Jonathan Ross
@JonathanRoss321
Followers
14,186
Following
99
Media
70
Statuses
513
Explore trending content on Musk Viewer
ドカ食い
• 99091 Tweets
自動車税
• 46482 Tweets
アイスの日
• 46089 Tweets
Saka
• 42633 Tweets
作者の性別
• 32399 Tweets
#素のまんま
• 23870 Tweets
Rodri
• 22419 Tweets
クソマロ
• 19586 Tweets
Foden
• 15792 Tweets
マシュマロ
• 14503 Tweets
ROCK 'N' ROLL
• 14090 Tweets
Mainoo
• 13213 Tweets
Saliba
• 12197 Tweets
LOREAL PARIS X ML
• 11657 Tweets
ブレマイ
• 11326 Tweets
Isak
• 10645 Tweets
Last Seen Profiles
Pinned Tweet
😉
10
4
83
Nvidia hit 100k developers in 7 years. Our goal was to hit 100k developers in 7 weeks. It's been 6 weeks, and...
64
62
1K
What do
@GroqInc
's LPUs cost? So much curiosity!
We're very comfortable with this pricing and performance - and no, the chips/cards don't cost anywhere near $20,000 😂
#Groqspeed
34
41
398
134
118
270
20
12
254
More LPU™ systems coming in hot! We hear you world you want that ultra-low latency speed. The team is building as fast as they can. We're increasing capacity and speed daily. More announcements coming next week, stay tuned!
42
40
486
10
8
165
We heard the demand for increased rate limits! This partnership adds another 21,600 LPUs this year to GroqCloud, with the option to add an additional 108,000 LPUs next year. For comparison. This will 8x the capacity this year vs. what's available on GroqCloud today, which will…
We're excited to partner with
@EWPESG
to build an
#AI
Compute Center for Europe in Norway. This puts us on track to deliver 50% of the world's
#inference
compute capacity via GroqCloud™, running on our AI infrastructure - the LPU™ Inference Engine. Read more:…
3
10
92
12
13
140
Looking forward to
@nvidia
GTC next week! Apparently there are numerous exciting and significant last minute changes to the event aimed at answering how Nvidia will respond to concerns raised by how popular some inference focused AI chip startup has recently become. 👀
What…
14
6
138
😉
A new partnership between
@aramco
and
@GroqInc
to build GroqCloud inference capabilities together announced at
@LEAPandInnovate
22
39
344
10
12
121
We broke
@ArtificialAnlys
's graph for Llama3 8B.
Sometimes we get a lucky shot on the benchmarks at the high end of our variance, but this is real, we have some we've logged that are even a liiiiiittle bit faster. Who knows, maybe someday this could be our mean perf 🤔
Maybe.
9
9
107
Wow on the press recently, thanks (
@Gizmodo
,
@stratechery
,
@SemiAnalysis_
) for the coverage of
@GroqInc
. 🙏 We're 2 months into providing early access to our LPU™ systems. Since recent publishings we've pushed a software update last night that gets more than 2x the throughput…
10
17
106
Thank you for the shoutout of
@GroqInc
on
@theallinpod
!
@friedberg
w.r.t. getting everything right all at once, when asked what our secret sauce is, we say, "No, we have 11 herbs and spices, and it takes every one of them to do what we do."
As for luck, here's a small excerpt…
E167: $NVDA smashes earnings (again), $GOOG's AI disaster,
@GroqInc
's LPU breakthrough & more
(0:00) bestie intros: banana boat!
(2:34) nvidia's terminal value and bull/bear cases in the context of the history of the internet
(27:26) groq's big week, training vs. inference,…
152
105
864
7
4
97
@notfncladvice
You're absolutely right. The reason we got to 100,000 developers >50x faster was we made it almost effortless to build AI applications. No low level C++ code, no CUDA kernels, no need to understand computer architecture.
We also deployed all the HW and automated the dev ops. 😂
3
3
96
Thank you for the shout out
@AndrewYNg
- at
@GroqInc
we're committed both to driving the cost of compute to zero, and the speed of compute to ∞
If you haven't tried ludicrous speed yet, try it at
And if you find yourself in Mountain View, you're invited…
Much has been said about many companies’ desire for more compute (as well as data) to train larger foundation models. I think it’s under-appreciated that we have nowhere near enough compute available for inference on foundation models as well.
Years ago, when I was leading teams…
43
248
1K
6
11
97
@adamscochran
Nvidia announced up to 50 T/s per GPU using 4-bit numerics with the B200. So 30x faster than what, an Arduino?
