Skip to the content.

The industry dimension catalog

Every dimension the field competes on, researched industry-first (not derived from what fak measured). ← back to the index.

Dimensions are grouped by the part of the stack they belong to. Each group is a section below; follow the group link for the full per-dimension analysis with fak’s position.

agent

See the full agent analysis for fak’s honest position on each dimension.

Category Dimension SOTA bar Leading systems Source Date
agent-fleet Cross-agent fleet serving time (N agents × T turns): work eliminated by shared-prefix fusion The competing floor remains tuned warm per-agent KV reuse (vLLM Automatic Prefix Caching / SGLang RadixAttention / OpenAI prompt caching); fak’s own measured 50x5 fleet-serving result (19 min vs 78 min warm-KV baseline, 4.1x from cross-agent prefix fusion) is the live number on this dim and no external SOTA has displaced the warm-KV-reuse baseline class. vLLM Automatic Prefix Caching, SGLang RadixAttention, OpenAI prompt caching (warm-KV reuse floor) link 2026-01
agent-fleet Marginal value of cross-agent prefix fusion ON TOP of an already-warm per-agent cache The conservative SOTA point is fak’s own already-warm per-agent KV arm (baseline 1.0x); cross-agent prefix fusion on top of it yields a measured ~2.4-2.7x marginal gain. No external benchmark redefines the already-warm-cache baseline for this dim. Tuned warm per-agent KV cache, already hot (vLLM APC / SGLang RadixAttention class) link 2026-01
agent-fleet Cross-agent reuse marginal value vs a LIVE tuned shared-prefix engine (head-to-head) Head-to-head marginal value of cross-agent reuse vs a LIVE tuned shared-prefix engine (vLLM APC / SGLang RadixAttention, which also fuses a shared prefix once) remains UNMEASURED; fak’s own WebVoyager stratification puts the marginal-vs-tuned win at only ~1.0-1.10x on that workload. SGLang RadixAttention (live process), vLLM Automatic Prefix Caching (live process) link 2026-01
client-compaction Long-session history compaction that preserves the provider prompt-cache prefix (drop-and-splice vs summarize-and-resend) The field splits two ways. Summarizers (Aider, LangChain ConversationSummaryMemory, Codex CLI, Copilot CLI) and Anthropic’s API context-editing REWRITE/clear content and break the prefix cache (OpenAI docs: ‘when you drop, summarize or compact earlier turns … you’ll break the cache’; Anthropic context-editing docs: ‘Invalidates cached prompt prefixes when content is cleared’). The cache-preserving sub-field is small: LangChain trim_messages (pure sliding-window drop, no rewrite) and Copilot’s DELIBERATE cache-boundary reset reason about it; no surveyed tool ships a PROVEN byte-identity splice on by default. Context-editing pairs clearing with a memory tool for recovery (a different axis: recoverability), reporting up to ~84% token reduction on an Anthropic agent eval. Anthropic API context-editing (clear_tool_uses_20250919), OpenAI Codex CLI (summary compaction), GitHub Copilot CLI (cache-boundary compaction), Aider (ChatSummary), LangChain (trim_messages / ConversationSummaryMemory) link 2026-01
model-routing Model routing granularity: per-aspect + first-class ensemble routing vs whole-request single-model selection Surveyed 2025-2026 routers/gateways (RouteLLM, Martian, NotDiamond, Unify, OpenRouter, Portkey, LiteLLM Router) all route the WHOLE request to ONE model; OpenRouter Fusion is the only shipped model ensemble and it is a fixed parallel-synthesize recipe, not a configurable per-aspect reduction. None routes a sub-request aspect to its own model. (Aurelio Semantic-Router routes to an intent, not a model; vLLM/SGLang routers balance replicas of ONE model for KV locality — a different layer.) RouteLLM (LMSYS), Martian, NotDiamond, Unify.ai, OpenRouter (+ Fusion), Portkey, LiteLLM Router link 2026-06
request-routing Cache-aware request routing across replicas (route to the worker that already holds the prefix) Cross-region KV-locality-aware routing (GORGO-proxy) makes network latency an explicit routing term and cuts median TTFT ~2.5x vs least-load / prefix-trie baselines (224ms vs 568ms) while preserving prefix-cache locality across regions. GORGO / GORGO-proxy (cross-region, arXiv 2602.11688), SkyWalker (locality-aware cross-region LB, EuroSys 2026, arXiv 2505.24095) link 2026-02
time-to-solution Long-horizon agent time-to-solution: wall-clock time for an agent (or agent fleet) to actually FINISH a multi-hour engineering task FrontierSWE (Proximal Labs, published 2026-04-16) is the field ultra-long-horizon coding benchmark: 17 real tasks (5 implementation, 9 performance-engineering, 3 ML-research) with a 20-hour-per-task budget, graded on a zero-to-one scale rather than binary pass/fail. It is UNSATURATED - the top model (Claude Fable 5, avg rank 2.35 / 90 percent dominance) still fails to solve almost all tasks, and agents burn many hours (e.g. Opus 4.6 ~6.6-13.8h/task) making little progress. There is no published time-to-solution winner: the SOTA bar is that long-horizon completion is an OPEN problem the field measures but nobody yet finishes fast. FrontierSWE (Proximal Labs) - the ultra-long-horizon coding benchmark, Claude Fable 5 / Opus 4.6 + agent scaffold (top FrontierSWE ranks), GPT-5.4 + agent scaffold link 2026-04

