Skip to main content

DPIIT-recognised startup · Hyderabad, India

Compressing the memory that limits AI inference.

Syncoril RSIA develops STRIDE, a patent-pending, licensable compression technology that cuts AI inference memory by 70–80%, so the same GPUs can serve 3.8× to 4.9× more conversations, with quality within 2.5% of the uncompressed model.

Get in Touch Explore STRIDE
  • Provisional patent filed, India
  • Validated on 4 model families
  • Licensable IP with reference implementation

KV-cache per active conversation

Illustrative · 7B-class model · 32,000-token context

Uncompressed (FP16) ≈ 4.2 GB
With STRIDE ≈ 0.9–1.1 GB

70–80%

smaller KV-cache footprint

< 2.5%

quality deviation, under 1% at standard

3.8×–4.9×

Measured compression: two recommended configurations, standard and high-density

70–80%

Reduction in KV-cache memory footprint, so far more conversations fit per GPU

< 2.5%

Perplexity deviation from uncompressed across both configurations (under 1% at standard)

4 families

Model families validated, unmodified, with 32 of 32 gated tests passed

About Syncoril RSIA

Deep-tech IP built for the economics of AI infrastructure.

Syncoril RSIA is a deep-technology startup based in Hyderabad, India, recognised by the Department for Promotion of Industry and Internal Trade (DPIIT), Government of India. We create and license foundational technology that makes AI infrastructure more efficient.

Our flagship, STRIDE, targets one of the least-addressed levers on inference cost: the memory consumed by the KV-cache. We pair rigorous multi-model validation with a protected IP position, so infrastructure operators and chip designers can evaluate and adopt it with confidence.

Our mission

To make large-scale AI inference materially cheaper and more scalable by reclaiming wasted memory, without compromising model quality.

Rigorous by default

Every claim rests on a gated test battery across multiple model families.

Protected by design

Patent-pending IP, with deep technical detail shared only under NDA.

Built for adoption

Drop-in integration that stays under the licensee’s control.

The problem

Memory, not compute, is the bottleneck in AI inference.

While a language model generates text, it keeps a running record of the conversation in its KV-cache, so earlier tokens need not be recomputed. That cache grows with every token, for every concurrent user, and it must live in scarce, expensive accelerator memory.

128 KB / token

KV-cache cost for a widely used 7B-parameter model.

4.2 GB+

Per active conversation at a 32,000-token context.

Bandwidth-bound

Cache reads often gate generation speed, not raw compute.

KV-cache size vs. context length

Illustrative · per active conversation · 7B-class model

Uncompressed (FP16) With STRIDE (3.8×–4.9×)
0 1 GB 2 GB 3 GB 4 GB 0 8K 16K 24K 32K Context length (tokens) Memory reclaimed ≈ 4.2 GB ≈ 0.9–1.1 GB
  • 1

    It grows without bound. Cache size scales linearly as conversations and context windows get longer.

  • 2

    It is multiplied by every user. The cache is held for each concurrent conversation a serving system supports.

  • 3

    Shrinking it pays directly. Reducing KV-cache is one of the most direct levers on AI infrastructure cost.

Solutions

STRIDE: adaptive compression for AI inference memory.

A software-first compression layer for the KV-cache, designed to fit the infrastructure operators already run, with a longer-term path to hardware.

KV-Cache Compression Layer

A drop-in compression method for existing GPU-based inference infrastructure. Delivers a measured 3.8× to 4.9× compression depending on the configuration you choose, with no hardware changes required.

Software-firstDrop-in

Licensable IP & Reference Implementation

We license the compression logic, with a reference implementation, for your engineers to integrate and operate inside your own stack: under your control, not a hosted dependency.

Patent pendingLicensee-controlled

Hardware Acceleration Path

A longer-horizon option: a synthesizable proof-of-concept has been designed and simulated, demonstrating the feasibility of an embeddable hardware IP block for accelerator designers.

Proof-of-concept stageFuture upside

Two recommended configurations

One layer, two settings. Choose the trade-off that suits your workload.

Both are validated on all four model families and recommended for production. They differ in how much memory they reclaim and how closely they track the original model; in both, measured quality stays close to the original model.

Maximum fidelity

Standard

3.8×

measured compression

Cache memory freed
about 74%
Perplexity vs. uncompressed
under 1% (max 0.76%)
Concurrent conversations per GPU
about 3.8×
Validated on
all four model families

Best when preserving the very last fraction of fidelity matters most.

