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...

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...

I'll be at SC25 next week repping @theregister.com for the fourth year running. Give me a shout if you're going to be in St Louis for the event.

Nvidia's biggest scale up domain is 72 GPUs. Google's is 9,216 TPUs. Historically TPUs have trailed on FLOPS, memory, & bandwidth. That's no longer the case with Ironwood. Google has a Blackwell-class TPU with absurd scale. More on @theregister.com ⬇️ www.theregister.com/2025/11/06/g...

I might be crucified for saying this, but OEM GPU servers are boring now. Everything is just a rebadged HGX box or NVL rack now. The only point of differentiation becomes whose lights out management interface does it have.

Say what you will about the hardware or the software, @tenstorrent.bsky.social's Blackhole QuietBox is a gorgeous bit of kit. Full review is already in the works #AI #Workstation #watercooling #Tech

This was a fun review. I don't think folks realize how big a deal the DGX Spark is. A year ago an #Nvidia #workstation with 128GB+ of VRAM cost tens of thousands of dollars. Is it better than AMD's Strix Halo? Now that's the real question. #DGXSpark #AI www.theregister.com/2025/10/14/d...

On the road again.

If you can't use AI then it's bye bye, Accenture tells staff.

I’m embarrassed to admit that I’ve never held a QSFP DD cable before. It’s enormous compared to the SFP+ DACs I’m used to.

I love my dog. He's usually a very good boy. But having adopted him as a 10 week-old puppy less than a year before COVID hit, his anxiety can be overwhelming at times. He had a vet visit today. He got his shots, but wouldn't let the Dr. do a proper exam. We get to come back in 4 weeks and try again

Look what just landed in the lab

Something curious I’ve noticed is I’m using virtualization less in my homelab preferring instead to run bare metal with containers. I still keep a PVE box for VMs when they’re warranted but a lot of the stuff I’m doing can be achieved using containers. #Linux #VM #Homelab #tech

Ever since Nvidia started talking about disaggregated inference architectures at GTC this spring, I had a feeling a HBM-less prefill accelerator was only a matter of time. My latest for @theregister.com www.theregister.com/2025/09/10/n... #Nvidia #AI #Datacenter #Servers #HPC

Oof. I can relate. Last night my email was filled with TrueNAS warnings. A drive reported Smart Errors. Logs: extended test failed. short test fails. Yep dead drive. 😟 Backup my core files and drop in the cold spare. 6 hours of resilvering left to go. www.theregister.com/2025/09/03/m...

New bench who this?

After three years of 24/7 service in my homelab my R9 3900X met its end on Sunday. During routine thermal paste change the cooler ripped it from the socket bending several pins. Alas even when bent back into position it refused to post. RIP my friend. #AMD #CPU #PC #Homelab #Tech

In my latest hands on for @theregister.com I break down everything you need to know to run large language models in the privacy of our home using Llama.cpp. www.theregister.com/2025/08/24/l... #AI #PC #LLM #HomeLab

Some stories are just made for puns, not to mention highlighting F-grade security. Yes, I did have fun with this.

So wait. Is Arm really getting into silicon? Seems like a sure fire way to piss off your customers. Maybe chiplets would be okay? Most Arm customers take off-the-shelf cores anyway. My latest for @theregister.com www.theregister.com/2025/08/19/a... #Arm #CPU #Chips #tech

The way the report describes the B30A it sounds a lot more like a B300 NVL than a H20 replacement... www.theregister.com/2025/08/19/n...

AMD Radeon Inference perf in Llama.cpp is an interesting conundrum. The Vulkan backend offers higher token gen but ROCm is vastly superior in prompt processing. Which do you opt for?

Big Dyson Sphere can be sung to Pink Pony Club

The funniest bit to me was the comparison of US chip tracking to the machines in The Matrix. Last I checked, Uncle Sam wasn't using human batteries to power major industry or AI development. My latest for @theregister.com www.theregister.com/2025/08/18/c...

Cisco, Arista, Broadcom and every other Ethernet equipment vendor is jazzed about AI and its not hard to see why when when ever GPU is an excuse to sell 3-5 of the fastest switch ports money can buy: My latest for @theregister.com www.theregister.com/2025/08/15/e...