cost

See the full cost analysis for fak’s honest position on each dimension.

Category Dimension SOTA bar Leading systems Source Date
cost-efficiency Inference unit economics ($ per 1M tokens) at realistic utilization GPT-4-equivalent quality now serves at ~$0.40 / 1M tokens vs ~$20 in late 2022 (≈10x/yr decline). Real production GPU utilization is only 30-60%, so realistic cost-per-token runs 2-3x above spreadsheet estimates; FP8/FP4 quantization roughly halves cost-per-token on H100/H200. Self-hosting a 7B breaks even only above ~50% utilization. Hosted frontier APIs, vLLM/SGLang self-host on H100/H200, FP8/FP4 quantization link 2025-09-01
cost-efficiency End-to-end effective request capacity / cost gain from the KV system under SLO Mooncake’s KVCache-centric disaggregated architecture increases effective request capacity by 59%-498% on real traces (up to 525% in simulation) while meeting SLOs; this remains the headline effective-capacity-under-SLO bar, now republished in ACM Transactions on Storage (2026). Mooncake (Moonshot AI / Kimi), NVIDIA Dynamo (KVBM + NIXL KV transfer) link 2024-07
cost-efficiency Energy efficiency (Wh per token / tokens-per-watt) and power-capped throughput Empirical inference energy is ~0.0001-0.002 Wh per output token depending on model size; a ~500-token GPT-4o-class query is ~0.3 Wh (~3e-4 Wh/token). FP4/FP8 and Blackwell-class hardware are the main levers. No neutral audited tokens-per-watt leaderboard yet exists, so the bar is a measured range, not a single SOTA number. NVIDIA Blackwell (GB200/GB300, FP4), AMD MI355X (FP4) link 2026-03-01
cost-efficiency Goodput under an SLA (requests/s meeting TTFT and TPOT SLOs simultaneously) DistServe’s headline 7.4x more goodput (or 12.6x tighter SLO) over prior systems remains the canonical bar; newer 2025-26 work (TaiChi, PD-Multiplexing, DOPD, NVIDIA Dynamo) reports only incremental gains relative to disaggregation itself rather than a new baseline-relative record. DistServe, NVIDIA Dynamo (incremental: 38% over best aggregated config, 2026), PD-Multiplexing/GreenContext (1.62x over PD-disagg, 2025-09) link 2024-01
cost-efficiency SLO-constrained goodput (per-GPU) DistServe’s 4.48x SLO-constrained goodput (the conservative per-workload figure under strict TTFT+TPOT SLOs) still stands as the canonical bar; no 2025-26 system reports a higher SLO-constrained goodput multiplier against the same baseline. DistServe, DOPD (>99% SLO attainment, up to 1.5x over vLLM/DistServe, 2026-03), TaiChi (90% SLO attainment, 29-77% over PD-disagg, 2025-08) link 2024-01

decoding

See the full decoding analysis for fak’s honest position on each dimension.

