Hard AI infra interview questions

Three things change at once: dense compute more than doubles with an fp4 tensor core, HBM3e reaches 8 TB/s, and NVLink stops at 72 GPUs instead of 8. Work through what each does to tensor-parallel degree, per-step weight reads and expert placement for a 671B-parameter mixture of experts.

The roofline, the sharding arithmetic and the parallelism trade-offs transfer unchanged, and the numbers are in the same units. What changes is who writes the kernels, how shapes must behave, and the interconnect topology you shard against. The v6e and v7 numbers worked through against an H100.

An N × N grid of multiply-accumulate cells holds a weight tile still and pumps activations through it, one row per cycle, with no register file or instruction fetch in the inner loop. The cycle count, the fill and drain, and the formula that says why a 128-row tile wastes two thirds of the array.

A Trn2 instance carries 16 Trainium2 chips with 1.5 TB of HBM and more dense bf16 FLOPS than an 8-GPU H100 node. Whether the cheaper FLOPS reach your workload depends on the Neuron compiler, the kernels you do not have yet, and an MFU number you must measure rather than assume.

192 GB and 5.3 TB/s per card beat the H100 on both memory numbers at a lower hourly price. Whether that wins in production is a question about achieved bandwidth and ROCm kernel coverage, and there is a short list of numbers to run before you sign the order.

Put the weights in on-chip SRAM and the HBM wall disappears: tens of terabytes per second per chip, and a batch-1 step limited by the pipeline rather than the memory. The price is capacity: a 70B needs hundreds of chips per replica, and the cost per token depends on keeping every one of them busy.

Four bits give eight magnitudes per sign, and the largest step between them is 50%. Block scaling rescues most of the tensor; a single outlier in a block of sixteen flushes its neighbors to zero. Where the precision goes, which layers fail first, and the measurements that separate acceptable from broken.

512 GPUs at 400 Gb/s each is 205 Tb/s of injection bandwidth a two-tier fabric must carry with no oversubscribed link. The port arithmetic that gives 16 leaves and 8 spines, the rail wiring that keeps NCCL traffic one hop away, and the line where 900 GB/s of NVLink becomes 50 GB/s of InfiniBand.

Four percentages from a profiler summary are enough to place a kernel on the roofline, estimate its arithmetic intensity, and rule out half the optimizations people reach for. The reading, the arithmetic that backs it, and the fix list in the order it pays.

An H100 has 50 MB of L2, split into two partitions, with a bandwidth several times HBM's. That is enough to change the roofline for a kernel whose working set fits and to do nothing at all for the ones that stream. The arithmetic of what fits, the residency controls, and the two ways the L2 lies to a profile.

A two-socket GPU node has two PCIe trees, two memory pools and one inter-socket link that every wrong-sided transfer must cross. A dataloader on the far socket shares that link with seven other GPUs' traffic, and its copies run at half speed. The topology map, the arithmetic and the pinning.

Decode throughput is bytes per step over bandwidth, price is dollars per hour, so tokens per dollar for a 70B falls out of a spec sheet in ten lines. In bf16 the three generations are within 10% per dollar; fp8 pulls the H100 ahead, and fp4 on the B200 doubles it again. The chain, the table and the caveats.

Every watt in comes out as heat, so the design starts from the load and one equation. The flow rate, the CDU count and redundancy, the water quality requirements that decide whether the loop survives three years, and the automatic response that has to fire before a human can read the page.

A rail-optimized fabric makes this fault findable because the eight rails are identical by construction, so any asymmetry is a defect. Bisecting by rail, then by switch, then by node, then by port narrows 512 links to one in four steps, and the counters say which of three physical causes it is.

Acceptance is a contract document, not a checklist, so every criterion has to be a number with a tolerance that a vendor can be held to. Comparing nodes against each other rather than against a datasheet is what finds real defects, and the duration is what finds the thermal ones.

A training cluster is a synchronized load, so the peak is far above the average and the average is what everyone budgeted. The immediate mitigation takes minutes and costs a few percent of throughput, and the four durable fixes trade cost against throughput in different places.

Single-node fine and two-node halved is a specific signature, and it points at one thing before anything else. Why the copy path costs exactly about half, the five causes that produce it, and the one log line that confirms which.

The endpoint count is larger than the node count and that surprises everyone building their first cluster. What has to be reachable, why this network must survive when every other one is down, and the security posture it needs because it can power-cycle the entire fleet.

The GPUs are almost never the critical path and planning backward from their delivery date is how deployments slip. The real ordering, what can run in parallel, the two dependencies that cannot, and the cost of a month of delay expressed in the unit that gets attention.

