The perfect data fabric to fatten starving GPUs

In the era of AI, companies are scrambling to buy the latest GPUs worth tens of millions of won.
However, when you actually open the lid, there are many cases where the GPU is idle and unable to demonstrate its full potential. Why is that?

The answer is “Because the data arrived too late”no see.

GPUs consume data at incredibly high speeds, but the storage and networks that store and transmit that data cannot keep up with that speed. This phenomenon, where the worker (GPU) goes hungry and waits because the food (data) is delayed, is called 'GPU Starvation'.

Today, three key technologies that have brilliantly broken through this frustrating bottleneck, VAST Data, HPE Alletra MP Storage, and QFX SwitchLet me talk about their perfect match.

1. VAST Data: “Perfectly Separating the Brain and Warehouse of Storage”

Simply put, storage is a huge vault for storing data.

In the past, when the number of storage devices was increased due to the volume of data to be stored, the actual work speed slowed down as the devices struggled to synchronize information (cache synchronization), asking each other, "What data do you have?".

VAST DataThis problem was solved with the power of 100% software.
storage's Calculation (brain) roles and Storage (warehouse) They completely physically split the roles.

• C-Box (Head): It is a smart, dedicated computing server responsible for data processing.
• D-Box (Body/Warehouse): It is a dedicated storage space that reliably stores data.

• E-Box (Universal Box): As the most recently released model, it integrates computing (head) and storage (warehouse) functions into a single small and compact chassis, capturing both high performance and scalability.

By clearly dividing roles in this way, we were able to extract data from the D-Box (warehouse) and deliver it at ultra-high speed without any hesitation, even when hundreds of GPUs simultaneously requested data.

2. HPE Alletra MP: “Software Only! A Versatile Framework That Changes Usage”

If VAST is smart software, it requires a 'container (hardware)' that can robustly house it and support it according to the enterprise environment.
This is where HPE Alletra Storage MP (Multi-Protocol) comes in.

In the past, the appearance and structure of storage equipment varied depending on the purpose, such as for files, backups, and databases.

However, HPE

“What if we used a sturdy frame (hardware) that looks exactly the same, but just installed different software?”

He came up with a very efficient idea called...

purpose	Solution Name	Core software installed	Role in the AI pipeline
AI model training	GreenLake for File	VAST engine	Pouring data into the GPU at ultra-high speed ‘Frontline striker’
Data Collection/Backup	Alletra MP X10000	HPE's proprietary cloud S3 engine	Collecting global data ‘'Giant Dam (Data Lake)'’
System/DB Operations	Alletra MP B10000	HPE's proprietary High Availability Block Engine	Stably supporting infrastructure ‘Defender’

When building an AI infrastructure, you can use the X10K for data collection and file storage running the VAST engine for full-scale GPU training. This allows you to unify the hardware form factor into a single form (Alletra MP) while optimizing it for your specific needs.

3. HPE Networking QFX Switch: “Opening a Congested Highway Instead of an Expensive Private Road”

Okay, the data (storage) is ready, and the mouth (GPU) to eat is also ready.
Now, a sturdy road (network) is needed to connect these two.

In the past, an expensive and closed dedicated road called 'InfiniBand' was used to achieve high speeds without data loss.
The performance was good, but you had to use equipment from a specific company, and it was quite tricky to handle.

However, as Ethernet, the internet technology we commonly use, has advanced tremendously, an alternative has emerged.

Right at the center of it all HPE Juniper Networking's QFX SwitchThere is.
Recently, Juniper QFX has received official qualification in the VAST storage environment and is establishing itself as the network standard.

• Takes responsibility for even the internal roads of the storage: Previously, QFX switches were used only for external communication connecting storage and GPUs. However, now QFX5230-64CD(Spine)와 QFX5130-32CD(Leaf) The switch has also been officially certified for use as an internal storage ultra-high-speed network (NVMe fabric) connecting VAST's C-Box and D-Box.

• E-Box Integrated Network Support: Even in an E-Box environment where compute (server) and storage are combined into one, the QFX switch smoothly handles both client traffic and internal storage traffic.

As a result, there is no longer a need to insist on the expensive and difficult-to-handle InfiniBand.

With just a single proven QFX switch, it is now possible to seamlessly traverse the entire AI network—from internal storage communication to external GPU connections—into a 'single Ethernet highway.'.

---

The three puzzle pieces to solve the chronic problem of AI data centers, 'GPU starvation,' have now all been put together.

Software that sends data without limits VAST Data, a versatile frame that transforms freely to suit the purpose HPE Alletra MP, and perfectly replaced expensive dedicated roads HPE Networking QFX Switchuntil.

This open, vendor-free full-stack combination will be the most realistic and powerful breakthrough for many engineers and companies grappling with AI infrastructure. Isn't it time to clear out our company's clogged AI data pipeline?