Category Dimension SOTA bar Leading systems Source Date
speculative-decoding Draft acceptance length / acceptance rate (mean accepted tokens per verify step) EAGLE-3 reports per-task acceptance rate ~0.80-0.88 (coding/instruction) and mean accepted length ~4.5-5.0 tokens per draft-verify cycle, exceeding the older DeepSeek-V3 MTP ~0.85 second-token figure. EAGLE-3 (SafeAILab), DeepSeek-V3 MTP, FastMTP / FastEagle (2025 successors) link 2025-03
speculative-decoding Breadth of speculative methods supported (draft-model, EAGLE/-2/-3, Medusa, MTP, lookahead, n-gram/prompt-lookup) TensorRT-LLM supports draft-target, Medusa, ReDrafter, Lookahead, and EAGLE(-1/-2, EAGLE-3 via two-model/disaggregated); vLLM supports EAGLE/EAGLE-3, Medusa, MTP, n-gram/prompt-lookup; SGLang supports EAGLE/EAGLE3 and standard draft TensorRT-LLM, vLLM, SGLang link 2025-12
speculative-decoding Retrieval / prompt-lookup (n-gram) drafting for input-grounded tasks SuffixDecoding (NeurIPS 2025, model-free suffix-tree drafting) is the new prompt-lookup-family SOTA: outperforms classic PLD/n-gram by 1.3-3x and delivers up to 5.3x speculative speedup (and ~4x end-to-end on SWE-Bench agentic tasks) over vanilla decoding. SuffixDecoding (Snowflake Arctic Inference + vLLM), vLLM PLD/n-gram (prior bar), TensorRT-LLM Lookahead link 2025-05
speculative-decoding Speculative-decoding behavior under high concurrency / large batch (break-even) EAGLE-3’s measured high-concurrency behavior: positive throughput uplift sustained to batch ~64 in SGLang (1.38x at batch 64), with the speedup approaching break-even beyond that as baseline throughput rises; EAGLE-2 already regresses (0.93x at batch 24). EAGLE-3 (SGLang), MagicDec (fixed-window draft, prior reference), vLLM/SGLang continuous batching link 2025-03
speculative-decoding Output-distribution losslessness of speculation (greedy and stochastic) EAGLE 3.1 (vLLM/EAGLE/TorchSpec, May 2026) is the current SOTA lossless speculative-decoding method: up to 2x longer acceptance length and up to ~2.03x per-user output throughput vs EAGLE-3 (Kimi-K2.6-NVFP4, GB200), fixing ‘attention drift’ robustness; correctness is preserved by rejection-sampling verification that matches the target distribution (greedy spec == greedy non-spec). EAGLE 3.1 (vLLM + TorchSpec), EAGLE-3, SpecForge training framework link 2026-05
speculative-decoding Speculative-decoding wall-clock speedup (single-stream / low-concurrency) EAGLE-3 reports up to 4.79x speedup on LLaMA-3.3-70B with no quality loss and is the de-facto industrial standard supported in vLLM and SGLang; vLLM’s P-EAGLE reaches ~4-5x over standard decoding (20-30% over EAGLE-3 alone) on coding benchmarks EAGLE-3 (SafeAILab), vLLM, SGLang, P-EAGLE (vLLM) link 2025-09
speculative-decoding Speculative decoding effect on ITL (acceptance-rate-bound) EAGLE-3 reports ~0.80-0.88 acceptance on coding/instruction tasks giving ~3-4x decode token-rate (ITL) gains on H100/H200, and outperforms EAGLE-2/Medusa-2 by 15-25% tokens/s at batch size 1. Speedup primarily reduces ITL, not TTFT. EAGLE-3, Medusa-2, TensorRT-LLM / vLLM speculative decoding link 2026-01
speculative-decoding Speculative decoding throughput effect across batch sizes (does it survive at high concurrency?) EAGLE-3 remains the named SOTA: ~1.38x throughput at batch 64 in SGLang on a single H100, peaking at ~1.81x at batch 2 and up to ~6.5x single-stream, ~1.4x over EAGLE-2. EAGLE-3 (SafeAILab), SGLang speculative backend, vLLM/TensorRT-LLM spec backends link 2025-03
structured-output Constrained/guided decoding per-token overhead (JSON-schema, regex, grammar/CFG) XGrammar-2 (arXiv 2601.04426, Jan 2026) drops grammar-mask per-token overhead to ~12-20us (Llama tool-calling ~12-13us, OpenAI Harmony ~15-20us) with <6% end-to-end overhead vs unconstrained and ~10ms (>100x faster) dynamic compilation; it is the new SOTA, far below LLGuidance’s ~250us (Harmony) / >1000us (tool-calling) on the agentic workloads. XGrammar-2 (vLLM/SGLang/TensorRT-LLM default backend), XGrammar, LLGuidance (Rust Earley) link 2026-01
structured-output Grammar/schema compilation latency and cache (cold-start for dynamic agentic schemas) XGrammar-2 compiles in ~10 ms vs >1000 ms for XGrammar (~80x faster), with >6x tool-calling compilation speedup via Earley-based adaptive mask cache + cross-grammar cache + JIT; integrated into vLLM and SGLang XGrammar-2, XGrammar, LLGuidance/Guidance link 2026-05
structured-output Sampler coverage and correctness (temperature, top-p, top-k, min-p) Reference samplers cover the full mainstream set: temperature, top_p, top_k, min_p, typical, presence/frequency/repetition penalties, logit_bias, seed/n/best_of, plus pluggable custom logits processors (argmax-invariant classification). vLLM and llama.cpp converge on Temperature+Min-P as the 2026 default power-user combo. vLLM SamplingParams + logits processors, llama.cpp sampler chain, SGLang sampler link 2026-01
structured-output Structured-output validity vs reasoning-quality tax (constraint-induced degradation) Constrained decoding guarantees ~100% syntactic/schema validity (JSON-Pass/Type-Safety >.95) but does NOT guarantee semantic quality and can REDUCE accuracy: ExtractBench (2026) shows structured mode dropping overall validity 51%->37% and GPT-5 credit-agreement pass rate 86.9%->70.0% vs prompt-based extraction. The honest SOTA bar is ‘validity is solved; value accuracy is the open gap (13.7pt spread across models).’ XGrammar (vLLM/SGLang/TRT-LLM default), LLGuidance (guidance-ai), OpenAI Structured Outputs link 2026-02
structured-output Tool/function-call parsing coverage and robustness (per-model formats) vLLM ships per-family parsers (hermes, llama3_json, mistral, pythonic, internlm, granite, plus qwen3/qwen3_coder); when constraints aren’t enforced it falls back to extracting from raw text, so args can be malformed - making grammar-backed (constrained) tool calling the robustness frontier vLLM tool-call parsers, SGLang, XGrammar-2 (grammar-backed tool calling) link 2025-12

