Sign in

Tobias Mann

@tobiasmann.bsky.social
868 followers 374 following 260 posts

Systems Editor at TheRegister / SitPub — hiker, animal lover, photographer, blogger, and tech journo.

PostsRepliesMedia
Tobias Mann @tobiasmann.bsky.social · 25/08/2026
When it comes to inference, compute is key, but memory bandwidth is king. With 128 accelerators OpenAI's (& Broadcom) Jalapeño offers ~2 PB/s of HMB4 B/W — more than either Nvidia's Vera Rubin or AMD's Helios My 1,000+ word analysis only @TheRegister www.theregister.com/systems/2026...
theregister.com
OpenAI's upcoming Jalapeño chip looks like it'll be an inference beast
128 chips, 1.7 exaFLOPS, and 27 TB of HBM give Altman and crew a leg up over Blackwell, and maybe even Rubin
000
Tobias Mann @tobiasmann.bsky.social · 24/08/2026
How potent are Nvidia's Groq-3 LPX racks? Very, but Gemma 4 31B tells an idealized story. The real trick will be how efficiency it scales to MoE. My full 1,200 word analysis only @theregister.com www.theregister.com/systems/2026...
theregister.com
What Nvidia's first Groq 3 LPU benchmarks tell us about its $20B gamble
Gemma 4 31B performance tests offer a best-case scenario for next-gen dataflow accelerators
151
Tobias Mann @tobiasmann.bsky.social · 06/07/2026
The hardware is neither fast nor cheap, but with up to 128 GB of memory, there aren't many AI+ML workloads AMD's AI Halo can't run. The question is: is it worth $4K: www.theregister.com/ai-and-ml/20...
theregister.com
AMD’s Ryzen AI Halo makes local AI look easy, but at $4K, easy doesn't come cheap
128 GB of memory! In this economy?
110
Tobias Mann @tobiasmann.bsky.social · 01/06/2026
This vulture is now circling over Taiwan. Stay tuned for a flood of Computex and GTC Taipei news this week.
000
Reposted by Tobias Mann
Matt Rosoff @mattrosoff.bsky.social · 04/05/2026
Where else are you going to find this much detail on rolling your own local agents? Only at @theregister.com courtesy of @tobiasmann.bsky.social with an assist from @lot49.com www.theregister.com/2026/05/02/l...
theregister.com
How to roll your own local AI coding agents
: Take those token limits and shove them by vibe coding with a local LLM
363
Tobias Mann @tobiasmann.bsky.social · 23/04/2026
ICYMI: AMD's 9950X3D2 isn't for gamers. It's also not for content creators. So who is it for? We get to the bottom of this weird little beast of a chip in our review over on @TheRegister www.theregister.com/2026/04/21/a... #AMD #CPU #PC #Workstation
theregister.com
AMD's Ryzen 9 9950X3D2 Dual Edition tested
Review: An $899 CPU? In this economy?
020
Tobias Mann @tobiasmann.bsky.social · 21/04/2026
An $899 CPU in this economy? AMD's Ryzen 9 9950X3D2 with its 208 MB of system cache hits store shelves tomorrow. But is it worth the $200 premium over the 9950X3D? I cover all of that in my latest review for @theregister.com www.theregister.com/2026/04/21/a... #PC #CPU #AMD #Tech
theregister.com
AMD's Ryzen 9 9950X3D2 Dual Edition tested
Review: An $899 CPU? In this economy?
000
Reposted by Tobias Mann
Matt Rosoff @mattrosoff.bsky.social · 24/03/2026
Big first for Arm: It's making its own chips - @tobiasmann.bsky.social has all the details here. tip @Techmeme
021
Tobias Mann @tobiasmann.bsky.social · 23/03/2026
I prefer systems level analysis, but if someone wants us to look at their CPU or GPU we'd certainly entertain it.
100
Tobias Mann @tobiasmann.bsky.social · 23/03/2026
In case you missed it, @theregister.com does CPU reviews now! Our review of Intel's Core Ultra 270K and 250K Plus processors dropped this morning. www.theregister.com/2026/03/23/i...
theregister.com
Intel Core Ultra 270K, 250K Plus aim for budget PC builders
Review: More cores, higher clocks, and lower prices? What's not to like?
130
Tobias Mann @tobiasmann.bsky.social · 17/03/2026
Here's a closer look at Nvidia's new Vera CPU rack blade. 8x 88-core Vera CPUs with 1.5 TB of LPDDR5x memory a piece. #GTC
100
Tobias Mann @tobiasmann.bsky.social · 16/03/2026
Sorry getting Love Death and Robots vibes from this keynote closer.
000
Tobias Mann @tobiasmann.bsky.social · 16/03/2026
Tell me you've never played Five Nights at Freddy's without telling me you've never played Five Nights at Freddy's.