A 44 percent failure rate is far above any plausible hardware defect rate, so the first hypothesis is that the test or the environment is wrong rather than the hardware. The three checks that separate those, the cohort analysis that localizes a real defect, and how to run the vendor conversation.

Switch radix decides it, and at 10,000 endpoints the answer sits right at the boundary for current high-radix switches. The arithmetic, what a third tier costs in switches, optics and power, and the two designs that avoid it entirely.

Same GPU count, same silicon, and a completely different machine. The domain size is the whole argument, it is worth an order of magnitude on specific traffic, and it costs a failure domain nine times larger plus a facility that most halls do not have.

A fixed power envelope turns hardware selection into an optimization with one objective and several overheads competing for the same watts. Where the 1 MW actually goes, why the answer is not simply the most efficient GPU, and the two overheads that can be cut to buy more compute.

Isolation has to be physical where it matters and logical where it can be, and picking the boundary wrong is either expensive or a security problem. What partitions cleanly, what does not, and the allocation unit that decides fragmentation and margin.

Acceptance measures the hardware and production measures the whole system, so the gap lives in everything acceptance did not cover. Four candidate layers, the order to eliminate them, and the two that account for most of these cases.

A fleet-wide version change is the most likely cause of the next unexplained performance regression, so the rollout is designed to make that attributable. Cohorts, a canary that measures rather than boots, and the rollback that has to be real before the first node is touched.

Two failures and a pass is not a passing node, it is a node with a demonstrated intermittent fault and a small sample. The arithmetic that shows how weak the evidence for acceptance is, what an intermittent fault costs once it is in a training job, and the policy that removes the argument.

Doubling per-rack density is four separate upgrades with different lead times, and the electrical one usually cannot be done while the hall runs. The phasing that keeps capacity available through it, the constraint that decides whether it is possible at all, and the number that says whether to upgrade or move.

Modules do not fail randomly at three times a published rate, so the excess has a cause and it is usually thermal or a batch. The cohort cuts that localize it, the counters that separate a dying module from a dirty connector, and the environmental cause that looks like a hardware defect.

Both run at the same port speed, so the discount is not buying slower hardware. It is buying a different operational model and a different failure mode, and whether that is worth 15 percent depends on what your team already runs and what the fabric has to carry.

The naive kernel reads two bytes per multiply-add; a 128x128 tile with register blocking reads a few hundredths of that. Each step (shared-memory tiles, register blocking, vectorized loads, double buffering, tensor cores) with its intensity arithmetic and speedup, and where a hand kernel stops and cuBLAS begins.

Each version fixed a different bottleneck: version 1 fixed memory traffic, version 2 fixed non-matmul work and parallelism, version 3 fixed the fact that softmax and matmul were waiting for each other on Hopper. The arithmetic that shows why a softmax costing one percent of the FLOPs can cost sixty percent of the time.

Four sections, read in a fixed order, and one number that usually names the bottleneck outright. The report of a kernel at 12 percent of DRAM bandwidth while its memory pipeline reads 82 percent busy, what that gap means, the fix it implies, and the numbers the fixed kernel reports back.

At batch 1 a decode step is a few milliseconds of GPU work behind several hundred kernel launches, and the CPU cannot issue them fast enough. The arithmetic showing the GPU starve, what replay changes, and the four constraints capture imposes on memory.

The KV for one sequence is scattered across fixed-size pages, so the kernel reads a table of physical block numbers before it can read any keys. What the indirection costs, what the table costs in memory, why decode must split the KV dimension across thread blocks, and how the partial softmax states combine.

A library call gives you a tuned matmul and nothing around it. Four things a template library gives instead: an epilogue fused into the output stage, one launch for a list of ragged problems, a matmul whose inputs are different types, and a tile schedule that fills the SMs on awkward shapes.

Normalization reads a row and writes a row, so a copy sets the ceiling and everything else is overhead you can remove. The six changes in the order a reviewer wants them, what each is worth, and the one that is a correctness fix rather than a speed fix, with the measured error that proves it.

The target runs one forward pass over all k drafted tokens, compares its distribution against the draft's at each position, and keeps the longest prefix that survives a coin flip. The acceptance rule that makes the output exactly a target sample, the resampling at the first rejection, and the one-integer rollback.

The matmul runs on tiny formats and the accumulator does not, so every low-precision GEMM is really a scaled integer or float product with a correction applied on the way out. Where the scales come from, the three granularities and what each costs, and why one of them needs a periodic promotion into fp32.