distributed

See the full distributed analysis for fak’s honest position on each dimension.

Category Dimension SOTA bar Leading systems Source Date
parallelism Data-parallel (DP) attention / hybrid attention-FFN parallelism SGLang’s DP Attention is the production technique that eliminates KV-cache duplication across devices for DeepSeek-V3, enabling the EP32/EP72 deployment to scale memory-efficiently; combined with PD disaggregation and MTP it delivers up to ~60% higher output throughput. SGLang, vLLM link 2025-07-17
parallelism Large-scale expert parallelism (EP) for giant MoE models EP320: Huawei CloudMatrix-Infer runs DeepSeek-R1 with 320-way expert parallelism (one expert per NPU die: 32 shared + 256 router + 32 redundant experts) over the UB network, superseding the 96-GPU large-scale EP (effective EP~72) demonstrated by LMSYS/SGLang. Huawei CloudMatrix-Infer (CloudMatrix384, DeepSeek-R1), SGLang + DeepEP (96-GPU, prior bar), vLLM expert-parallel deployment link 2025-06
parallelism Multi-node scale-out (combined TP x PP x EP x DP across hosts) A single GB300 NVL72 is 72 Blackwell-Ultra GPUs in one NVLink domain; MLPerf Inference v6.0 (2026-03-30) composed FOUR of them – 288 GPUs over Quantum-X800 InfiniBand, NVIDIA’s largest-ever submission – into one coherent DeepSeek-R1 671B MoE serving run at ~2.494M tok/s aggregate offline. Disaggregated serving via NVIDIA Dynamo composes the TPxPPxEPxDP dimensions across the fabric (up to 1.5x on Llama-3.1-405B interactive). NVIDIA GB300 NVL72 x4 (288 GPUs over Quantum-X800 InfiniBand, MLPerf v6.0), SGLang/TensorRT-LLM + NVIDIA Dynamo multi-node, GB200 NVL72 (prior generation) link 2026-04-01
parallelism Prefill-decode disaggregation & per-phase SLO isolation Phase disaggregation for SLO isolation (separating compute-bound prefill from memory-bound decode so each phase’s SLO is met independently) is now production-standard via NVIDIA Dynamo, which reports 38% goodput over the best aggregated config and 7x with wide expert parallel on GB200 NVL72; DistServe’s 7.4x remains the founding goodput figure. NVIDIA Dynamo 1.0, DistServe, TaiChi (unifies agg+disagg to fix both TTFT and TPOT violations, 2025-08) link 2026-03
parallelism Prefill-decode disaggregation: independent scaling of compute-bound prefill and memory-bound decode SGLang PD disaggregation achieves up to 6.9x more goodput (Chatbot 3200in/800out, TTFT<=1000ms, TPOT<=25ms) on MI300X; this remains the headline open-framework PD-disaggregation figure, now corroborated by NVIDIA Dynamo’s 2026 production results (38% goodput over best aggregated config; 7x with wide expert parallel on GB200 NVL72). SGLang PD (MI300X), NVIDIA Dynamo 1.0 (38% over best aggregated config; 7x w/ wide-EP on GB200), vLLM disaggregated prefilling (NIXL) link 2025-08

memory

See the full memory analysis for fak’s honest position on each dimension.

