly high.
Next year I suspect we'll be hearing about IBM's next Power chip which if I had to bet will formally adopt CXL now that CAPI is dead.
Wild card for next year will be new FPGAs from both AMD and Intel as well as any plans to integrate them into main processor packaging.
Similarly I would expect AMD, ARM and Intel to all be discussing aspects of their smaller core designs destined for the server market.
17:40 Nah i just went to my local takeaway for Hot Chips 😜
Pet peeve of mine: when ppl say megatransfers per second instead of effective memory speed 😋
I agree re tiles vs chiplets. They arent the same thing. And calling the same thing is (intentionally?) misleading.
When will I be able to map out a chip on my pc and order a half a dozen prototypes the way I can with pcbs? Surely that's on the horizon. Overall I felt without ibm and with Intel present in body only hot chips was a bit... just not up my alley I guess. I don't want to call it a "dud" cause that's not really true, and certainly not fair.
I would be interested in hearing more about silicon photonics. I remember reading something from jds uniphase about that 2 or 3 decades ago.
should you do a video on it: yes
all of it :)
Did Tesla solve their compiler problem with the DOJO tile architecture?
To respond, I actually would like to hear your take on DOJO ...
The photonics thing sounds like an interesting topic. What I am wondering is what needs to happen for it to make sense. Should we expect future motherboards to replace coper traces with fiber? xD It sounds crazy!
On a smaller note, wondering what your definition of chiplet is then... AMD has core complexes, I/O chiplets - puts it on a substrate, and there is your processor.
If Intel is doing the same with making cores, I/O, GPU chiplets... seems the same to me.
The only notable difference is that intel aims to have several versions of those dies - where AMD bins and disables broken cores but fundamentally uses a single 8 core design.
Yeah okay I see your point...
Thinking about it some more, I dont understand this intel strategy, to be honest...
AMDs single design chiplet for the cores allow it flexibility in what they produce and there is less double work. Intel has to have "one production line" per chip size. Thats 3x the work, compared to AMD sticking to one 8 core chiplet. At this point I am wondering if intel is going to disable cores on bigger chiplets to fit them into lower skus... And the only benefit I see if less latancy between the cores. AMD just idles its 2nd chiplet most of the time to keep the work together.
Forget tesla idk how AI works but I do care about lightmatters key note please do that instead.
What do you have on the notebook?
Thanks
I think it's a rode microphone
@@christopherjackson2157 you are right, thank you
Tesla Dojo is great and etc, but could you compare it to other systems and alternatives, because looking at Elon Musk comments they are still not outperforming Nvidia chips for they needs. Or it is only matter of time for software to mature.
Yeah it's still in strong development, but at least we're getting to peer behind the curtain for now
@@TechTechPotato yeah, hope they give more about performance on Tesla AI Day 2, and show any progress in mapping networks to chips, because it was question from public at first AI day which make me worried most about.
BTW from moon shots, do you think if Tesla use photonics for transport layer (between chips/tiles) would be good idea?
PS I would love compilation about what we can do (prove by working concept chips) in photonics, because this is looking more and more promising, but hard for people not from this segment to know what are today limitation, because I read about logic operation are not currently possible, but next I saw presentation where they use interference to do so... summary would be great, with how they are controlled. Because one thing light could allow is true 3D processing, CPUs in shape of cubes with interface on each side (20-40years in future?).
Get Tesla Dojo video done please
Cool chips, but can they run crysis?
It should be fish and chips 🍟 being English
It seems like Intel 3 and Intel 18A are very minor improvements over Intel 4 and Intel 20A, respectively. With the use-case being IFS. That's just my thought since Meteor Lake to Arrow Lake jumps from Intel 4 to Intel 20A, skipping Intel 3. It me wonder if Intel does plan to keep its internal and IFS nodes separate. Also, I don't recall Intel 16 ever being used for anything in the past. It seems like a minor modification of 22FF for the purpose of IFS again.
most tanned man on earth delivers flawless sum up of elitist marketing campaign disguised as a conference
I would love a deeper dive on LightMatter Passage.
fyi the 'sponsor' chapter is too long
Did Jensen Huang not say? "The more you buy, the more you save!" Speaking about NVIDIA GPUs.
We all saw how that became true with crypto mining lol.
Go in-depth more please.
More of what?
yes.
You’re making the same mistake I did in my first term studying Physiology in thinking binocular vision is needed for depth perception when I said people with only one eye shouldn’t drive. Stereoscopic vision, as you said is very limited in depth, beyond this your brain uses motion, to assess depth. If you don’t believe me close one eye and wiggle your head from side to side like you have a medical problem. You still retain depth perception, no? That’s the skill they should replicate with their “hot chips”. At low speed they should just wobble their camera from side to side but at higher speeds everything is moving anyway.
Does Tesla plan to sell time or hardware? If not, is this a show and tell to attract engineers?
AMD mobile 6000 CPUs have a Microsoft security device onboard... nice to know.... o.O
Intel's tile/chiplet approach is looking half cooked. Unlike AMD, Intel's core is still monolithic in MTL and ARL. They just carved out the I/O & Graphic pieces to separate tiles. They should be looking to carve out the core itself (P & E) into tiles like AMD's CCX/CCD. Sapphire Rapids is taking a different approach with EMIB.
It may seem that way at first glance, but Intel does have a useful benefit here with Foveros: each piece of the completed CPU tile can be fabbed independently. This should result in better overall yields since defect rates will be lower for each component.
It always bears keeping in mind that the reason we have "F" SKUs in the Intel lineup these days is because of defect rates. The IGPU is present on every "F" CPU, it just doesn't work well enough to be saleable and is disabled and sold as a cheaper F model. They introduced the "F" in the 9th generation so that they wouldn't throw away an otherwise completely functioning 9900K. I'm sure they'd like to get back to where they were five years ago where every CPU that left the factory had working graphics.
For a while I thought MTL was the manifestation of Intel’s 2017 “‘Mix and Match’ heterogeneous design” concept, i.e. evict the System Agent and Graphics from the main die and join them back together using EMIB. But after some reading, I’ve realized that MTL is not that simple - in fact that concept aligns more with AWS Graviton 3.
1. Since Sandy Bridge, Intel GT has been tightly integrated with IA. What we think of as L3 in an Intel client processor actually serves double duty as GT last level cache, and is more accurately described as System Level Cache. Evicting the GT is easier said than done.
2. MTL’s Media Engine and Display Engine - both typically tied to the GPU - actually exist as part of tSoC, not tGPU.
3. (rumor / speculation) there are some whispers that MTL will introduce a third core type as part of tSoC. Leaked spec sheet published by Igor contains references to LP E-cores. Architectural details are scarce, these can be simply Next-Mont reimplemented on TSMC N6, or a new design. But - if true - it is likely that MTL based systems can shut tIA down entirely, conserving power when the compute load is low.
SPR is really more of a “this design couldn’t fit inside the reticle limit, so we cut it up and welded it back together” situation. MTL fundamentally re-examines how a SoC is put together.
MTL and ARL is certainly welcome and progressing in the right direction compared to their monolithic designs. But if Intel really wants to engineer a scalable (client to enterprise) and competitive solution, they must come up with a way to disaggregate their cores (P & E) into multiple tiles and make them work in tandem via something like AMD IF. Combine that with the process node and packaging advancements along with EUV/High NA EUV embrace, they could achieve Pat G's vision of unquestioned leadership by 2025. They just need a lot of stars to align for them.
I despise Tesla too much to consume any news about them. Even if from a silicon perspective, it might be interesting, I had to skip the segment.