Routing hands each expert a different number of tokens, so the layer is a list of ragged matmuls rather than one. Three separate costs a loop pays, the measured share of compute that skewed routing wastes on tile padding, and what a single grouped launch fixes and what it cannot.

Compute the utilization first so you know how much time is unaccounted for, then read one timeline in a fixed order: GPU idle, then what the host was doing in the gaps, then whether the gradient all-reduce overlapped the backward pass, then the optimizer. Four causes, the evidence for each, and what each is worth.

The source translation is a script and takes an afternoon. The assumptions underneath it are the problem: a wavefront is 64 lanes rather than 32, so a warp-level reduction written for 32 ports cleanly and reduces half the data. What breaks silently, what breaks loudly, and the numbers that change every tuning decision.

Six things decide whether a kernel measurement means anything, and a script that ignores them can report a number that is wrong by a factor of two in either direction. What to control, which statistic to report, and the six ways a large speedup turns out to be a measurement artifact.

At 128k tokens one layer's activations are 73 GB per sequence and attention grows with the square of the length. The arithmetic that forces the sequence onto sixteen GPUs, what a ring of KV chunks costs per layer, why the communication hides, and where Ulysses and Megatron sequence parallelism fit around it.

Every MoE layer routes each token to eight experts that live on other GPUs, twice per layer, forward and back. The bytes per rank per layer for DeepSeek-V3, why that all-to-all runs close to one-to-one against compute, and why a single hot expert stalls every rank waiting at the combine.

A router that sends 90% of tokens to 26 of 256 experts makes those ranks do nine times their share while everyone else waits. The arithmetic of the imbalance, what a capacity factor drops on the floor, why the auxiliary loss fights the model, and the bias-only balancing DeepSeek-V3 used instead.

Tensor parallel 8, pipeline 16, data parallel 128: the layout Meta trained Llama 3 405B on, rebuilt from the memory floor, the compute budget and the bandwidth of each link. Why the axes go in that order, what each one costs per step, and the bubble that the schedule has to hide.

A loss that stops moving has five different causes and one of them is not a training problem at all. The checks in order: is the step counter moving, what are the learning rate and the grad norm doing, is the data repeating, has the loss scaler collapsed, and is one rank sending zeros.

Meta's Llama 3 report counted 419 unexpected interruptions in 54 days on 16,384 GPUs, one every three hours. The failure-rate arithmetic, the loss as a function of the interval, the square-root formula that minimizes it, and why the write must be sharded and asynchronous first.

A dead node at 4 a.m. costs twenty minutes if a human restarts it and under five if the system does. The four stages of an automatic recovery, the reason a lost node takes a whole pipeline replica with it, why hot spares beat resharding, and the minutes that each stage still costs even when everything works.

Every collective ends when the last rank arrives, so a single throttled GPU taxes 16,383 others. The arithmetic of that tax, the per-rank timing that finds the rank in one step, and the ranked list of causes from a hot GPU to a NIC with symbol errors, each with the command that confirms it.

Bucketed all-reduce in backward, FSDP's one-layer-ahead prefetch and split-GEMM tensor parallelism hide most of the bytes, and none of it works below a certain number of tokens per GPU. The ratio that decides it, the SMs the NCCL kernels take, and the tail that is never hidden.

On a TPU pod you write one program, annotate how each array is sharded across a named mesh, and the compiler inserts and schedules every collective. What that removes from the engineer's job, what the torus changes about layouts, where Pathways fits, and what a compiler-owned schedule costs.

LoRA shrinks the trainable state and leaves 141 GB of frozen weights untouched. QLoRA attacks the frozen half: NF4 takes it to 35 GB, double quantization compresses the scales themselves, paged optimizers absorb the spikes. The byte accounting, what dequantization costs, and where DoRA adds a term.

Generation is a decode workload and the update is a training workload, and they want different software on different GPUs. The per-token arithmetic for each side, the 141 GB broadcast that has to happen every step, the ratio of actor to learner GPUs that follows, and the idle time that on-policy training builds in.

Sixteen samples per prompt, a verifier that runs code, no value model, responses to 16k tokens: GRPO moves the cost from the learner to rollouts and rewards. The token arithmetic for one step, the tail that holds a batch for the longest sample, and the three places a fleet idles.

Three mantissa bits and a range of 448 make fp8 unusable with one scale per tensor. The arithmetic of what an outlier channel destroys, the 128-element block scaling and fp32 promotion that the DeepSeek-V3 report used, the ops that stayed in bf16, and what the 2× peak bought in practice.

Every rank writes its own shard, a metadata file maps every chunk to a global tensor, and a loader intersects the old layout with the new. The bytes per rank for a 405B, the manifest that makes the write atomic, the metadata that makes resharding possible, and the host-memory staging that keeps the GPUs running.

