# GCP instance:
176 vCPU (88 Cores - AVX512) y 1.4TB de RAM... ~4.64$/hour (spot)

## Mainnet code

| Size | Pre-processing [s] | Batch proof [s] | memory [Gb] | $proof |
|:--------:| :------------------: | :---------------: | :-----------: |:-----------: |
| $$2^{23}$$ | 61 | 104 | 490 |0.134 |

Version being executed in mainet

## Optimized code with new intermediate polynomials

| Size | Pre-proc [s] | Batch proof [s] | Mem [Gb] | $proof | #cells | MCells/segon/core | $ cell |
|:--------: |:------------------: |:---------------:|:-----------:|:-------:|:------:|:----------------------:|:------:|
| $$2^{23}$$ | **26** | **81.5** | **295** | **0.105** | $$ 984 \cdot 2^{23}$$ | 1.15 | $1.27 \cdot 10^{-11}$ |
| $$2^{24}$$ | **41** | **151** | **580** | **0.195** | $$ 984 \cdot 2^{24}$$ | 1.24 | $1.18 \cdot 10^{-11}$ |
| $$2^{25}$$ | **69.5** | **299** | **1130** | **0.385** | $$984 \cdot 2^{25}$$ | 1.26 | $1.17 \cdot 10^{-11}$ |

**Note:** #cells = (#constPols + #commitPols)*N
**Hardware used**: 176 vCPU (88 Cores - AVX512) y 1.4TB de RAM… ~4.64$/hour (spot)

-------------------------------------
# GCP production instance

n2d-standard-224 - 224 vCPU y 896GB de RAM… ~2.78$/hour (spot)

## Mainnet code

| Size | Pre-processing [s] | Batch proof [s] | memory [Gb] | $proof |
|:--------:| :------------------: | :---------------: | :-----------: |:-----------: |
| $$2^{23}$$ | **90** | **122** | **507** |**0.094** |

## New prover without reduced memory

| Size | Pre-processing [s] | Batch proof [s] | memory [Gb] | $proof |
|:--------:| :------------------: | :---------------: | :-----------: | :-----------: |
| $$2^{23}$$ | **28** | **100** | **295** | **0.077** |
| $$2^{24}$$ | **43.5** | **200** | **580** | **0.154**|
| $$2^{25}$$ | **-** | **-** | **-** | **-** |

## New prover with reduced memory

| Size | Pre-proc [s] | Batch proof [s] | Mem [Gb] | $proof | #cells |Mcells/second/core | $ cell |
|:--------: |:------------------: |:---------------:|:-----------:|:-------:|:------:|:----------------------:|:------:|
| $$2^{23}$$ | **28** | **122** | **221**| **0.094**| $$984 \cdot 2^{23}$$ | 0.60 | $1.14 \cdot 10^{-11}$ |
| $$2^{24}$$ | **43.5** | **240** | **430**| **0.185**| $$984 \cdot 2^{24}$$ | 0.61 | $1.12 \cdot 10^{-11}$ |
| $$2^{25}$$ | **82** | **515** | **850**| **0.398**| $$984 \cdot 2^{25}$$ | 0.57 | $1.21 \cdot 10^{-11}$ |

**Note:** #cells = (#constPols + #commitPols)*N
**latency 1M cells per core** is: core seconds used (=batch prof time $\times$ #cores) divided by #cells/1M
**Hardware used**: n2d-standard-224 - 224 vCPU y 896GB de RAM… ~2.78$/hour (spot)

## Improvements
* **57% pre-processing latency:**
* Load data from files in parallel
* Mekelization of constants done at runtime instead of loading from file. This is slower than using the optimized files load, however ends up saving a lot of data-tranfer times
* **22% batch-proof latency:**
* Activation of AVX512
* Optimized intermediate polynomials

* **26% memory requirements:**
* Optimize utilization of memory buffer along the proof, avoid data transpositions for temporal polynomials

* **400% increase of steps capacity** ($N=2^{25}$)