Graid Technology Inc.

Your NVMe array is quietly running at 12–18% of line rate. ‼️ That's the throughput half of the storage tax — the silent percentage you lose to a hardware RAID controller that can't process parity at NVMe speeds. Rip out the controller and go to software RAID? You've traded throughput for CPU. Real customer workloads show 18–28% of host cores burned on RAID I/O processing. Cores you provisioned for databases, inference pipelines, and VMs — now doing storage housekeeping.

Either way, the array you paid for runs at a discount.

And with enterprise NVMe pricing up ~257% since Q2 2025, the same tax rate now hits a much bigger check. Every underperforming drive is more expensive than ever. Every burned CPU core competes with AI workloads for the same silicon.

SupremeRAID™ is a third path: software RAID with the I/O workload running on an NVIDIA GPU instead of host CPU. Full NVMe line rate. CPU returned to your applications. Enterprise-grade RAID 5/6/10 protection. One card.

Garrett McKibben, Senior Director of Technical Marketing, wrote up the full breakdown — why the tax exists, what it actually costs at today's drive prices, and how GPU-accelerated RAID eliminates both halves of it. Read the blog here: https://zurl.co/1f87e

And while you're at it — the Summer 6 Pack Promotion is live through September 30: buy 5 SupremeRAID™ licenses, get the 6th free! Details in the blog.

Storage Tax: Fix Underperforming NVMe Arrays | Graid Technology Hardware RAID and software RAID both charge a tax on your NVMe investment — one in throughput, one in CPU. With enterprise NVMe pricing up ~257%, here's wh

🏆 We are proud to be recognized in the 2026 Taiwan AI Awards, standing out among 137 outstanding companies.

As AI infrastructure evolves, the memory hierarchy is being redefined: when HBM is not enough, AI turns to DRAM; when DRAM is not enough, SSD becomes the next layer of scale.

In this new AI computing architecture, storage is no longer a passive backend resource. It is becoming a critical performance layer that determines how AI systems scale, accelerate, and operate efficiently.

Graid Technology is proud to extend GPU-powered storage acceleration into the AI era, helping organizations unlock higher performance, stronger data protection, and greater infrastructure scalability.

This recognition reflects the dedication of our team and the continued trust of our customers and partners. Thank you for being part of our journey.

🚀 We look forward to continuing to push the boundaries of AI storage and helping shape the future of AI together.

Learn more: https://zurl.co/VlHz8

GPU memory pressure is a real infrastructure bottleneck — and it's getting worse as context windows grow. 📈

As long-context LLM inference scales, KV cache demand scales with it. When the cache working set outgrows GPU memory, serving stacks either degrade or teams are forced to scale compute nodes just to get more local SSD headroom. That's an expensive and inflexible answer.

We built the SupremeRAID™ KV Cache for Rack to offer a better one.

The architecture is straightforward: a dedicated SupremeRAID™ NVMe storage appliance connects to GPU servers over high-speed Ethernet and exports a shared NFS cache path via LMCache. GPU servers stay focused on inference. Cache capacity scales independently — without touching the compute fleet.

We validated it with a production-grade benchmark:

🖥️ Model: Qwen3-235B on 4x NVIDIA H200 GPUs

⚙️ Stack: vLLM + LMCache + EvalScope multi-turn workload

📦 Storage: 10 × KIOXIA CM7-V NVMe SSDs in RAID 5 on Supermicro SSG-221E-DN2R24R

🔗 Network: Dual 200 Gb/s Ethernet

Total Throughput improvement vs. no KV cache offload:

→ +32.7% at 512-token prefix length

→ +47.0% at 2,048-token prefix length

→ +51.6% at 4,096-token prefix length

→ +53.4% at 8,192-token prefix length

Zero fails across all 512 requests, at every prefix length tested.

The trend is the story: the longer the reusable context prefix, the more the external cache tier contributes. And inference workloads are moving toward exactly this shape — longer contexts, higher concurrency, more multi-turn depth.

Graid Technology has qualified SupremeRAID™ KV Cache for Rack on 20 storage server platforms across AIC, Dell, Giga Computing, Lenovo, and Supermicro — giving architects validated options across a range of form factors, processor platforms, and NVMe densities.

📄 Read the full white paper — benchmark methodology, test configuration, and complete results — here: https://zurl.co/vSyi1

GPU memory pressure kills LLM throughput. SupremeRAID™ KV Cache for Rack externalizes the cache tier — and delivers up to 53.4% higher throughput on Qwen3-235B + NVIDIA H200. Get the full benchmark results here: https://zurl.co/XWNmm

White Paper: SupremeRAID™ KV Cache for Rack | Graid Technology SupremeRAID™ KV Cache for Rack delivers up to 53.4% higher LLM throughput by externalizing the KV cache tier. See the benchmark results.

GPU memory pressure is a real infrastructure bottleneck — and it's getting worse as context windows grow. 📈

As long-context LLM inference scales, KV cache demand scales with it. When the cache working set outgrows GPU memory, serving stacks either degrade or teams are forced to scale compute nodes just to get more local SSD headroom. That's an expensive and inflexible answer.

We built the SupremeRAID™ KV Cache for Rack to offer a better one.

The architecture is straightforward: a dedicated SupremeRAID™ NVMe storage appliance connects to GPU servers over high-speed Ethernet and exports a shared NFS cache path via LMCache. GPU servers stay focused on inference. Cache capacity scales independently — without touching the compute fleet.

We validated it with a production-grade benchmark:

🖥️ Model: Qwen3-235B on 4x NVIDIA H200 GPUs

⚙️ Stack: vLLM + LMCache + EvalScope multi-turn workload

📦 Storage: 10 × KIOXIA CM7-V NVMe SSDs in RAID 5 on Supermicro SSG-221E-DN2R24R

🔗 Network: Dual 200 Gb/s Ethernet

Total Throughput improvement vs. no KV cache offload:

→ +32.7% at 512-token prefix length

→ +47.0% at 2,048-token prefix length

→ +51.6% at 4,096-token prefix length

→ +53.4% at 8,192-token prefix length

Zero failed requests. Across all 512 requests. At every prefix length tested.

The trend is the story: the longer the reusable context prefix, the more the external cache tier contributes. And inference workloads are moving toward exactly this shape — longer contexts, higher concurrency, more multi-turn depth.

Graid Technology has qualified SupremeRAID™ KV Cache for Rack on 20 storage server platforms across AIC, Dell, Giga Computing, Lenovo, and Supermicro — giving architects validated options across a range of form factors, processor platforms, and NVMe densities.

📄 Read the full white paper — benchmark methodology, test configuration, and complete results — here: https://zurl.co/eAAJT