BEST SELLERS

In Server Components

Absolutely fantastic card all around. I'm both gaming and doing AI workloads with this GPU. I'm also using 2 of them in parallel on vLLM and it works great.

10GB ethernet port is cool.

I was an early adopter and grabbed a b70 on day 1, it was horrible. It barely worked, pcie 5.0 would not negotiate, drivers timed out, power would drop out, performance was barely above b580 in games, the only thing that was fast was some edge case pytorch training. Even the vllm performance was not great and very limited models supported. After a few months, now in the gamer drivers in windows we get fan control and overclocking, gaming performance really is about on target with a rx 6800 xt or 9060 xt, DCS world runs great, COD Warzone with FG on 1080p XESS Native = 250+ fps. I'm sure not all issues are fixed especially on the model side with llamacpp and vllm but this is now good enough and well beyond day 1. Sure my 7900 xtx smokes it in all gaming aspects but this gpu really is faster than a 4090 in some pytorch training scenarios and 35% faster than a 5080 in bert tokenization. One of the biggest issues for daily life was the fan control on the day 1 intel version of the b70 I had. The gpu was idling 60-70c. The asrock version has the same vapor chamber as the radeon ai pro r9700 which is definitely a better GPU for AI/Training/Inference and about twice as fast gaming. Now with fan control the gpu idles at 28-30c in windows and gaming hits 58-60c with a relatively inaudible (with headphones) fan speed. 270k tuned at 185w tpp pro b70 games and workloads less than 200w Doesnt heat up the room

Bought two of this for my rig with amd ryzen 9950x3d. Very easy to install, nice 2 slot sizing and great quality. Codex did all the configuration and optimization. Main benefits - great price - memory size combination (even at 1700 per card) - good performance- I get 50-60 tokens with 2 gpu and most of large models, up to 70bn - really compact design comparing to most 4090 and 5090 Very happy with this card

Lots of expansion capability for memory, disks, pci-e, and fans IPMI is wonderful for getting Board feels solid and well built (no flex)

96GB of GDDR7 ECC memory per card, Blackwell architecture, and dedicated hardware for AI and ray tracing. Serious workstation capability for demanding projects.

Very spacious Comes with screws Can fit multiple fans and storage drives If you're a Linux nerd like me who wants to do a GPU Pass-through for a Windows VM, this case is it (I Pass-through the Nvidia card btw) Yes, I have mismatch fans, it was all I had for now

New LGA 1851 Lots of room for PCI 5 expansion Onboard BT and WiFi Tons of space for M.2 memory (gen 4 at least)

Keep in mind this card has its limitations, I am looking at it from the stand point of this verse a RTX3090 used (about the same price). No way I am spending $1000 on a used card that can die tmw even with Cuda support. So now I get 32GB vs 24GB ram. I am primary running windows and LMstudio and Ollma worked fine after installing all the intel software (like 4 large installs) Vulkan support gave me around 100T/KS for qwen 3.5-35b. Out of the box with standard setup. This running on an Intel i7-11th gen with 64GB DDR4 and PCI 4 x8 interface.. The biggest hit is on how fast I can load the model in memory. The card is very quite with the fan. I suppose other people got loud fans.

Professional video card. Works great with games as well Ultra low power and efficient video card. Performance on par with a Nvidia 4060, but in a single slot low profile 50 watt card

cost efficient server that performs. x86 is almost at its scaling limit with 5nm process unmatched 5ghz per core performance optimal for running cryptocurrency validators great ipmi

96Gb of vram and energy efficient

-lots of parts -very flexable motherboard mounts -easy to work in -comes with pc speaker

- Good packaging from Newegg, just a smallish box with no branding on the outside. - Obviously it performs quite well. Will use the full 600W under load (but you can limit wattage using `nvidia-smi`). Have a good PSU with a 600W connector (not included) and very good airflow setup. - Working in Arch Linux with just `nvidia-open` and `nvidia-utils` packages.

Minimum memory for models I wanted to work with, even quantized, is 24gb. $2,000+ just to experiment seemed like a lot in today's economy. Intel offered the B60 Pro for $600! So far, it's been everything I'd hoped. Gemma 4 26b and 31b models run comfortably and it's pretty amazing what capabilities they provide for local AI. I found that, for me, LM Studio was a little easier to use than AI Playground. YMMV Be sure to attach a monitor, even if you plan to use remote or onboard graphics. If no monitor is detected the board may not load the drivers. Turn off ECC memory or you'll be working with 21 instead of 24gb. It's astonishing how much free software is out there for GPUs (like the B60) that have enough VRAM to run that software. It's an incredible value proposition for anyone interested in learning about / experimenting with local AI.