Maximum value

High-density

up to 4.9×

measured compression

Cache memory freed
about 80%
Perplexity vs. uncompressed
under 2.5%
Concurrent conversations per GPU
about 4.9×
Validated on
all four model families

Best when capacity and cost matter most: more compression, with quality that still closely tracks the original.

A drop-in layer, not a rebuild.

STRIDE sits at a single, well-defined integration point in the inference pipeline. No architectural rework required.

  1. IncomingRequests
  2. ModelForward pass
  3. KV-cacheGrows with context
  4. STRIDECompression layer
  5. OutputCompressed cache in memory
  • More concurrent users
  • Higher throughput
  • Lower cost per token
  • Model-agnostic by validation, not just by design.
  • Composable with the quantization a stack already runs.
  • No hardware changes needed on existing GPU infrastructure.

Business value

What a smaller memory footprint is worth.

AI answers are produced on GPUs, and GPU memory is the scarce, expensive part. STRIDE shrinks the memory each conversation needs. The result is simple: the same GPUs serve many more conversations, or fewer GPUs serve the same conversations.

  • More customers, same hardware

    Imagine a warehouse where every box suddenly shrinks to a quarter of its size: the same shelves hold about four times as many. STRIDE does this for AI memory, so each GPU serves 3.8× to 4.9× more conversations at once.

  • A smaller GPU bill

    If memory is what limits a workload, the same demand needs up to about 80% fewer GPUs. That is hardware you do not have to rent, buy, power or cool.

  • Longer conversations, no bigger machines

    The longer a conversation, the more memory it uses. Compressing that memory lets long documents and extended chats fit on hardware you already run.

  • Quality you can keep

    Answers stay within 1% (standard) or 2.5% (high-density) of the original model on the standard quality measure, tested across four model families.

Estimate the savings for your fleet

Illustrative. Move the controls to match your own situation; every figure is explained below.

STRIDE configuration
1,000

From 8 to 10,000 GPUs.

$1.91
50%

Only GPUs whose capacity is held back by cache memory benefit. Be conservative: 50% is a prudent starting point.

Illustrative annual saving

$6.2Mper year

That is about 368 GPUs you would no longer need, or 36.8% of your total GPU bill.

Annual GPU bill today$16.7M
Annual GPU bill with STRIDE$10.6M
GPUs needed for the same work
1,000 → 632
Or keep the GPUs and serve
about 3.8× the conversations

In plain English: if half of your GPUs are held back by memory, shrinking memory use by 3.8× lets that half do the same work with roughly a quarter of the hardware.

Worked scenarios

Annual saving at the calculator’s starting assumptions: A100 80GB at $1.91 per GPU-hour, always on, with 50% of GPUs memory-limited.

FleetBill todaySaving, standardSaving, high-density
100 GPUs$1.7M$616K$666K
1,000 GPUs$16.7M$6.2M$6.7M
10,000 GPUs$167.3M$61.6M$66.6M

How the numbers are built

  1. 1Capacity scales with compression. In testing on an NVIDIA A100 80GB, the standard configuration held 3.80× to 3.90× more simultaneous 5,031-token conversations. The calculator uses 3.8× and 4.9×, rounded down.
  2. 2Only memory-limited GPUs benefit. GPUs bound by raw compute do not change, so savings apply to the share you choose.
  3. 3The arithmetic. Saving = GPUs × memory-limited share × (1 − 1 ÷ compression) × price per GPU-hour × 8,760 hours.
  4. 4Gross, not net. Figures are the operator’s saving before any licence fee, and are an illustration rather than a quote, forecast or guarantee. Real results depend on workload, context length, batch sizes and hardware.

Price sources (public on-demand list prices, USD, November 2026)

GPU prices move frequently and vary widely by provider and commitment. Compression and capacity figures come from Syncoril RSIA’s validation report (September 2026).

Validated performance

Validated, not just designed.

STRIDE was tested unmodified across four independent model families and multiple attention architectures, using a gated nine-test validation battery with pre-declared acceptance criteria.

32 / 32

gated test results passed · 36 executions · zero exceptions

Models
Four open-weight model families
Evaluation text
WikiText-103 validation split, 10 held-out passages per model
Hardware
NVIDIA A100 80 GB, bfloat16
Method
Paired comparison on identical text, pre-declared ±5% tolerance

Both of STRIDE’s recommended configurations are shown: Standard (3.8×) and High-density (up to 4.9×). Production-readiness tests below were run at the standard configuration.