Category Dimension SOTA bar Leading systems Source Date
kv-cache Attention architecture impact on KV footprint (MQA/GQA/MLA bytes-per-token) Multi-head Latent Attention (MLA, DeepSeek-V3) holds ~70 KB/token KV footprint vs 192-328 KB/token for GQA models (a 2.7-4.7x reduction, ~9x vs MHA, ~2x vs 8-group GQA) at higher quality; the frontier is now MLA + learned sparsity (DeepSeek NSA, 2025) pushing toward sub-100 floats/token for long contexts. MLA (DeepSeek-V3), MLA + Native Sparse Attention (NSA, DeepSeek 2025), Qwen-Latent / MiniMax-Text-01 MLA-style link 2025-02
kv-cache Automatic prefix caching / RadixAttention prefix reuse and cache hit rate Automatic prefix caching now reaches ~85-95% cross-request hit rates on shared-prefix workloads (SGLang RadixAttention few-shot 85-95%, agentic ~88%; vLLM APC ~87% warm-cache hit, ~88% faster TTFT); the original RadixAttention 6.4x throughput claim (LMSYS 2024-01) remains the canonical headline number. SGLang RadixAttention, vLLM Automatic Prefix Caching link 2024-01
kv-cache Automatic cross-request prefix / KV-cache reuse (shared system prompts, agent scaffolds, few-shot) Provenance: SGLang-published KV plane. SGLang RadixAttention’s automatic cross-request KV reuse via an LRU radix tree gives up to 6.4x throughput on shared-prefix workloads with 50%-99% cache hit rates; this remains the canonical cross-request prefix-reuse bar in 2026. SGLang RadixAttention, vLLM Automatic Prefix Caching, Mooncake KV pool link 2024-01
kv-cache KV-cache-aware / prefix-aware request routing across a fleet KV-cache-aware routing across a replica fleet now delivers ~50% lower TTFT, ~34% lower TPOT, ~61% more RPS and ~62% more output tok/s (Baseten on NVIDIA Dynamo, Qwen3 480B), with llm-d reporting 3x throughput / 2x TTFT vs round-robin and GKE Inference Gateway 13.9x faster TTFT than a 3rd-party service. Baseten + NVIDIA Dynamo KV router, llm-d v0.5 (GAIE/EPP prefix-cache-aware), GKE Inference Gateway prefix-cache-aware routing link 2025-10
kv-cache KV-cache memory efficiency and max concurrent sequences (paging / fragmentation) PagedAttention keeps KV-cache memory waste under 4% (near-zero internal/external fragmentation), enabling 2-4x throughput at the same latency; this remains the canonical paged-KV efficiency reference as of mid-2026. vLLM PagedAttention, SGLang RadixAttention (radix-tree paged KV) link 2023-09
kv-cache KV-cache quantization (FP8 / INT8 / KIVI sub-4-bit) accuracy at long context Ultra-low-bit KV quant frontier (2026): Together AI’s OSCAR reaches 2.28 effective bits within 1.42 points of BF16 (Qwen3-8B; ~0.02-pt gap on Qwen3-32B) with ~8x KV memory reduction at 100K context and up to 7.83x job-level throughput; TurboQuant (ICLR 2026) is near-lossless at 3-4 bits / full-precision NIAH recall at 4x. OSCAR (Together AI, 2026), TurboQuant (Google/NYU, ICLR 2026), Kitty 2-bit dynamic channel-wise (arXiv 2511.18643) link 2026-06