Prefill and decode fight over the same GPU and each ruins the other's latency. Putting them on separate pools ends the fight, at the price of shipping every request's cache across the network. The break-even is a number you can derive.

A flat median with a broken tail means one in a hundred requests is hitting something the others do not. There are five usual suspects, and the order you rule them out in is the answer.

Three engines, one hardware roofline, and the difference between them is which part of the roofline each one reaches first on your workload. The decision is a table, and the table has reversal conditions.

Three hundred fine-tunes as three hundred deployments would need three hundred GPUs. Batched adapters put them on one base, and the arithmetic says what each extra adapter costs in memory and in step time.

A replica takes a minute to become useful and the traffic changes in seconds. Scale on the wrong signal and you buy GPUs you never use, or you drop requests while they boot. The design is three numbers and two timers.

141 gigabytes have to move from somewhere to eight GPUs, and every hop has a bandwidth. Add the CUDA init, the engine warm-up and the graph capture, and the minute is gone unless you design each step.

671 billion parameters that touch 37 billion per token: the memory bill is for the big number and the bandwidth bill is for the small one, until batch and routing collide. The cache is tiny, and that changes the whole fit.

Least-loaded routing sends a conversation's tenth turn to a replica that has never seen it, and the whole history prefills again. Affinity fixes that and creates hot spots. The router that does both is a scoring function with two terms.

At 128k tokens one request's cache is 43 gigabytes and its prefill is measured in seconds. Capacity breaks first, then TTFT, then the scheduler. Each has a fix, and the numbers say which fix you need at which length.

All three guess tokens for the big model to check. They differ in where the guess comes from, how many candidates they verify per step, and how often the target agrees. Acceptance length per unit of draft cost is the number that decides.

Two hundred milliseconds is about four thousand prompt tokens of prefill on a full node, before any queueing. Every design choice is about not spending that budget twice.

An image is not one token; it is hundreds to thousands of them after a second model has run. That changes the prefill budget, the batch shape and where the first bottleneck appears.

A paused conversation holds gigabytes of cache that its next turn will need in thirty seconds. Recompute costs prefill; reload costs bandwidth. The comparison is one division per tier, and the answer flips with prompt length.

Expert parallelism reads fewer bytes and pays two all-to-alls, so it loses when the interconnect is wrong or when the routing is uneven. Three candidate causes, the measurement that separates them, and why the third one is the answer more often than the first two.

Two point eight trillion parameters fits in one node, because the experts ship in a four-bit format. The footprint arithmetic that reproduces the project's own published hardware minimum, the hybrid cache that needs specific engine support, and why batch-one throughput is a tenth of the bound.

The bound assumes weights stream contiguously and nothing else costs time, and at batch one both assumptions fail badly. Four terms that make up the gap, why sparse models suffer most, and the two fixes that recover most of it.

The two phases want opposite machines, so running them together makes every setting a compromise. What splitting buys, the KV transfer that couples the pools, and the arithmetic that says whether the coupling is affordable for this model.

Forming and then hanging is a different fault from failing to form, and it points at a small set of causes. The rank field that is a cumulative count rather than an index, the collective that never completes, and the three checks that separate configuration from fabric.

It helps most exactly where these models are weakest, which is single-user latency at low batch, and the published gains are larger than on dense models for a reason. The acceptance arithmetic, what it costs at high concurrency, and the measurement that decides.

A tail that wide is queueing or prompt-length variance, and the two are distinguishable in one measurement. Four causes ranked by how often they are it, the decomposition that attributes the wait, and the fix that is usually a flag rather than hardware.

Two steps consume most of the week and neither is the one people plan for. The gate that can end the project on day one, the multi-hour download, the correctness check almost nobody runs, and the canary that has to see real traffic before Friday.

Four things break in order and only one of them is the attention math everyone expects. What a single million-token sequence costs in memory, why the first token takes minutes, and what the model's own architecture already did about the parts that would otherwise be impossible.

Twenty models of different sizes is a packing problem with a cold-start penalty, and the two pull in opposite directions. What stays resident, what loads on demand, the arithmetic that decides which is which, and the interface that stops every team from asking for a dedicated replica.

Engine upgrades change defaults, kernel selection and scheduling at once, so attribution is the whole problem. The bisection that works on a version rather than a commit, the four categories of cause, and the pre-upgrade baseline without which none of this is possible.

Most layers stop having a growing cache, which changes capacity planning by a large factor and demands a cache manager most engines did not have. What the memory model becomes, what the engine must implement, and the two features that were disabled by default while it settled.

