Watad

System integrator & solution provider for #Cybersecurity, #AI, ML, Smart Cities, and Industrial Automation

[Watad @WatadDigital]

Long contexts consume GPU memory through KV cache, reducing concurrency even when model weights fit. Capacity planning must include sequence length distribution, not only parameter count.

[Watad @WatadDigital]

Prefill is compute-intensive; autoregressive decode is often constrained by memory bandwidth and KV-cache traffic. Benchmark them separately before choosing a serving topology.

[Watad @WatadDigital]

A cluster can report high GPU utilization while useful tokens per second fall. Compare end-to-end throughput, queue time and p95 latency under the actual request mix.

[Watad @WatadDigital]

Watad operates at system scale.

The systems that keep the Kingdom running are not single products. A desalination plant is intake, pretreatment, membranes, pumping, power, control, communications and security, built by different vendors over different decades, operated as one. A city is lighting, traffic, utilities, surveillance and emergency services that have to share data and still fail independently. An AI platform is compute, storage, fabric, facility and people.

Our work sits at the level where those parts meet: the integration, the boundaries, the commissioning and the operation of systems too large for any one vendor to own.

We are built for environments where failure matters, and in critical infrastructure the environment is the whole system.

#CriticalInfrastructure #SystemsIntegration #SmartCities #SaudiArabia #Watad

[Watad @WatadDigital]

Know where your plant actually stands against OTCC, with evidence from the cabinets.

Watad's OT security gap assessment combines three things most assessments skip: passive traffic capture on the control network, a physical walk of every cabinet and remote site, and interviews with the operators and vendors who actually connect. The findings are mapped to the National Cybersecurity Authority's OT controls and ranked by consequence to the process, not by severity score.

You receive an asset inventory built from evidence, a gap report an auditor can read, and a remediation plan sequenced around your turnaround calendar and your budget.

Two to four weeks per site, delivered for refineries, utilities and industrial plants across the Kingdom.

Need an OTCC readiness picture you can defend? Contact Watad to scope an OT security gap assessment.

#OTSecurity #NCA #OTCC #ICS #Watad

[Watad @WatadDigital]

Sovereignty is a physical address.

Every conversation about sovereign AI eventually becomes a conversation about buildings. Where are the weights stored. Where does the inference run. Where do the logs go. Who holds the keys to the storage encryption, and under which jurisdiction.

The Kingdom has made its direction clear through national programs, data regulation and investment in domestic compute. For an enterprise, the practical version is simpler: a platform inside the country, inside your own security boundary, operated by people you can call, on hardware you control the lifecycle of.

That does not exclude cloud. It defines which workloads may live there. The regulated core stays home.

Watad builds the home.

#SovereignAI #DataSovereignty #SaudiArabia #AIInfrastructure #Watad

[Watad @WatadDigital]

More security tools. Same blind spots.

A typical enterprise in the Kingdom runs twenty to forty security products. Each one sees its own slice. Few of them see the control network, the vendor laptop, the contractor's cloud account, the service account created in 2019, or the file share that holds the plant drawings.

The blind spots are not where the tools are weak. They are where nobody decided who owns the view.

Before recommending another product, we map what the organization can actually see: every asset class, every network, every identity type, and which tool, if any, covers it. The map usually produces a shorter shopping list and a longer list of configuration work on what is already paid for.

Coverage is a design, not a purchase.

#Cybersecurity #SecurityArchitecture #SOC #Watad

[Watad @WatadDigital]

A single rack now draws what a small data hall used to.

A conventional enterprise rack sat around 5 to 8 kilowatts. Dense air-cooled GPU racks pushed that to 30 or 40. Rack-scale liquid-cooled AI systems now sit above 120 kilowatts, and the roadmaps point higher.

The implication is physical. Power delivery to the rack moves from cords to busway. Cooling moves from air to water. Floor loading, fire suppression and even the door widths get reviewed.

It also means a small number of racks can carry an entire organization's AI capacity, which is good news for facilities with limited floor space and a sound electrical supply, a common profile in industrial sites across the Kingdom.

The room is the project. The rack is the delivery.

#DataCenter #AIInfrastructure #LiquidCooling #Watad

[Watad @WatadDigital]

سؤال واحد ممكن يفتح لك عالم كامل من المعرفة.

تبدأ بسؤال عن الفضاء، وتلقى نفسك تستكشف الفيزياء والتاريخ وأسرار الكون.

مع يحيى من نفيد، تقدر تسأل بالعربي، تتعمق في أي فكرة، وتكمل فضولك بسؤال جديد.

كل معرفة تبدأ بسؤال.

#يحيى #نفيد #الذكاء_الاصطناعي

[Watad @WatadDigital]

InfiniBand or Ethernet is really a question about your team. InfiniBand still leads at the largest scale. Purpose built AI Ethernet is defensible for enterprise inference and mid scale training. #Networking #Watad

[Watad @WatadDigital]

المصنع في النهاية عبارة عن آلة ضحمة مصنوعة من آلاف الآلات، كل واحدة تحرك سائلًا أو غازًا أو مادة ضمن حدود يضبطها متحكم يقرأ حساسًا مرات كثيرة في الثانية.

حين تنحرف الأجهزة، أو يتقادم المنطق، أو تتعثر الشبكة، يستمر المصنع في العمل، أسوأ قليلًا فقط كل ساعة، لسنوات!