kv-cache KV compression / token eviction under a fixed cache budget Reasoning-aware eviction now beats the old H2O/Quest budget cliff: R-KV achieves lossless compression at 34% KV budget on MATH-500 and 10% on AIME-2024, and at 16% budget exceeds FullKV at 105% accuracy; Apple’s EpiCache (2026) gets near-full-cache accuracy at 4-6x compression with up to 40% higher accuracy than eviction baselines and 3.5x lower peak memory. R-KV (redundancy-aware, reasoning, arXiv 2505.24133), EpiCache (Apple, 2026), SnapKV / H2O / Quest (baselines now superseded) link 2025-05
kv-cache KV-cache offloading to CPU/host/NVMe/remote (multi-tier hierarchy depth) LMCache (graduated to production Jan 2026, OSDI/arXiv paper Oct 2025) delivers up to 15x throughput improvement over vanilla vLLM by tiering KV cache across HBM/DRAM/NVMe with a global prefix index; llm-d production benchmarks report up to 57x faster TTFT and 2x throughput vs round-robin under high prefix reuse on 16 H100s. LMCache (vLLM, production Jan 2026), llm-d native KV offload (57x TTFT), Mooncake Store link 2025-10
kv-cache KV-cache transfer for prefill/decode disaggregation (NIXL / NCCL / UCX) NVIDIA Dynamo 1.0 (GA March 16, 2026 at GTC) delivers up to 7x higher throughput with disaggregated prefill/decode serving combined with wide expert parallelism, measured on DeepSeek-R1 on GB200 NVL72; KV blocks move prefill->decode over RDMA via NIXL at wire speed. NVIDIA Dynamo 1.0 + NIXL (GA Mar 2026), Mooncake Transfer Engine, SGLang EPD disaggregation link 2026-03
kv-cache Non-prefix / cross-document KV reuse (RAG chunk caching) LMCache CacheBlend reuses NON-prefix KV caches (any RAG chunk, not just the first) and selectively recomputes ~15% of tokens, achieving near-100% KV-cache hit rate with 2.2-3.3x lower TTFT and 2.8-5x higher throughput vs full recompute; won ACM EuroSys’25 Best Paper. LMCache CacheBlend, Mooncake (non-prefix KV pool) link 2025
kv-cache Paged KV block management and memory utilization (anti-fragmentation) PagedAttention achieves ~96.3% KV-cache memory utilization (the complement of <4% waste); no 2026 system reports a higher published paged-KV utilization figure. vLLM PagedAttention, SGLang RadixAttention link 2023-09
kv-cache Prefix/KV-cache reuse impact on TTFT (cache-hit latency) Qualitative: APC reuses exact KV pages via hash-based block matching, cutting prefill tokens and TTFT for repeated/long-prefix prompts with negligible steady-state overhead; overhead from hash compute grows with input length and concurrency. Distributed prefix-cache scheduling (llm-d) extends the win across replicas. vLLM Automatic Prefix Caching (APC), SGLang RadixAttention, llm-d distributed prefix cache link 2025-01
kv-cache Prefix caching / automatic KV reuse across requests (shared-context workloads) SGLang RadixAttention delivers ~29% higher end-to-end throughput than vLLM on H100 (16,200 vs 12,500 tok/s, Llama 3.1 8B) on shared-prefix workloads, with up to 6.4x on prefix-heavy RAG/multi-turn traffic; the 29% SGLang-vs-vLLM edge still holds in 2026 benchmarks. SGLang RadixAttention, vLLM Automatic Prefix Caching link 2026