The conversion is the easy part and the evaluation is the project. What calibration data does and why yours should come from production, the per-category evaluation that catches what an average hides, and the parts of the model that should not be quantized at all.

Model evaluations answer whether the weights are good and deployment evaluations answer whether your configuration serves them correctly, which is a different and more common failure. What to sample, why category averages hide the regressions that matter, and the reference that makes a result mean something.

One tenant can consume a shared deployment's capacity without exceeding any limit you set, because the limits are usually on requests and the cost is in tokens. What actually needs limiting, the three isolation levels by strength and price, and the fix that works in minutes.

Graphs remove the per-kernel launch cost, which is a large share of a decode step at low batch and a small one at high batch, so whether it matters is a question about your operating point. The arithmetic, the four reasons an engine turns them off, and the measurement that settles it.

Fifty full copies is fifty deployments and is unaffordable. Fifty low-rank adapters over one shared base is one deployment, because an adapter is a tiny fraction of the weights and adapters from different tenants can share a batch. The arithmetic that makes it work and the two limits that decide how far it goes.

Two constraints per replica, memory and throughput, each derived from the model and the traffic; whichever binds sets the users per replica, and the replica count follows with headroom. The full chain from 1,000 users to a number of nodes.

Memory follows total parameters and compute follows active ones, so the two models pull the GPU count in opposite directions. The chain for a 235B-total, 22B-active MoE against a dense 70B, for serving and then for training.

Capex plus two years of power and operations, minus what the cards are worth at the end, against a rental rate at a utilization. The chain that produces a break-even utilization, and the two inputs that swing it from 60% to over 100%.

The token stream is tens of megabytes a second and the checkpoints are tens of gigabytes a second; the two workloads differ by a thousand and want different tiers. The chain from tokens per second to bytes, the checkpoint burst, and the restore that sets the real spec.

Peak sets the fleet, average sets the bill, and the ratio between them is idle money. The chain from a 3x diurnal peak to a replica count, the utilization it implies, what autoscaling can and cannot recover given model load times, and what to do with the trough.

Dollars per FLOP from the GPU price and the MFU, times 6N: a 70B training token costs about three quarters of a microdollar, and the whole 15T-token run follows in one multiplication. The chain, the comparison to an inference token, and why the training token is cheaper.

Each card reads an eighth of the weights in 5.3 ms, then the step pays 160 latency-bound all-reduces and hundreds of kernel launches that do not shrink with sharding. The chain to a 9 to 12 ms step, the communication floor, and why TP8 gives 4x rather than 8x at batch one.

For bf16 weights the decode intensity equals the batch, so the ridge is at batch 295 on an H100. Then the KV cache enters and the answer changes shape: the projections go compute-bound at 300 while the attention reads never do, and at 8k context the step as a whole stays memory-bound at any batch.

Activations are the term the 16-bytes-per-parameter rule leaves out, and at 8k context they are bigger than the weights: about 34 × sequence × hidden bytes per layer, 2.3 GB, 182 GB per micro-batch across 80 layers. The chain, what FlashAttention removes, and what checkpointing does to the number and the compute.

Concurrency equals arrival rate times residence time, and residence time on an LLM replica grows with concurrency. The chain from 50 requests a second to 1,000 in flight, the replica count with a stated utilization, and the feedback loop that turns 90% utilization into a queue that does not drain.

Decode is bandwidth-bound, so cost per token tracks dollars per terabyte per second, and on spec the MI300X wins by 2x. The table, the memory-capacity term that changes the replica shape, and the software-efficiency discount that decides whether the spec ranking survives a benchmark.

Port counts decide everything: a two-tier Clos of 64-port switches tops out at exactly 4,096 endpoints at full bisection, which is why that number appears in so many cluster designs. The switch and optics count both ways, what oversubscription actually slows down, and where the money goes.

Thirty-two senders aimed at one port fill a switch buffer in microseconds, and what happens next depends on which of two mechanisms fires first. The buffer arithmetic, why the slow mechanism is the safe one and the fast one is the dangerous one, and the counters that tell you which is happening.

A tokenizer moves about a megabyte of text per second per core, which makes this a seven-hundred-core-day batch job rather than something to run during training. The throughput arithmetic per stage, the shard format the loader needs, and the determinism requirements that let you resume without corrupting a run.

Six and a half terabytes leaving sixteen thousand GPUs at once is a burst no shared filesystem absorbs, and the fix is to stop trying. What the barrier actually has to wait for, the per-node arithmetic that makes it seconds, and the manifest rule that decides whether a checkpoint is usable at all.

