
2 meetings sous NDA avec 2 clients différents :
- CA-GIP le 19/11/2024
- BPCE le 20/11/2024
### General
Nom officiel devrait être z17 (M/T 9175 ME1)

3 main topics :
- AI
- Security
- Automation for Operational Efficiency

### AI

1 AIU per chip / 4 AUI per machine in z16
z17 will allow for processors to use AIU in a different chip.
8 Spyre cards per set, 1 set minimum if configured. 
Each Spyre cards ships 32 AIU procs, total 256 AIU per set of cards.
Full config ships 1536 AIU cores.

Youtube : IBM Talia Gherson (IBM Research), explains Spyre before it was called Spyre

We are on the road to get (hybrid - Traditional + GenAI) AI on the platform. The idea is to be able to run GenAI directy on the platform.

Fabrice demande si l'exploitation peut se faire directement (nativement) dans z/OS ou si LoZ est un pré-requis : l'exploitation "native" se fera au travers de z/CX.

Spyre cards : François demande comment se fait le distinguo entre TraditionalAI and GenAI ?
Answer : the goal is it will be automatic, but we're not there yet.
Subject needs to be revisited closer to GA or shortly after.
For now, exploitation through PyTorch or TensorFlow will be through LoZ or z/CX. 
TBC for z/OSCP.

Tous les zSW qui sont AI-enabled pourront utiliser les ressources IA dispo dans la machine, AIU ou Spyre : 
- zSCC
- TDz
- SDTz
- ......
### Automation 

Market demand : 
Opentelemetry
Infra as Code, management through Ansible and Terraform rather than plain HCD. 
	Might explain the drastic increase in HSA in z17 (270%)

Infra as Code : ZIM appliance ? TBC with Carl and Walter before to the call.
Does it have anything to do with DPM ?
Matthias to follow-up.
### Processor : 

5nm, 5.5 GHz
8 cores + IO DPU / chip
	8 L1 caches
	10 L2 caches
16 cores per socket (a socket = a DCM)
1 DCM = 2 chips
4 DCM per drawer
1 to 4 drawers : 
- Max43 (1 drawer)
- Max90 (2 drawers)
- Max136 (3 drawers)
- Max183 (4 drawers) - Factory Build only
- Max208 (4 drawers) - Factory Build only
### Memory 
16TB mem / drawer (except 2nd drawer = 15TB)
More cache (40% more per core) : 
- 36MB L2
- 360MB vL3
- 2.8GB vL4
Memory encryption déjà dans le z16, sans que ce soit paramétrable (c'est là) : was AES128, now AES256.
Memory encrypted, including L4, L3, L2 caches. 
Long term goal is to be able to run everything encrypted with FHE (Fully Homomorphic Encryption), when it reached a point where overhead is acceptable.
### Refactored IO:
2x port capacity
2nd GenAI accelerator (8x dedicated AI core per chip)
z16 : "Kraken" processor (qui se trouve ds les cartes IO, qui fait la traduction FICON => FC) a disparu en zNext.
Facilitera les améliorations sur les protocoles de transport, qui seront développés en même temps que le processor.

La disparition du Kraken a libéré de la place ds la carte ce qui a permis de rajouter des ports => Next Gen FICON seront 4 ports.
##### DPU

I/O adapter management change : 
- Design to move processing into firmware
- Supports new I/O adapter innovations
- Improved RAS
- 
One DPU per CP
Support:
	- FC / FCP / zHPF
	- OSA (EQDIO)
	- OSC

Disparition des cartes RoCE en tant que telles, convergence OSA / RoCE, CHPID type OSH.
Action : Description de l'existant CA-GIP, architecture et IO definition, et Matthias vérifiera avec le lab. 
Il est nécessaire de connaitre / comprendre les changements nécessaires pour migrer de l'infra actuelle (OSA + RoCE) vers la nouvelle (carte convergée).
Supportées en z/OS 2.5 uniquement !
Si z/OS 2.4, on est limité aux cartes OSA Express "traditionnelles"

Les cartes Network Express seront supportées sur z17 pour apporter la convergence RoCE/OSA.
Les cartes OSA Express seront supportées pour certains cas d'usage, dont les cartes OSC (les consoles OSA-ICC par exemple).

Pas de chgt sur la capacité IO globale : 12 IO drawers * 16 cartes = 192 cartes.
### Infrastructure
Amélioration de la communication inter-drawers, d'un facteur à confirmer (ds les proof points, Feb / Mar timeframe).
Francois demande si les TKE sont un requirement pour zNext : Non, mais dépend de "local regulations".

Capacité : 2478 MIPS uniproc (about +10%) **ATTENTION** ce sont des chiffres calculés, benchmarks to come Feb / March
Freq: 5.5GHz (+5%)
Full cap : 246062 MIPS (+14%), Max208 model.

Changement du design électrique : Power Dev lab has shared that z16 design was conservative because it was a new design, and lab was reluctant to keep usage going at high usage rate / consumption.
Matthias proposed to organize a call with Power Lab to take client through the details.

Quiesce : Mathhias has not heard about changes, but proposed to follow-up.

### Future considerations : 

Make sure Z frame remains free of I/O drawers, to allow easier upgrades by :
- either replacing the Z frame entirely without touching the I/Os
- or plugging in a next generation CPC drawer before removing one for older generation.

Moving down from 4 to 3 drawers (while keeping the numbers of cores similar) might be an idea to accomodate the growing need for large zIIPs configurations in a given LPAR. This would help keep the LPAR in a single drawer.

### Pricing

Workload Pricing Classification : Long term plan is to be able to classify workloads based on what kind of executable they are using (Omegamon, Db2, DFsms.....) and apply different price points. At some point in the future, intent is to remove the need for zIIP.
**ATTENTION** : removing zIIP means removing SMT2 capability, which can be a problem with certain workloads.

La classification sera faite au niveau du HW, requiert une recompilation du binaire (donc c'est à la main de l'éditeur) pour inclure le tag et le certificat nécessaire pour tracker l'exécution.

Pre-reqs : TFP for HW