Gloom

research @CORSAIR | local compute is fun :)

[Gloom @Xyoskeleton]

@elliotarledge nice, gl on the build. if you ever have any issues with the corsair psu do lmk!

[Gloom @Xyoskeleton]

@mov_axbx some cases include accessories to help support SSI-CEB form factor at least our cases do/did but it's niche so wouldn't be shocked if many didn't these days.

[Gloom @Xyoskeleton]

@elliotarledge Fair enough, Corsair PSUs 👀?

[Gloom @Xyoskeleton]

@Hikari_07_jp Yes! We designed it this way. Even if you had our RM-series power supplies, they use what we call "Type-4" cables so if you ever upgrade to an HX or HXi series (besides the "Shift" models), those same cables will work. If you have any questions feel free to reach out anytime

[Gloom @Xyoskeleton]

@MikeBradleyAI @claudeultramax @TheAhmadOsman 9000D case ?

[Gloom @Xyoskeleton]

Not surprised lol. Has happened in 2002 and 2017 just search up "DRAM price fixing scandal" I believe GamersNexus has a video on this too from back then

[VGC @VGC_News]

The three largest RAM manufacturers in the world are being sued for allegedly working together to fix the supply and price of memory, leading to the recent surge in price for consumers.

https://t.co/C66OWNd083

[Gloom @Xyoskeleton]

Been using @Baidu_Inc Unlimited OCR for channel + marketing automation workflows and it’s been great.

I made an optimized STRIX Halo port since the original release was NVIDIA-focused. Tested it over the last couple days and really enjoying it :)

Link: https://t.co/It94402ZWo

[Baidu AI @BaiduAI_News]

We’re open-sourcing Unlimited OCR — built to read long documents in one pass.

With 3B total parameters and only 500M activated, Unlimited OCR sets new end-to-end SOTA results on OmniDocBench v1.5 and v1.6.

The key innovation is Reference Sliding Window Attention (R-SWA), inspired by how humans transcribe books: keeping the source, recent context, and next words in focus, while softly forgetting what’s no longer needed.

With constant KV Cache size and lower attention cost, Unlimited OCR can transcribe 40+ pages in a single forward pass — without losing context or slowing down.

Explore the model👇:
--GitHub: https://t.co/5ZJBsEldKd
--Hugging Face: https://t.co/4FKFr9EfOu

[Sandro @pupposandro]

A few weeks ago, with @davideciffa, we experimented around power capping our 3090s.

And we discovered a sweet spot at 220W where you can get ~92% of max throughput at ~58% of the power.

Qwen3.5-35B-A3B Q4_K_M on a single RTX 3090:
• 320W: 115.2 tok/s, 0.381 tok/s/W, 76°C
• 220W: 105.4 tok/s, 0.436 tok/s/W, 64°C
• 200W: 95.8 tok/s, 0.438 tok/s/W, 60°C

10% throughput loss for 40% less power.
Fans basically silent.

We think this is a no-brainer. Lower bill, lower temps (particularly important in the upcoming summer), longer GPU life.

[Gloom @Xyoskeleton]

@loktar00 imo a good buy now, I think intel will start maturing and supporting their drivers for AI like AMD is doing for ROCm esp with whatever they decide to cook up next for their next gen AI hardware

[Gloom @Xyoskeleton]

@theo used market is really good right now! I deff think memory prices will get better next 2 years but I think we'll see periods were pricing is good in one area (i.e memory) then storage will get worse and then swap

[Gloom @Xyoskeleton]

One of my favorite demos from @computextaipei was this @CORSAIR AI300 / AMD Strix Halo node cluster we put together.

The idea was pretty simple: show what local AI can realistically look like for businesses and entrepreneurs that do not need to spend heavily on API calls or use models that are overkill for the work they are doing. We used the Xeneon Edge to demonstrate the difference in token costs compared to running things locally, and to show how local AI can be a more practical and cost-effective option for certain use cases.

It was really exciting to see how well people responded to it and to watch that “light bulb” moment happen when we explained it in this way. I still think local AI has a long way to go, but proof-of-concept work like this helps show people that, in the near future, this kind of compute will be much more accessible and can make a meaningful difference in bringing local AI to life without breaking the bank.

[Gloom @Xyoskeleton]

@BlueJayvideos You bought a real good chip the CL30 SK Hynix chips we packaged on that SKU is so good can't believe it's gotten so expensive now. If you ever have any issues (shouldn't ever but ya never know) feel free to reach out anytime

[Gloom @Xyoskeleton]

@RealDBell golden era fr, I miss those days, that was around the time I joined Corsair completely different world now hahahah

[Gloom @Xyoskeleton]

@7384254b brother you used to be able to buy 128GB of memory for sub 300 at one point

[Gloom @Xyoskeleton]

It’s not really cost cutting, it’s supply. Vendors sometimes quote what they have, and you either take it or gamble on a better option next time. DDR5 single vs dual-stick performance is nbd but imo they prob decided this so they can launch on time. So I get it.

[Osvaldo Pinali Doederlein @opinali]

The detail about the Steam Machine configured with single-channel RAM, despite board supporting a second stick, is really the last straw. Amazing you'll pay up to $1.4K and they still do this last-minute cost-cutting move to further nerf that CPU/GPU. (Image from @digitalfoundry)

[Gloom @Xyoskeleton]

@gabriel1 i thought about this a lot, this is like one of the best use cases for LLMs is to automate things like this. we automate the validation process for things like keyboards/mice same should be applied to things like web-apps etc using AI, it's just too slow (getting better tho).

[Gloom @Xyoskeleton]

@marvv_eth i don't blame you i'm the same way, it's gotten hard even for me to get product LMAO but it'll get better