Three parallelism dimensions, two link speeds eighteen times apart, and one mapping that makes the job fast. The traffic each dimension generates per step, why the loudest one must stay inside the node, and the arithmetic showing what a wrong rank order costs before anyone notices.

RoCE inherited a transport that assumes one ordered path per connection, which is why it needs a lossless network built by hand underneath it. What a redesigned transport changes about that, which of today's operational burdens it removes, and why the honest answer to a purchase decision is still about what ships.

Three networks, not one, and the mistake that hurts most is putting two of them on the same wires. The port and switch counts for each, the cable plan, the split of the bill between switches and optics, and the two decisions that are hard to change once the cluster is racked.

A collective stops when one participant stops, and a job of a thousand ranks has no way to distinguish a link that will come back in two seconds from a rank that has died. What the job actually does during those seconds, the cost per flap, and why the fault passes every test you would run.

Exact duplicates are a hash and a group-by. Near-duplicates are a similarity search over billions of documents, which becomes a shuffle rather than a computation. The signature arithmetic, why the shuffle is the expensive stage, and where a GPU classifier fits in a pipeline that is otherwise all CPU.

Two all-to-alls per layer across sixty layers is over a hundred collectives on the critical path of every decode step, each one small and latency-bound. The per-GPU volume worked out, the time on each link type, and why a single large NVLink domain changes the design rather than just improving it.

The signature is unusual: many jobs degrade together, throughput collapses, and the drop counters stay at zero. Reading pause counters backward along the paths to find the origin, the three conditions that produce one, and the configuration change that stops it happening again.

Past the two-tier limit there are two ways forward and they are not close. One triples the switch count and adds fifty percent more optics for two extra hops of latency; the other is the same shape with bigger switches. The counts for both, and the reason the industry follows switch radix.

Every tenant bursts independently, so the shared tier sees the sum of uncorrelated peaks rather than the sum of averages. The provisioning arithmetic that follows, the three isolation boundaries, and why per-tenant local storage solves most of it before any policy is written.

A naive cold start for a 7B function is minutes: pull, start, load, initialize. Under a second means deleting links, not speeding them up: lazy image loading, weights from a local cache, a snapshot of the initialized process with its GPU memory, and a warm pool sized from arrival rate. The chain with a number per link.

Containers share a kernel with the code they run, the wrong boundary for code you did not write. gVisor, Firecracker microVMs and full VMs each buy a stronger one at a cost in memory, start time and GPU access. The overhead arithmetic that turns 50,000 sessions into a host count, and the tiered design.

A shared pool recovers the GPUs inference holds for its peaks, but a preempted training gang takes minutes to give them back and an SLO breaks in seconds. The utilization of each design, the reclaim-time arithmetic against the traffic ramp, and the floor-plus-borrow split most fleets land on.

A 100,000-entry queue cannot be scanned every cycle, a big gang cannot wait forever behind small jobs, and a preemption must not discard an hour of work. The loop in order (admit, order, place, reserve, backfill, preempt), the queue-wait arithmetic, and the checkpoint-aware victim rule.

Allocated is not usable: a cluster full of half-empty nodes has hundreds of idle GPUs that no gang can take. The arithmetic of how mixed job sizes fragment a fleet, the packing policies that prevent it, the defragmentation moves that repair it, and what each costs the jobs already running.

You cannot remove one rank from a synchronous job; you replace the node and restart from checkpoint, and the whole craft is making that take three minutes instead of thirty. The detection signals, the drain-restart-quarantine sequence, the spare-pool arithmetic, and where elastic training changes the answer.

The expensive thing is a base model in GPU memory, not the customer's job. Keep bases resident, pack LoRA jobs onto them, isolate tenants at the process and device level, checkpoint adapters in seconds so jobs run on cheap capacity, and attribute cost per tenant. The arithmetic behind each rule.

Allocated means a scheduler handed the GPU out; utilized means it did work. A 60-point gap is jobs holding GPUs they do not use: idle notebooks, placeholder jobs, a loader-bound run at 20% tensor-active. The three metrics that separate the causes, the per-tenant table that names them, and the policies that close it.

A synchronous training job across two clouds runs at the speed of the link between them, a thousand times slower than the fabric inside either. The arithmetic that kills the idea for one job, the two cases where it is tolerable, and what a platform does instead: partition the work, not the job, and move data ahead.

A pod's default network is an overlay that no RDMA NIC will speak. The three ways to give a pod the fabric (host networking, SR-IOV virtual functions via a NIC device plugin, or a macvlan secondary interface), what each costs in isolation and speed, how NCCL finds the right NIC, and the failure that halves bandwidth.