الأتمتة الصناعية في وتد هي عمل الحفاظ على تلك الدقة...

أنظمة تحكم مهندَسة ومُشغَّلة كما ينبغي، أجهزة تقرأ بصدق، شبكات حتمية، وتحديث يتم دون إيقاف الإنتاج.

#IndustrialAutomation #SCADA #ProcessControl #Watad

[Watad @WatadDigital]

Design the platform before the purchase order.

Most AI infrastructure projects in the Kingdom start with a server quote and work backwards. The quote fixes the GPU count, and every other decision, storage, fabric, cooling, power and operations, gets squeezed into whatever the quote assumed.

Watad runs a two day AI architecture workshop with your infrastructure, data, security and facility teams in one room. We size the platform from your actual workloads, survey the room you intend to use, and produce a design, a bill of materials and a commissioning plan you can put in front of procurement.

If you are planning an AI platform in the next twelve months, this is the cheapest two days of the project.

Message us to book the workshop.

#AIInfrastructure #Architecture #SaudiArabia #Watad

[Watad @WatadDigital]

Storage can starve compute.

A GPU that waits for data costs exactly as much as a GPU that computes. The difference shows up only in the training curve and the invoice.

A training job reads its dataset again every epoch, writes checkpoints measured in terabytes, and reloads them after every interruption. Eight GPUs can easily demand ten to twenty gigabytes per second of sustained reads, with write bursts far above that. A conventional NAS built for file shares delivers a fraction of this and collapses under concurrent clients.

So we size storage from the GPU side: throughput per GPU, metadata operations per second, checkpoint write time as a percentage of step time.

The cluster is only as fast as its slowest feed. Usually that feed is storage.

#AIStorage #AIInfrastructure #HPC #Watad

[Watad @WatadDigital]

Not every connection is visible.

The network diagram shows the plant isolated from the enterprise. Then the walk-through finds a cellular modem a vendor left behind for remote diagnostics, a laptop that bridges the office and the control network for convenience, a historian with a second interface nobody documented, and a wireless access point installed during the last turnaround.

None of these were malicious. All of them are paths.

The first phase of every OT security program we run is discovery: passive traffic capture on the control network, a physical walk of every cabinet, and a reconciliation of what exists against what the diagram claims. The gap is always larger than the diagram's owner expects.

You cannot protect a connection you do not know you have.

مو كل اتصال ظاهر.

الرسم يُظهر المصنع معزولًا عن الشبكة المؤسسية. ثم تجد الجولة الميدانية مودم خلوي تركه مورّد للتشخيص عن بُعد، وحاسوبًا محمولًا يربط المكتب بشبكة التحكم للراحة، وسجلًا بواجهة ثانية لم يوثقها أحد، ونقطة وصول لاسلكية رُكبت في آخر صيانة شاملة.

لا شيء منها خبيث. وكلها مسارات.

المرحلة الأولى من كل برنامج أمن تشغيل نديره هي الاكتشاف: التقاط سلبي للحركة على شبكة التحكم، وجولة مادية على كل خزانة، ومطابقة ما هو موجود مع ما يدّعيه الرسم. الفجوة دائمًا أكبر مما يتوقع صاحب الرسم.

ما لا تراه، لا تستطيع حمايته.

#OTSecurity #ICS #CriticalInfrastructure #Cybersecurity #Watad

[Watad @WatadDigital]

Your GPUs are waiting.

Most enterprise clusters we walk into run somewhere between 30 and 50 percent utilization. The models are not too small. The data cannot reach the accelerators fast enough, the scheduler leaves gaps between jobs, and the network stalls on the collective operations that training depends on.

The GPU was the expensive part of the purchase order. It is rarely the part that is actually busy.

Before approving the next rack, measure the current one. Storage throughput per GPU, time to first batch on the data loader, queue wait times and link error counters tell you exactly where the money is sitting idle.

Watad designs AI infrastructure as one system: compute, storage, fabric, cooling and operations, so the accelerators spend their time computing.

#AIInfrastructure #HPC #DataCenter #SaudiArabia #Watad

[Watad @WatadDigital]

A GPU purchase is a component decision. An AI platform is an architecture decision.

Teams often start with accelerator specifications because they are visible and easy to compare. The harder work sits around workload shape, data movement, network topology, storage behavior, power density and the operating model.

If those layers are designed independently, expensive compute spends time waiting. Start with the workload, then work backward into the infrastructure.

The useful question is whether the whole system can keep expensive compute productive.

[Watad @WatadDigital]

The rack is where AI runs. The platform is everything that keeps it useful.

A rack can arrive on time, power on correctly and still fail to deliver the expected business outcome. Production AI depends on orchestration, data pipelines, observability, storage, networking and support.

Acceptance should test the full path from data to model to user under realistic load, not only whether the equipment powers on.

Architecture should remove waiting before procurement adds more capacity.

[Navid | نفيد @Navid_co]

The next limit on AI compute may not be the GPU. It may be heat.

As rack densities rise, cooling stops being a supporting system.

It becomes part of the compute architecture.

More: https://t.co/PupFaaQL7f

[Navid | نفيد @Navid_co]

Owning compute is one thing. Monetizing it well is another.

More: https://t.co/KQjYlDsvBY