models

See the full models analysis for fak’s honest position on each dimension.

Category Dimension SOTA bar Leading systems Source Date
hardware-coverage Hardware backend breadth (NVIDIA, AMD, Intel, TPU, AWS, Apple, CPU) vLLM runs on NVIDIA, AMD ROCm (MI300X/MI325X/MI350X/MI355X), x86/ARM/PowerPC CPU, plus plugin backends for Google TPU, Intel Gaudi 2/3, AWS Inferentia/Trainium (Neuron), Apple Silicon, Huawei Ascend, IBM Spyre/Z, and more; llama.cpp targets the widest portability surface (CUDA, HIP/ROCm, Metal, Vulkan, SYCL, CANN, OpenCL, CPU SIMD on x86 AVX-512/AMX and ARM Neon/SVE/SME). MLC-LLM uniquely adds WebGPU/browser via TVM compilation. vLLM, llama.cpp, SGLang, MLC-LLM link 2026-01-01
hardware-coverage Hardware / format coverage and portability of quantized kernels FP8 W8A8 is hardware-accelerated on NVIDIA Hopper/Ada/Blackwell and AMD MI300X (hipBLASLt FP8 GEMM + aiter attention); NVFP4 is Blackwell-only; MXFP4 acceleration on MI300X/MI355X is on the roadmap, not yet broadly shipped. GGUF k-quants give the widest CPU/consumer-GPU portability. vLLM (FP8 on Hopper/Ada/Blackwell + MI300X via hipBLASLt/aiter), SGLang (FP8 MI300X), TensorRT-LLM (NVFP4 Blackwell), llama.cpp (broad CPU/GPU GGUF + FP4) link 2026-01
hardware-shape-neutrality Hardware-shape neutrality: explicit fences for host-CPU assumptions A hardware-neutral stack exposes backend contracts, capability discovery, layout/dtype/residency boundaries, and explicit fallback behavior so new accelerators do not require a fork of the model loop. Mature serving stacks cover more device families, but most public scorecards still report breadth without a seven-assumption shape-neutrality ledger. vLLM plugin backends, llama.cpp backend matrix, MLC-LLM / TVM compilation link 2026-06-30
model-coverage Embedding & reranker (non-generative) model coverage vLLM and SGLang both serve embedding and reranker/cross-encoder models alongside generative models with pooling support and OpenAI-compatible embeddings endpoints, covering the common open embedding/reranker families (BGE, E5, GTE, Qwen-embedding, etc.) on the same engine. vLLM, SGLang, TensorRT-LLM link 2026-01-01
model-coverage Long-context serving (128K-1M tokens) & KV/prefix-cache reuse Million-token windows are production-ready (Qwen2.5-1M open weights; 128K-1M is mainstream). SGLang’s RadixAttention prefix caching delivers up to ~6.4x gains on prefix-heavy RAG/multi-turn workloads and a ~29% throughput edge over vLLM on H100 (16.2K vs 12.5K tok/s) on shared-prefix traffic. SGLang (RadixAttention), vLLM (PagedAttention/prefix caching) link 2026-01-01
model-coverage LoRA / multi-LoRA hot-swap serving The peer-reviewed SOTA is S-LoRA (MLSys’24, arXiv:2311.03285): ~2000 LoRA adapters served over one base model on a single A100-80GB at up to 4x vLLM throughput, via unified paging of adapter weights + custom MBGMM/MBGMV kernels. vLLM Multi-LoRA integrated S-LoRA-style techniques as a follow-on (max_loras GPU-resident + max_cpu_loras LRU CPU cache + runtime hot-swap); SGLang/TensorRT-LLM offer comparable multi-adapter batching. Practical resident count is bounded by GPU memory and adapter rank. S-LoRA (MLSys’24, ~2000 adapters/A100-80GB), vLLM Multi-LoRA, SGLang, TensorRT-LLM link 2025-08
model-coverage Maximum parameter count & frontier-MoE architecture coverage Kimi K2 (Moonshot AI) – 1 trillion total parameters / 32B active, 384 experts (8+1 active) – is the largest open-weight MoE routinely served by vLLM and SGLang, exceeding DeepSeek-V3/R1’s 671B. Kimi K2 / K2.x (Moonshot AI, 1T MoE; vLLM + SGLang + KTransformers), DeepSeek-V3/R1 671B (prior bar), Qwen3 MoE / Llama-4 (other frontier MoE) link 2025-07
model-coverage Multimodal / vision-language (VLM) model coverage SGLang is explicitly a serving framework ‘for large language models and multimodal models’ and extends RadixAttention prefix caching to multimodal inputs with vision-embedding reuse; vLLM also provides broad multimodal model support. Coverage spans the major open VLM families (Qwen-VL, Llama Vision, InternVL, Pixtral, etc.). SGLang, vLLM link 2026-01-01
model-coverage Capability/quality of the served model (SWE-bench Verified, GPQA, AIME, LiveCodeBench) Frontier models served by these engines define the quality bar buyers cite: SWE-bench Verified ~80.6% (Gemini 3.1 Pro) / 79.2% (Claude Opus 4.5 + Live-SWE-agent, Nov 2025); GPQA Diamond effectively saturated ~93-94% (Gemini 3 Deep Think 93.8%, GPT-5.4 92.8%); AIME 2025 up to 100% (GPT-5.4). These are model-served-by-engine results, not engine claims. Gemini 3.x (SWE-bench/GPQA), Claude Opus 4.5 + agent scaffold, GPT-5.x (AIME) link 2025-11-24

numerics

See the full numerics analysis for fak’s honest position on each dimension.