A driver upgrade touches every layer at once: kernel module, CUDA runtime, container toolkit, NCCL, fabric driver, and every job's image. The compatibility matrix that decides whether a job can run on the new node, the canary that proves it, the wave arithmetic for 2,000 nodes, and the rehearsed rollback.

A hang is a collective that one rank never entered, and finding it means asking which rank is missing rather than what is broken. The flight recorder that answers that in seconds, what to do when it is not enabled, and the four causes that a missing rank turns out to be.

A link that corrupts occasionally does not fail; it retransmits, and each retransmission costs bandwidth that shows up as a slow job with no error anywhere. The counters that reveal it, the arithmetic turning a replay rate into lost bandwidth, and the threshold where watching becomes draining.

The timeline is the easy part and the detection gap is the valuable part: not what broke, but how long it was broken before anyone knew and why. A worked example with its five sections, the distinction between the trigger and the cause, and the property that separates action items that ship from ones that do not.

The controller's real design problem is not detecting faults but deciding when a signal is strong enough to remove a node from a job. The six states and their transitions, the cost asymmetry that sets every threshold, and the arithmetic showing what a one-percent false-positive rate costs at fleet scale.

Each rank keeps a ring buffer of its recent collectives, and on a timeout every rank dumps it. Comparing sequence numbers across a thousand of those buffers finds the missing participant in one pass, and the three patterns those comparisons produce point at three different causes.

Power rises faster than clock speed does, so giving up thirty percent of the power costs closer to eleven percent of the throughput. The relationship behind that, the uniformity requirement that matters more than the level, and why an uncoordinated cap is worse than a deeper coordinated one.

A cold plate holds a few tens of grams of water against seven hundred watts, so temperature rises about two degrees a second and the limit arrives in under twenty. What that timescale rules out, the protection layers that act inside it, and the choice that decides whether you lose a job or the hardware.

Every replacement costs a spare, a maintenance window and a return process; every device kept costs the risk of a job it will fail. Four triggers that decide it without a case-by-case argument, the arithmetic that sets the repeat threshold, and why the device's history beats any diagnostic result.

Three separate reserves get merged into one number and then argued about. The repair pipeline from Little's law, the restart pool that has to be instantly available, and the correlated-failure buffer sized by the largest thing that can fail at once, each derived and then added.

Ten thousand queries a second through an embedder, a sharded vector index, a reranker and a 70B answer model. The latency budget per stage, the two fleets sized from tokens and pairs rather than queries, and the cache that decides whether the LLM tier is 60 nodes or 200.

Ten thousand GPUs is 1,280 nodes, ten pods, about 700 switches, a filesystem that swallows a 5 TB checkpoint in under a minute and a 14 MW feed. The bill of materials with a reason on every line, from the FLOPs the model needs to the spare pool the failure rate demands.

A hundred thousand jobs is a queue you cannot scan per tick, a preemption policy that has to know when each victim last checkpointed, and a state machine with one transition that most designs get wrong. The data model, the tick loop, the checkpoint-aware eviction and the arithmetic behind each.

A serverless GPU is a promise to run a customer's model within seconds of a request while charging only for the seconds it runs. The cold-start chain from image pull to first token, the snapshot that collapses it, the warm pool sized by Little's law, and the isolation and packing decisions that set the margin.

Fifteen trillion tokens starts as a few petabytes of raw text and ends as 30 TB of shards a training job reads at 1.7 million tokens per second. The stages, the bytes at each boundary, the throughput each one needs to finish in two weeks, and the two stages where the pipeline actually spends its time.

A 70B model's KV cache is 320 KB per token, and a 30k-token conversation is 10 GB that GPU memory cannot keep between turns. The three tiers with their bandwidths, the break-even where reloading a prefix beats recomputing it, the eviction and lookup design, and the traffic shape where the tier is worth its complexity.

Two regions at 60% is not one region with a spare: the question is what happens in the 30 seconds after a region drops. Capacity sized for N-1 with the arithmetic, latency routing with a residency override, failover that does not stampede, weights warm everywhere in advance, and the state that must not cross a border.

Every 30 minutes a thousand GPUs write a terabyte in a burst that must finish in a minute, then nothing until the next. The burst arithmetic that sizes the write tier, the two-stage path to durable storage, retention that keeps the right checkpoints, resharding so a 512-GPU checkpoint resumes on 256, and lineage.

A GPU cloud sells whole nodes and fabric slices to tenants who expect them in minutes and expect them to work. The node state machine, the reconciler that keeps thousands converging, the isolation lines on shared fabric, the health loop that pulls a bad node first, and billing that survives a control-plane outage.

