On November 5, 2026, Supermicro announced a shared, persistent KV-cache storage solution with IBM and NVIDIA. IBM Redbooks’ reference paper, published June 5 and last updated August 20, already described the architecture and benchmark findings later cited in the announcement. Supermicro’s announcement and IBM’s technical paper describe the system and its test results.
What the announced system combines
The announced stack pairs Supermicro Petascale all-flash storage servers using AMD EPYC 9005 Series processors with IBM Storage Scale, NVIDIA Dynamo inference software, and NVIDIA Spectrum-X Ethernet. The idea is to keep reusable model context in shared storage, extending the cache beyond the memory available on a GPU.
A KV cache stores key-value states calculated while processing a prompt. Reusing those states can avoid recalculating the same context when it is needed again. In the IBM reference architecture’s G4-only test, Dynamo fetched cached data from shared storage to host RAM, then transferred it to GPU memory in chunks. The test disabled the G1 GPU-memory and G2 system-memory cache tiers.
What IBM’s benchmark tested
The benchmark used IBM Storage Scale Erasure Code Edition (ECE) v6.0.0.1, NVIDIA Dynamo v0.9.0+, and vLLM v0.14.1 with the openai/gpt-oss-120b model, pre-quantized to MXFP4. Eight Supermicro Petascale ASG-1115S-NE316R storage nodes used AMD EPYC 9535 processors; the GPU client was one Supermicro SYS-212GB-FNR node with four NVIDIA RTX PRO 6000 GPUs installed. The paper describes the model run as using one of those GPUs.
IBM reported results from separate single-request time-to-first-token (TTFT) and multi-request throughput tests. The table keeps their workloads and units distinct.
| Test | Workload and conditions | IBM-reported result |
| Single-request TTFT | Prompt length reached 130,000 tokens; the test used a 120-billion-parameter model. | Up to 56× speedup at 130,000 tokens. Cached TTFT ranged from 0.193 to 0.57 seconds across tested prompt sizes; without the cache, TTFT at 130,000 tokens was 32.14 seconds. |
| Throughput | 200 requests across 28 concurrent connections, selected from 100 prompts; the G4 cache started empty. | 0.19 requests per second (RPS) with GPU recomputation and 4.26 RPS with Storage Scale ECE. The cached run took 46.94 seconds. |
| Noisy-neighbor throughput | Four clients generated a reported 200 GB/s of competing network I/O during a 200-request cached run. | 3.6 RPS; the run took 55.56 seconds. |
The throughput figures describe one defined workload, not a general speedup for other models or deployments. In the noisy-network test, IBM’s displayed clean-network result of 4.26 RPS and noisy result of 3.6 RPS work out to about a 15.5% decrease. The paper describes that decrease as 18%, which does not match its displayed values.
When you compare cache results, keep the model, prompt length, software stack, workload, and cache state in view.
IBM’s reference paper came before the announcement
IBM Redbooks dates its reference paper to June 5, 2026, and lists August 20, 2026, as its last update. The paper contains the architecture and benchmark findings later cited in Supermicro’s November 5 announcement.
