← Back to Signal
Technical PositionResearch ReviewsPosition on TurboQuant

Orthogonal by architecture

Why TurboQuant and PrizmalRun solve different problems at different layers of the inference stack.

Audry Larocque and Jonathan Pilault · April 5, 2026 · 7 min

Key takeaways
  1. 01TurboQuant compresses the KV cache, which is memory. PrizmalRun controls the forward pass, which is compute.
  2. 02TurboQuant has no mechanism to avoid full-depth execution per token. PrizmalRun is that mechanism.
  3. 03The two systems stack: KV memory savings compound with per-token compute savings.
  4. 04TurboQuant's ceiling is the size of the KV cache. PrizmalRun's ceiling is the cost of the forward pass, an order of magnitude larger.
  5. 05Nothing in TurboQuant overlaps with Prizmal's patent position or GTM. The orthogonality is structural.

01 · What it does

A KV cache compression algorithm.

TurboQuant is a KV cache compression algorithm published by Google. It applies a random orthogonal rotation to stored key-value vectors, exploits the resulting Beta-distributed coordinates to fit a precomputed Lloyd-Max scalar quantizer, and uses a 1-bit residual correction stage. Output: 6x memory reduction, 8x attention logit speedup on H100, zero calibration, zero fine-tuning.

Its operating surface is what is stored between forward passes. It does not read model signals during inference. It does not modify the compute graph. It has no mechanism to route compute based on token difficulty. It is precisely a data representation scheme over a memory buffer.

02 · The stack

Two layers. Two systems.

Model · LLM weights

Neither system modifies them.

Prizmal · PrizmalRun

Reads attention entropy, logit margin variance, and hidden state norms per token. Modulates compute execution in real time: layer skipping, attention head scaling, FP16 and INT8 precision, sparsity injection.

TurboQuant · KV cache layer

Compresses stored key-value vectors to 3-bit representations. Acts on data in memory. Does not observe or influence the forward pass compute graph.

Engine · Qualcomm, Intel, AMD, NVIDIA

Silicon execution substrate.

A transformer running TurboQuant still executes every layer, every attention head, at full depth, for every token. TurboQuant has no mechanism to avoid that waste. PrizmalRun is that mechanism.

03 · Comparison

Memory representation vs runtime control.

DimensionTurboQuantPrizmalRun
Acts onStored KV vectors, data in memoryCompute graph, forward pass execution
TimingAt storage, before attention scoresReal time, token by token, mid-inference
Signal usedNone, data-oblivious by designAttention entropy, logit variance, hidden norms
OptimizesDRAM footprint, bandwidthFLOPs and R/E, reasoning quality per joule
Compute allocationUnchanged, every layer runs fullyDynamic, hard tokens get full depth and easy tokens skip
Quality proofNear-optimal distortion bound99% reasoning retention at 50% sparsity, +251% R/E vs dense
ComposableYes, stackable with PrizmalRunYes, compatible with TurboQuant

04 · Investor implications

Composable, not competitive.

Differentiation intact

TurboQuant does not enter PrizmalRun's operating domain. It cannot adapt to token difficulty. It cannot route compute.

R/E thesis strengthened

TurboQuant reduces memory bandwidth cost. It does not improve R/E. Reasoning yield per joule is a compute allocation problem, not a compression problem.

IP portfolio, no overlap

Prizmal's provisional patents cover guided blockwise structured pruning, disagreement-aware adaptive pruning, layer-sensitivity pruning caps, and oracle-guided routing policy training. Zero overlap with vector quantization prior art.

Composability additive

A joint deployment addresses both memory-side and compute-side efficiency simultaneously. Additive positioning, not conflict.

The structural ceiling TurboQuant cannot cross

KV cache compression is near the ceiling for its problem class. Beyond that point, further compression destroys effective information nonlinearly. The compute allocation problem has no such ceiling. Every token is different, every context is different.

05 · Bottom line

The runtime layer is unsolved.

TurboQuant is a strong result in KV cache memory compression. It is mathematically rigorous, calibration-free, and close to the theoretical optimum for its problem class. It also has a hard boundary: it cannot observe the forward pass, cannot read token difficulty, and cannot modulate compute. Its ceiling is physical. PrizmalRun operates above that ceiling entirely, reading live model signals and modulating compute in real time, token by token.

R/E vs dense
+251%

Reasoning quality per joule.

Reasoning kept
99%

At 50% sparsity.

Compute
-60%

FLOPs reduction at inference time.

FAQ

Does TurboQuant compete with PrizmalRun?
No. TurboQuant compresses KV cache memory between forward passes. PrizmalRun modulates compute inside the forward pass. They operate on different axes of the inference stack and can be deployed together.
Could Google extend TurboQuant to do what PrizmalRun does?
Not without rebuilding it. TurboQuant is a data representation scheme over a memory buffer. Reading live model signals and routing compute per token requires a different architecture.
What does TurboQuant validate for Prizmal?
That the largest AI labs now treat inference efficiency as a first-order research problem. TurboQuant addresses the memory layer. PrizmalRun addresses the compute layer. Both will be required for production inference at scale.

Run the numbers on your own traffic

The R/E Audit measures reasoning per unit of energy on your real workload, read only and reversible.