More conversations on the same GPU

Simultaneous 5,031-token conversations · one NVIDIA A100 80 GB

Uncompressed Standard High-density

Mistral-7B

90
342 · 3.80×
≈445 · ≈4.9×

Llama-3.1-8B

87
333 · 3.83×
≈435 · ≈4.9×

Qwen2.5-7B

203
772 · 3.80×
≈1,000 · ≈4.9×

Gemma-2-9B

31
121 · 3.90×
≈155 · ≈4.9×

Capacity = 80 GB, minus the model’s own weights, minus a flat 10 GB reserved for activations and runtime overhead: a deliberately conservative buffer, so figures are a floor. Standard figures are measured; high-density figures use the same method on the measured cache sizes and are rounded down.

Quality, measured against uncompressed

Perplexity deviation on identical text · both configurations

Standard High-density

  • Mistral-7B

    Standard−0.76%
    High-density+2.12%

  • Llama-3.1-8B

    Standard+0.16%
    High-density+0.28%

  • Qwen2.5-7B

    Standard+0.75%
    High-density−0.39%

  • Gemma-2-9B

    Standard+0.06%
    High-density+0.96%
−5%0+5%

Shaded band: ±1% · full width: ±5%

Largest deviation at the flagship setting: 0.76% (standard) and 2.12% (high-density). The ±5% tolerance was pre-declared for the standard test; high-density results are shown on the same scale.

  • 0.40–0.55 ms

    Decode latency

    Worst case per model, including the full cost of compression. Sub-millisecond on every model tested.

  • 150 steps

    Stable live generation

    Continuous real-time recompression with no compounding error: peak loss within 1.04× of uncompressed.

  • Zero spillover

    Outlier containment

    A deliberately injected extreme outlier (100× normal scale) did not affect the rest of the cache, on any model.

  • Zero leakage

    Correct batched serving

    Batched results matched sequential exactly, with no cross-sequence leakage between concurrent requests.

Figures are drawn from Syncoril RSIA’s own validation report (September 2026) under the test conditions above. Results depend on model, workload and configuration. The compression method is proprietary; the full validation report and technical detail are available under mutual NDA.

Licensing & partners

Licensing built around your own stack.

STRIDE is licensed as logic, not rented as a service: a reference implementation your engineering team integrates, tests and operates under your control. Commercial terms are structured around the value realised.

  1. 1

    Technical review under NDA

    Full documentation and the complete validation report, shared under a mutual non-disclosure agreement.

  2. 2

    Pilot integration

    Reference implementation provided, with testing carried out on your side against your own workloads.

  3. 3

    Licence & deployment

    A licence for the IP and reference implementation, integrated into your production stack.

Who we work with

Cloud AI inference providers

Serve more users per accelerator dollar.

Accelerator & chip designers

A licensable compression IP block for future silicon.

Long-context & agentic platforms

Where the memory advantage grows largest.

Edge & on-device AI

Relaxes tight on-device memory constraints.

Also: large-scale enterprise inference platforms evaluating ways to lower memory and serving costs.

Intellectual property & trust

Protected IP. Transparent process.

We protect what makes STRIDE valuable, and make it straightforward for serious partners to evaluate it.

Patent pending

A provisional patent has been filed with the Indian Patent Office, covering both the software method and the hardware embodiment, and establishing a priority date.

NDA-gated technical deep-dive

Full documentation, the complete validation report and the patent specification are available to qualified parties under a mutual non-disclosure agreement.

DPIIT-recognised startup

Recognised by the Department for Promotion of Industry and Internal Trade, Ministry of Commerce & Industry, Government of India (Certificate No. DIPP237142, issued 3 January 2026).

Roadmap

From ideation to patent-pending, and now to market.

  1. Ideation

    KV-cache identified as a high-leverage, underexploited lever on AI infrastructure cost.

  2. Method developed

    Initial software method designed and validated on a single model.

  3. Multi-model validation

    Four model families, nine-test battery, 32 of 32 gated tests passed.

  4. Hardware proof-of-concept

    Synthesizable proof-of-concept designed and simulated.

  5. Patent filed

    Provisional filed with the Indian Patent Office; priority date secured.

  6. Go-to-market We are here

    Pilot licensing engagements and commercialisation ahead.

Contact

Let’s talk.

For licensing and IP partnerships, pilot integrations, NDA-gated technical reviews, or to support Syncoril RSIA as a mentor or programme partner, write to us. We read every message and reply by email.

Email contact@syncorilrsia.com

Start the conversation

Licensing & IP partnership Pilot integration NDA technical review Mentorship, programme or ecosystem partnership
Location
Hyderabad, Telangana, India
Registered entity
Syncoril RSIA (OPC) Private Limited
CIN U62099TS2025OPC201105
Technical deep-dives
Shared only under a mutual NDA. Mention this in your email and we will start there.