- we're still a liiiiiitle bit faster. Just a little 😉
3
6
64
There are at least five levels of AI App Autonomy:
1. Ideate: draft, template, etc.
2. Inform: retrieve information reliably
3. Verify: check our work
4. Update: make requested changes
5. Decide: make decisions
What level does your app target?
5
6
63
We thought the universe needed to bend a bit more in both space and time. This will do it.
Groq news!
Happy to announce that
@DefinitiveIO
has been acquired by
@GroqInc
so that we can accelerate Groq’s cloud offering. Was lucky to be the seed and A for both of these companies.
@sundeep
will now be GM of the Groq Cloud and work closely with
@JonathanRoss321
to…
76
50
623
5
8
62
@rowancheung
Hey
@rowancheung
, another competitive difference is responsiveness. LLMs run faster on Groq®'s LPU™ chips than any other hardware, so if you want better answers fast let
@elonmusk
know that you want
@xai
to run at
#GroqSpeed
. That, or you can wait, and wait, and wait for them to…
26
25
36
Did you notice our URL is ? 😉
7
6
51
@IanCutress
@nvidia
@groq
I've met Jensen before, and he's had his team updating GTC specifically in response to Groq this week, so not knowing enough about Groq seems unlikely. That said, ***
@GroqInc
runs 70B parameter models faster than
@nvidia
runs 7B parameter models. *** Try it:…
2
9
52
Excited to launch Groq's first consumer product!
10
2
45
@lulumeservey
Lulu, have you been reading our internal comms plans? 😂
Credit goes to
@lifebypixels
our head of Brand and
@andycunningham4
our fractional CMO, and the whole
@GroqInc
team.
2
0
42
@kami_ayani
@adamscochran
Yes, we're do our arithmetic in FP16 and store weights in FP8 on now. That said, it's a 70B and 8x7B on that site. We do have a (not released) 180B parameter model running at 200T/s with FP16 multiplies, and the performance of an MoE tends to be similar to…
2
6
40
When you start a startup, someday you hope to get popular. How popular? Well, having a fake crypto currency named after you is a good sign you're popular.
There is no crypto currency related to Groq.
20
15
32
Cute kid. DM me, we'll get you access, and you can just include him in whatever demo you do, that'll be enough 😂
4
1
34
Hope this isn't on account of us...
How far they run…
How fast they fall.
And no, don’t panic. $NVDA will be absolutely positively 👌
$SMCI looks even worse. 👀
6
1
19
4
1
33
@Ahmad_Al_Dahle
We credit Llama3 with accelerating it a week faster than we projected.
Thank you
@AIatMeta
🦙🦙🦙
1
0
32
Deepgram knocked it out of the park! Enabling real-time voice to text from 12,000km away. Part of any serious voice solution.
Huge shoutout to Jonathan Ross at
@GroqInc
for this incredible LUI demo on
@CNN
! We're honored to be a part of your stack & we can't wait to see where this astounding technology will go next 🚀
Thanks again
@JonathanRoss321
for the mention 😄
Demo:
2
6
36
2
5
28
@PatrickMoorhead
@AMD
@nvidia
@Signal_65
Looks like Nvidia went big on input tokens and small on output tokens because the MI300X has better memory bandwidth, and will look better than Nvidia on output. I wouldn't call these "typical" settings, typically the output is longer.
#BenchmarkShenanigans
2
2
28
A little over 24 hours left until my talk at
@WorldGovSummit
- first time speaking to a Prime Minister, let alone multiple, let alone from a stage.
As usual,
@GroqInc
has something new exciting to show. Wish us luck!
😁👍
#GroqSpeed
7
2
28
Thrilled to talk with
@SavIsSavvy
and
@LisaMartinTV
for
#SuperCloud5
. Great conversation about why whiteboards were banned at
@GroqInc
,
#AI
safety, and how
#GenAI
may increase empathy in the world.
Oh, also, why the
#Groq
™ brand is "wow!"
The llama is back! 🦙
#Supercloud5
rock stars.
A curious wardrobe change for
#theCUBE
's Savannah Peterson as Jonathan Ross, CEO of
@GroqInc
, known for its mascot llama, stops by our studios today.
💬 Join the real-time conversations!
#LiveTechNews
…
0
3
12
4
10
25
@lukaszbyjos
@GroqInc
We'll save you the electric bill, and send you some foot warmers instead.
1
0
26
2
5
23
@ramahluwalia
Correction -
@GroqInc
doesn't compete with
@OpenAI
, we uniquely provide a service to affordably run very, very large language models ultra fast. Anyone can run on Groq and get a massive speed boost.
#GroqOn
🚀🚀🚀
4
8
23
got compute?