110
Tobias Mann @tobiasmann.bsky.social · 16/03/2026
I know right!!!!!
000
Tobias Mann @tobiasmann.bsky.social · 16/03/2026
NemoClaw is "Safe and secure." Those words will never come back to haunt Jensen, will they?
020
Tobias Mann @tobiasmann.bsky.social · 16/03/2026
If OpenClaw terrifies you, it should. But don't worry, Nvidia is going to make it "enterprise ready"
110
Tobias Mann @tobiasmann.bsky.social · 16/03/2026
Who needs SaaS when you could have AaaS!
000
Tobias Mann @tobiasmann.bsky.social · 16/03/2026
Every time I hear about OpenClaw this scene pops into my head.
media.tenor.com
buzz lightyear from toy story says the claw while woody watches
ALT: buzz lightyear from toy story says the claw while woody watches
011
Tobias Mann @tobiasmann.bsky.social · 16/03/2026
In Space, no can hear... your tokens stream:
000
Tobias Mann @tobiasmann.bsky.social · 16/03/2026
Oh great. AI that runs the datacenters that train and run AI. Can't see how that could go off the rails.
media.tenor.com
a robot with a purple light behind it says who me
ALT: a robot with a purple light behind it says who me
000
Tobias Mann @tobiasmann.bsky.social · 16/03/2026
Jensen says the time for optical scale up has come.
000
Tobias Mann @tobiasmann.bsky.social · 16/03/2026
Groq 3 LPU shipping in 2H "Probably about Q3," Jensen says. Manufactured by Samsung.
000
Tobias Mann @tobiasmann.bsky.social · 16/03/2026
That's quite a return on investment if Nvidia can actually pull it off.
000
Tobias Mann @tobiasmann.bsky.social · 16/03/2026
This tells you everything you need to know about Nvidia's $20B acquihire of Groq.
000
Tobias Mann @tobiasmann.bsky.social · 16/03/2026
Here's a closer look at Nvidia's Groq 3 LPU systems: www.theregister.com/2026/03/16/n...
000
Tobias Mann @tobiasmann.bsky.social · 16/03/2026
Jensen predicts a future in which inference providers will be able to charge $150 per million tokens
001
Tobias Mann @tobiasmann.bsky.social · 16/03/2026
There in the center is Nvidia's new Vera CPU rack system. It has eight 88-core Vera CPUs and 12TB of LPDDR5X memory. 32 of them form a 256 node rack system. www.theregister.com/2026/03/16/n...
000
Tobias Mann @tobiasmann.bsky.social · 16/03/2026
Tokens are the new commodity of the AI age. I feel like a snake oil salesman every time I try to explain this to my non-technical friends and family.
000
Tobias Mann @tobiasmann.bsky.social · 16/03/2026
Nvidia's NVL72 racks may not be cheap ($3.5M+ for Blackwell) but the data shows just how much more efficient they are.
000
Tobias Mann @tobiasmann.bsky.social · 16/03/2026
You want to put [our compute products] in any country, anywhere, we're delighted to support you.... Unless Uncle Sam says no, that is.
000
Tobias Mann @tobiasmann.bsky.social · 16/03/2026
Jensen claims $1 trillion of AI demand through 2027
011
Tobias Mann @tobiasmann.bsky.social · 16/03/2026
110 Robots hiding around GTC? Why does that make me a little uneasy? Oh... right: www.theregister.com/2026/03/13/c...
theregister.com
Bot harasses woman, led away by cops
: An incident in Macau
000
Tobias Mann @tobiasmann.bsky.social · 16/03/2026
Jensen: "Moore's Law is running out of steam." I could have sworn it was dead? Was it not? You told me it was, Jensen.
010
Tobias Mann @tobiasmann.bsky.social · 16/03/2026
Isn't using AI to turn unstructured data into the structured data necessary to make AI trustworthy a bit of a chicken and egg problem?
000
Tobias Mann @tobiasmann.bsky.social · 16/03/2026
Now I get it. Trust worthy AI was just green spaghetti the whole time. According to Jensen its structured data, but I just see spaghetti.
000
Tobias Mann @tobiasmann.bsky.social · 16/03/2026
Who needs raster performance when you upscale and enhance with tensor cores? DLSS 5 announced:
000
Tobias Mann @tobiasmann.bsky.social · 16/03/2026
Jensen: "This is the house that GeForce made."