Called it! GPT-5 was a cost cutting measure. "We have better models, and we just can't offer them because we don't have the capacity. We have other kinds of new products and services we'd love to offer," Altman said. ICYMI: www.theregister.com/2025/08/13/g...

Sneak peek at what I've got cooking:

Compiling code on a Raspberry Pi 5 is painfully slow.

OpenAI's decision to use MXFP4 datatypes as the standard for gpt-oss is a big deal and sets the tone for the rest of the industry. Is it perfect? No, but it's a huge improvement over FP4 or INT4. Find my deep dive @theregister.com www.theregister.com/2025/08/10/o...

Main thing confusing me: why give up an optimized training stack built around Nvidia to take a chance on Huawei? Like other than national pride. The 910C doesn't support FP8, so back to BF16. Were they using them for RL? Latest for @theregister.com www.theregister.com/2025/08/14/d...

Gentle reminder that GPUs aren't just for AI. In the right hands they can advance life saving science. Boffins at LLNL used the No. 1 ranked supercomputer, El Capitan, to build a real-time tsunami forecast system. My latest for @theregister.com www.theregister.com/2025/08/13/t...

Got GPT-OSS 120B running tonight in Llama.cpp. ~60Tok/s with minimal expert offload to DRAM. Had to compile a pull request to make it play nice to Open WebUI, but I have tool calling now. Need to figure out a system prompt cause my god this thing loves tables.

Meanwhile, if you haven't figured it out already GPT-5 is a huge improvement, just mostly for OpenAI's bottom line. This thing is a big cost cutting measure and not necessarily in a bad way: My latest for @theregister.com www.theregister.com/2025/08/13/g...

Still need to do some cable management and it remains to be seen whether that PSU will be adequate, but it posted. 24 Zen4 cores 128GB DDR5 5600 RDIMM 2x RTX 3090 TI 24GB w/ NVL And don’t worry those GPUs are power limited to 250W. The PSU shouldn’t explode.

Well of course the Chinese don't want H20s anywhere near their sensitive government workloads. Why give Uncle Sam any more reasons to think kill switches are a good idea. My latest for @theregister.com www.theregister.com/2025/08/12/c...

Challenge: go a week with gpt-oss as your only AI assistant. Could you do it? I have a feeling it's gonna be painful, but I'm going to try.

I did a thing…

When Llama 4.1?

OpenAI’s gpt-oss really loves tables and it’s exceedingly obnoxious on mobile.

After having a few days to mess with OpenAI's gpt-oss — specifically the 21B parameter version — I gotta admit, these are the most impressive open models to come out of the US so far this year. A 🧵 [1]

Sigh... This is the lasting Intel needs right now www.theregister.com/2025/08/07/r...

gpt-oss-20b has some rough edges. Like insisting that Biden is president. It's not just the hallucinations. OpenAI's protections against prompt injection work a bit too well. Once it locks in an answer, good luck changing its mind. Story = @theregister.com www.theregister.com/2025/08/06/o...

I’m currently fighting with OpenAI’s 20B param open model. It insists that Star Trek (1966) aired on CBS. And it refuses to admit it’s wrong.

So real talk. Is this really a second renaissance for nuclear or a distraction from all the gas plants we’re building to fuel AI? I want so badly for nuclear to make good on all the promises of my youth. And if can… space next? www.theregister.com/2025/08/04/f...

The problem with AI isn’t just that it uses a lot of power, it’s how it uses it Like a volume dial that jumps straight from 1 to 11 with nothing in between. Modern grids weren’t designed for that. Which makes Googles latest move all the more important. www.theregister.com/2025/08/04/g...

AMD preps rack-scale Helios systems to contend with Nvidia's Vera Rubin NVL144

AMD bets on rack-scale compute to boost AI efficiency 20x by 2030

AMD's MI355X is a 1.4 kW liquid-cooled monster built to battle Nvidia's Blackwell And the House of Zen wants to put 128 of them in your rack www.theregister.com/2025/06/12/a... #AdvancingAI #AMD

What again are we advancing? I can’t remember. #AMD #AI

Just an FYI for those that don't know. AI needs lots of optics not because light is faster. After the electro-optical exchange at either end, its usually slower than copper. Optics are used because at 400G or 800G, copper can only go a few meters. www.ft.com/content/f08d...