Recent content by bwang

and finally we have an explanation why these are no longer available. the voltage glitching attack allows overwriting the device ID so in the future 170HX will be indistinguishable from a retail A100

3x 5090 will prefill faster than 1x Pro 6000 and decode about the same. With 4x 5090 you can try something like PP=2 TP=2 which should give you ~2x decode at the expense of some prefill, but it also requires a 4x PCI-e 5.0 x16 board (expensive RDIMMS, $30/GB!) and patched Linux kernel modules to...

Finally got all the pieces of the 1 TB system put together and deployed MiMo-V2.6-Pro on it. I have to say, I'm impressed - I think it may have solved an open math research problem on the first try. It also absolutely flies at over 6000 tokens/sec prefill and 70 tokens/sec decode.

I am so confused. Why is this worth $600 shipped?

arena is A/B comparisons in short form chat which tends to favor things such as style, brevity, and formatting. it also doesn't test multi-turn tool call performance, error recovery, or planning which are important for coding.

oof, come on Nous who is doing your marketing, I literally cannot tell what the new feature is from this post. It seems like they've added the ability for the main session to chat with the sub-sessions? That's a good feature - it prevents main session context from getting poisoned with failed...

There is something reassuring about being able to run frontier models at home even if it makes no financial sense. To achieve this, we describe two configurations capable of serving GLM-5.2/5.3 in a modest (sub-$20K) budget that fit in 220V 50A residential service. The goal was to achieve...

What is it? The Arc Pro B50 is a unique bus-powered low profile card with 16 GB of VRAM. It is very slow: 85 TFLOPS bf16 XMX throughput, 170 TOPS int8 XMX throughput 224 GB / sec of memory bandwidth However, "slow" is relative - the B50 has more throughput and bandwidth than a $5000...

Doing things a bit differently: - Client hardware + models: stuff which runs on your existing gaming rig or something that looks like one. Good idle, OK cost, and the computer still usable for other things. All the Qwen-27B people will wind up here - Datacenter hardware + models: stuff which...

if it is anything like the nvidia stuff figure max power during prefill for a couple minutes at a time, 80w during decode, and max power generating images/videos

the static pressure drop across the server cards is insane. have Claude code you up something that does full-RAM-sized bf16 GEMMs or spin up ComfyUI and have it generate some long videos with Minimax-H3 and keep an eye on the HBM temps as you tune cooling, that is a $4000 card so I'd err on the...

I just spent all Saturday chasing a 1-in-a-million ghost serving GLM-5.2 on my new 640 GB system (10x bootleg A100) and let me tell you, I've never hated computers this much... As for actual system specs, see sig.

what's the goal here? collectibles or building a working rig? it sounds like you want the original packaging out of a Rampage or Extreme11; that's gonna be really hard to come by ten years in lol. arguably the box is worth more than the motherboard itself now, because I can't imagine anyone...

ygpm

ygpm on the laptop