200 milliseconds at p99 is a budget, and prefill alone spends it on a long prompt. The decomposition into network, queue, routing, prefill and first token, the prompt length the budget allows, chunked prefill in the tail, admission control that rejects what it cannot serve in time, and the utilization the SLO costs.

An agent thinks for a second, runs a tool for a minute, and repeats a hundred times over an hour. The session state machine and where each piece lives, the sandbox per session with its cold-start budget, the KV cache that should survive the tool call, and checkpoints that resume a step rather than an hour.

A video request is not a stream of tokens: it is 40 denoising passes over a latent the size of a small film, and the user waits a minute for the whole thing. The memory and FLOP arithmetic per request, why batching works differently, the queue and progress design for minute-long jobs, and where the cost goes.

Speculative decoding buys latency with spare compute, and a busy fleet has no spare compute. The speedup formula with acceptance rate and draft cost inside it, the batch size where the gain turns negative, where the draft runs and what it costs, the monitoring that catches a silent regression, and rollout per class.

A three-month run on 8,000 GPUs is a large bet against everything that can happen to one building. The recovery point and the recovery time as numbers, the checkpoint replication that sets the first, the second cluster and its warm state that set the second, and the drills that make the numbers true.

Prefill is compute-bound and decode is bandwidth-bound, and one GPU running both does each worse. The pool ratio from the traffic's prompt and output lengths, the KV transfer cost per request and the link it needs, the router that pairs a prefill worker with a decode worker, and the load at which it beats a monolith.

The first question is whether 1T is total or active, because storage follows one and compute follows the other. The state budget, the layout that falls out of it, why expert parallelism belongs inside NVLink, and the two failure modes a dense-model plan does not have: router imbalance and all-to-all congestion.

A free list, a per-sequence block table, and a reference count are the whole structure. What forking costs, why the first write after a fork must copy, and the measured waste that this design accepts in exchange for eliminating the far larger waste of pre-allocation.

Admission, a step that grows every sequence, and preemption under memory pressure. The loop is short and it has a failure that only appears when the cache is tight: in a measured run it preempted seventeen times and completed nothing, because every preempted sequence restarted from zero.

A plain hash ring with one point per replica measured a 47-fold load imbalance across eight replicas. With 150 virtual nodes each it fell to 1.19-fold, and adding a ninth replica moved 11.9 percent of keys against an ideal of 11.1. Both numbers, and what they cost.

Two ends, two access patterns, and one design choice that removes almost all contention. Why the owner takes the newest task and thieves take the oldest, what that does to cache locality, and the measurement showing why this cannot be demonstrated in Python at all.

Fair sharing of one resource is obvious and fair sharing of three is not, because a GPU-heavy job and a memory-heavy job are not comparable. The rule that makes them comparable, the greedy loop that implements it, and the tie-breaking bug that gave one user everything.

Metrics arrive in time order and are queried by time window, which is the pair of facts the whole design turns on. Sealed segments with a sparse index, pruning that skipped 82 of 84 segments in the executed test, and TTL compaction that drops whole files without rewriting a byte.

Both easy answers fail. Refusing without an alternative makes you the obstacle and guarantees the next person routes around you. Complying makes the check meaningless. The four-step response that holds the line and still gets the researcher moving today.

The interesting content is the order things broke in, because that order is a property of the system and it repeats. What breaks first at each scale step, why the control plane usually goes before the data plane, and the thing you would skip if you did it again.

You cannot win this with an opinion, and a flat objection to a purchase reads as an engineer who does not understand the business. The utilization evidence that carries it, the smaller counter-proposal that usually wins, and what to do once the decision goes against you.

This is a test of whether you have a position you can defend, not of which problem you pick. What makes a thesis defensible, two worked examples with the arithmetic behind them, and the counterargument you have to be able to state before the interviewer does.

Not depth, which is where most people push and where the returns flatten. What changes is the size of the problem you are handed, how ambiguous it is when you get it, and whether things happen because you asked rather than because you built them.

A confident plan built before you know anything is the failure mode, and so is refusing to answer until you have looked. The structure that handles both: a dated shape with the decision points named, one committed win, and the conditions that would change everything after day 30.

How you reported it matters more than how you found it. The internal disclosure that gets a fix instead of a defensive reaction, the four places isolation gaps hide in GPU infrastructure, and the test that keeps the fix from regressing.

Bad news delivered as a problem gets you a meeting. Delivered as a forecast with options, costs and a decision date, it gets a decision. The four-part structure, the projection that has to be defensible, and the mistake of waiting until you are certain.