We’re all frustrated – even Nvidia. The pandemic upended manufacturing plants resulting in a global shortage of the GDDR6 SDRAM needed to make RTX 30-series GPUs.

I’ve received a number of requests from L7 readers asking me to share my upcoming MIT ChipBrain RTX 30-series GPU deep learning build; but like many of you, I couldn’t purchase the RTX 3090s for the last three months.

Today, I successfully purchased four RTX 3090 GPUs – in a few weeks, I’ll share how to build a multi-GPU RTX 3090 deep learning workstation that outperforms Lambda’s $34k machine for 1/3 the cost. You can subscribe to get notified when it comes out [here].

In this post, I give five “hacks” to obtain RTX 30-series GPUs (3060 ti, 3070, 3080, 3090) + where you can buy an RTX 3090 GPU right now (as of 12/8/2020 at 4:15pm EST). I’ll update this page as I learn other sources, so please post your own tips and tricks in the comments below.

RTX 3090 GPUs in stock for $2000 now

As of 12/8/2020 at 4:15pm EST, I confirmed with an employee at Computer Upgrade King (CUK) that they have 1,700 PNY RTX 3090 GPUs available to ship in the United States. (Yes, the company is named CUK, but let’s move past that.) The easiest way to purchase one a PNY RTX 3090 GPU is on Amazon – please use [this link] to support the L7 blog at no extra cost to you (via Amazon Affiliate link).

If you are unable to find these on Amazon, try calling CUK and purchasing over the phone.

These GPUs sell out faster than I can write a blog post, so please comment below if updates are needed.

Hack 1: Street Merchant – scan websites automatically

An L7 reader, Dan Vachalek, reached out to share his secret for finding RTX 30-series GPUs – a command line tool called: Street Merchant. The tool has been pre-configured to constantly search for RTX GPUs across Amazon, MicroCenter, Walmart, Newegg, etc. When they become available, you can modify the .env settings file to send you a slack notification or email. It takes <10 minutes to set-up.

Vendor websites change frequently, so in practice, I found the Street Merchant tool to give too many false positives to be useful, but it does work, with patience.

Hack 2: Micro Center employee

If you have a Micro Center near you – visit in person and chat with as many of the floor staff as possible. Ask them to send you an email when they get GPUs in stock. Eventually, someone will add you on an email list – when they arrive, your new best friend will let you know and you can come by in-person and pick one up. That’s one of the easiest ways to purchase a single RTX 30-series GPU. Note Micro Center has a “one per household” policy.

Hack 3: NextWareHouse.com backordering

You can backorder purchase these GPUs from NextWareHouse.com. Don’t let their low reseller rating entirely dissuade you. This is a backorder website, so it can take some time for them to fulfill requests, but two of my colleagues successfully obtained 6 RTX 3080 and 10 RTX 3090 GPUs from NextWareHouse.com (after waiting a week longer than expected), and all GPUs arrived in working condition.

My guess is that NextWareHouse.com’s negative reviews on nearly every company review website disuades people form purchasing from them, hence why they sell out slower. But purchase at your own risk (I have no affiliation with NextWareHouse) and you’ll need patience as it may take a few weeks to get off backorder and there can be delays.

Hack 4: eBay.com

You can purchase the RTX 309-series GPUs on eBay, but I don’t recommend it. Many eBay sellers are upset about the scalpers and in retaliation, create fake pages to fool bots into purchasing GPUs that don’t exist. Be careful on eBay for the next few months.

Hack 5: Share with others here on L7

If you find a vendor with RTX 30-series GPUs in stock, please share in the comments – use this to crowd-source solutions to buy these GPUs. If you find them, buy what you need, then share with other engineers and researchers so they can, too!

  1. Relative positioning of GPUs for optimal speed: [view this post]
  2. How to build a multi-GPU deep learning machine: [view this post]
  3. Build Lambda’s state-of-the-art 4-GPU rig for $4000 less: [view this post]
  4. Benchmarking RTX 20-series GPUs for deep learning: [view this post]

Acknowledgements

The GPUs and machines used for these benchmarks were purchased using a grant from the The MIT-IBM Watson AI Lab, MIT Quanta Lab, and the MIT Quest for Intelligence.