Groq Says It Can Deploy 1 Million AI Inference Chips In Two Years
2
6
21
@ylecun
@tegmark
@RishiSunak
@vonderleyen
Meta's Open Source models saved Groq
2 years ago we got a panicked call from a CEO deploying the then largest LLM service, he couldn't get any GPUs. They couldn't share the models with us so for 2 years...Nothing.
Then Llama. Now we have insane demand for LPU chips.
Thank you!
0
8
22
🫶
2
3
22
Try
#GroqSpeed
for yourself at .
A question we at Groq used to get before we had LLMs running on our
#LPU
™ Inference Engine was why do
#LLMs
need to run faster than reading speed. No one asks that anymore.
7
4
20
@pmddomingos
We love Nvidia, for training models. For inference, well, we have another offering from
@GroqInc
.
1
4
16
Clever new benchmark from Artificial Analysis.
Per Token pricing is not always 🍏-to-🍏, when normalized provider pricing increases up to +19%
When tokenizing the same text, different tokenizers output a different number of tokens and therefore have different ‘tokenizer efficiency’. Because we pay per token, this means we…
3
2
20
0
1
19
@GroqInc
Thanks for access! The speed is insane!
Note how similar the calls are. Makes it super easy to get going..
3
2
31
0
1
18
2
2
15
@canipeonyou
@nvidia
Nvidia's been talking to a lot of CFOs recently - explaining why they're still economical for inference.
We believe in show rather than tell, so let's count how many LPUs get deployed by the end of 2025. That'll answer that question 😁
0
0
15
We now have definitive proof that fast tokens are worth more than slow tokens.
3
2
14
@ramahluwalia
and I start getting philosophical around timestamp 41:17. What's next after the Turing test? What is insight and can AI have it? Can an AI experience a Eureka! moment? What is sentience, and are we even sentient?
4
3
11
2
1
12
@samlakig
@satnam6502
Your math seems to be accurate :)
The weights for Mixtral are smaller, the KV cache is a little larger (32K), but overall, close.
The reason we're so low power is that we keep the model in the on-chip memory (SRAM). HBM requires almost 10x the energy per token at this latency.
2
1
9
@ramahluwalia
Hey
@satyanadella
,
Have you considered
@GroqInc
's LPU™ Inference Engine? ;)
#GroqSpeed
1
4
9
We love the work
@lmsysorg
is doing! Paged Attention, Chatbot Arena, LMSYS-Chat-1M. Very cool. I'm really impressed with their performance on
#llama2
70B in the video.
Of course, LPU vs. GPU, I think we all know how this video is going to end.
#GroqSpeed
2
1
10
@tomjaguarpaw
@Russell_AGI
@GroqInc
We have plenty of capacity:
As for our next-gen chip (in addition to those 1 million LPUs), we're the lead customer for Samsung's 4nm plant in Taylor Texas, which is what Tom Ellis shared.
We're going very, very high volume 😁
2
1
10
@samlakig
@satnam6502
We measure in joules per token. Under full load we're expecting it to get down to less than 3J / Token
1
0
9
Love the wow factor!
#GroqSpeed
We love when
#Groqsters
show people our demo out in the wild. This was last night at Schiphol airport and it's really fun to see the reaction of end-users to
#groqspeed
. Try it yourself over at and reach out for API access requests.
0
2
11
0
2
8
Please no one name any models grizzly or honey badger.
0
1
9
Choose your own startup adventure:
0
3
8
@sharongoldman
@GroqInc
@mattshumer_
@nvidia
Thank you
@sharongoldman
. I asked the LLM to summarize, but it said the content was best read in full 😉 That said, trying anyway:
#LPU
== Insanely Fast Inference
6
1
7
@NaveenGRao
@chamath
Thanks! We're huge fans of what you're doing over at Mosaic/Databricks :)
0
0
7
@GroqInc
's LPU™ chips run LLMs at crazy speeds, e.g. 70B parameters models at 300 tokens / second. How fast do you think we can run a 7B parameter model like
@MistralAI
's 7B model? Watch below to find out
👇 😂
2
1
7
@TonyHawkersCRC
-- We built our own chips - that's how we're faster. The video you saw was Meta's 70B parameter model running on our LPU™ Inference Engine, rather than on GPUs.
2
3
5
1
0
7
@JoshMiller656
@IntuitMachine
It was the only picture of a GPU we could find that hadn't been photoshopped to look more regular than it is 😉
1
0
6
Hey
@SamA
, glad to see you back at
@OpenAI
.
As you've said, technology magnifies differences, and the difference here👇 is clear.
#BetterOnGroq
When the dust settles, let's build something together.
#GroqOn
1
3
7
0
0
5