000
Tobias Mann @tobiasmann.bsky.social · 16/03/2026
20 years of CUDA. My does time fly
010
Tobias Mann @tobiasmann.bsky.social · 16/03/2026
Here we go again #GTC 2026
2641
Reposted by Tobias Mann
Matt Rosoff @mattrosoff.bsky.social · 09/03/2026
Welcome back to the Kettle, The Register's weekly podcast, hosted anew by @bvig.bsky.social. This week, he's joined by @tobiasmann.bsky.social and @jessicalyons.bsky.social to discuss the role of tech in the war in Iran: www.theregister.com/2026/03/09/k...
theregister.com
Iran is the first out-loud cyberwar the US has fought
Kettle: Cyber is no longer the hush-hush thing it used to be, as team Trump invades Iran with hackers taking the lead
143
Tobias Mann @tobiasmann.bsky.social · 29/12/2025
#ICYMI I tested Nvidia's GB10-based DGX Spark against the AMD Ryzen AI Max+ Pro 395-based HP Z2 Mini G1a. For AI development, the GB10 is hard to beat, but for an multi-purpose PC, AMD's Strix Halo is more flexible. Full story: @theregister.com www.theregister.com/2025/12/25/a... #AI #PC
theregister.com
Tested: AMD's Strix Halo vs Nvidia's DGX Spark
Hands On: Two tiny boxes, 128 GB apiece – but very different strengths
020
Tobias Mann @tobiasmann.bsky.social · 26/12/2025
#AMD's Strix Halo (Ryzen AI Max+ 395) and #Nvidia's DGX Spark (#GB10) are an interesting class of hardware for local AI #development. Which is right for you depends on whether you want a #PC that runs AI or a PC for #AI. My latest for @theregister.com www.theregister.com/2025/12/25/a...
theregister.com
Tested: AMD's Strix Halo vs Nvidia's DGX Spark
Hands On: Two tiny boxes, 128 GB apiece – but very different strengths
011
Tobias Mann @tobiasmann.bsky.social · 08/12/2025
The one exception, of course, is Google's TPU pods which are still using optically switched torus topologies.
000
Tobias Mann @tobiasmann.bsky.social · 08/12/2025
If you hadn't noticed, AI infra all pretty much looks like #Nvidia's NVL72 now. #AWS' Trainium3 UltraServers are the latest example For more on the convergent evolution of #AI infra and the future of the #datacenter, find my latest on @theregister.com forums.theregister.com/forum/all/20...
130
Tobias Mann @tobiasmann.bsky.social · 05/12/2025
Here’s what you need to know about Graviton5 - 192 cores - 2MB L2 per core - 192MB of L3 - DDR5 7200 support - 8800MT/s planned - 25% higher perf than AWS dual G4 M8g instances (192 cores vs 2x96 cores) My latest for @theregister.com #reinvent #reinvent25 www.theregister.com/2025/12/04/a...
theregister.com
Amazon keeps pressure on Intel, AMD with 192-core Graviton5
re:invent: The homegrown chips now account for half of all new CPUs added to AWS over the past three years
020
Tobias Mann @tobiasmann.bsky.social · 04/12/2025
And that’s a wrap on Re:Invent #reinvent #reinvent25
030
Tobias Mann @tobiasmann.bsky.social · 02/12/2025
Funny how Amazon's Trn3 UltraServer looks a lot like Nvidia's NVL72... #reinvent #reInvent25
031
Tobias Mann @tobiasmann.bsky.social · 02/12/2025
To compete with Nvidia, Amazon is embracing Nvidia, fusing the GPU giant's NVLink interconnects into its next-gen Trainium4 UltraServers. My latest for @theregister.com theregister.com/2025/12/02/a... #reinvent #reInvent25 #aws #ai #cloud
theregister.com
Amazon to fuse Nvidia's NVLink into Trainium4 accelerators
Re:Invent: Meanwhile, Trainium3 makes its debut promising million-chip training clusters
031
Tobias Mann @tobiasmann.bsky.social · 27/11/2025
The Tenstorrent's Blackhole QuietBox is an impressive bit of #AI kit. - 128GB of GDDR6 - 2TB/s of MEM BW - 3 petaFLOPS of FP8 - 12.8 Tbps of interconnect BW - and an ambitious software strategy that... needs some polish Find my full review on @theregister.com www.theregister.com/2025/11/27/t...
theregister.com
Blackhole QuietBox, Tenstorrent's AI workstation reviewed
hands on: $12K machine promises performance that can scale to 32 chip servers and beyond but immature stack makes harnessing compute challenging
040
Tobias Mann @tobiasmann.bsky.social · 17/11/2025
Europe has joined the US as an exascale super power. EuroHPC's biggest iron has crested 1 EF on the Top500's HPL benchmark. Is HPL still the bench to watch when Blackwell offers more than 200x the FP8 perf as FP64? #sc25 My latest for @theregister.com www.theregister.com/2025/11/17/e...
theregister.com
Europe joins the US as an exascale superpower
SC25: EuroHPC's biggest iron still has more to give with Universal Cluster expansion expected to come online next year
171