Category Dimension SOTA bar Leading systems Source Date
numerical-correctness Deterministic / bitwise-reproducible inference (batch invariance) Batch-invariant kernels (invariant normalization, matmul, attention reductions) yield 1000/1000 bitwise-identical completions under dynamic batching, vs dozens of distinct outputs without; cost ~60% slower than default kernels. Now exposed as a vLLM batch-invariance feature. Thinking Machines batch-invariant-ops (on vLLM FlexAttention), vLLM batch-invariance mode, LLM-42 (deterministic speculation) link 2025-09
numerical-correctness Numerical-correctness error metric vs reference (perplexity, KL-divergence, recovery %) Best practice pairs perplexity with KL-divergence against the FP16 baseline (llama.cpp emits both); KLD captures per-token distribution shift perplexity hides. Empirically: 5-bit GGUF is a strong near-lossless trade-off, 3-bit shows clear degradation; imatrix is essential below 3-bit, marginal at Q4_K_M+. Generation text-similarity is reported as a supplementary correctness check. llama.cpp (perplexity + KL-divergence tooling), Neural Magic / vLLM recovery-% methodology, ModelOpt accuracy harness link 2026-01
numerical-correctness Dense GPU compute correctness with ZERO vendor GEMM (cuBLAS-free) cuBLAS is the tuned vendor GEMM every other stack carries; matching it bit-for-bit (cosine=1.0) is parity by construction. cuBLAS (NVIDIA tuned vendor GEMM) link 2026-06
quantization Full 4-bit (W4A4 + KV4) via rotation/outlier removal Learned-rotation methods now lead 4-bit weight+activation: SpinQuant (ICLR 2025) closes the gap to full precision by up to 45.1% relative to QuaRot on hard-to-quantize Llama-3 8B, and 2025-2026 successors (Qronos, ResQ PCA-rotations, TurboQuant for KV cache, WUSH adaptive transforms) advance further; Hadamard/learned rotations are the standard outlier-removal primitive feeding NVFP4/W4A4. SpinQuant (learned rotations), Qronos (2025), ResQ PCA-rotation (2025), TurboQuant (2026, KV cache), QuaRot (Hadamard baseline) link 2025-04
quantization 4-bit microscaling float (NVFP4 / MXFP4 / FP4) accuracy vs FP8 on Blackwell NVFP4 (4-bit, 16-value blocks, FP8 E4M3 fractional block scales) now achieves <=1% accuracy drop vs FP8 on large reasoning models (DeepSeek-R1-0528: MMLU-PRO 85->84, GPQA 81->80) and ~99% recovery vs BF16 for 70B-235B models, at ~2.3x higher throughput than weight-only 4-bit; clearly the SOTA 4-bit format on Blackwell, well ahead of MXFP4 (32-value, power-of-two scales). NVFP4 via NVIDIA TensorRT Model Optimizer + LLM Compressor (vLLM, Blackwell B200/GB200), NVFP4 (vLLM Marlin) link 2025-09
quantization FP8 (W8A8, E4M3) accuracy retention vs bf16 reference FP8 (W8A8, E4M3) weight+activation quantization is effectively lossless: ~99.9% accuracy recovery on HumanEval and <=1% degradation vs BF16 across all model scales, including Llama 3.1 405B fully recovering accuracy. FP8 W8A8 via LLM Compressor (vLLM), NVIDIA TensorRT Model Optimizer FP8, AMD Quark FP8 link 2024-10
quantization INT4 weight-only (W4A16: AWQ, GPTQ) accuracy and calibration sensitivity INT4 weight-only (W4A16) via AWQ or GPTQ keeps accuracy degradation within negligible-to-moderate range (<4% typical, often <1-2% on HumanEval); AWQ is the de-facto best-practice INT4 format for vLLM in 2026 and generally edges out GPTQ on perplexity. AWQ (AutoAWQ) via vLLM, GPTQ / GPTQModel + Marlin (vLLM), W4A16 LLM Compressor link 2026-01
quantization int8 / Q8_0 SIMD decode throughput vs the same-rung int8 peer HuggingFace dynamic-int8 is the standard same-rung CPU int8 reference (1×); llama.cpp Q8_0 CPU is the harder peer (~parity). HuggingFace dynamic-int8, llama.cpp Q8_0 link 2026-06
quantization INT8 W8A8 (SmoothQuant) accuracy under activation outliers INT8 W8A8 (per-channel static weight + per-token dynamic activation, optionally SmoothQuant) recovers ~99% of BF16 accuracy, i.e. ~1% average loss; vLLM/LLM Compressor report negligible Open-LLM-leaderboard drop. INT8 W8A8 SmoothQuant via LLM Compressor (vLLM), SmoothQuant (Xiao et al.) link 2025-03
quantization KV-cache quantization (FP8 / INT8 / INT4 KV) FP8 (e4m3) KV-cache quantization is the production standard: ~46% KV memory reduction (per-token cost ~54% of BF16), +14.9% output throughput / -14.8% median ITL on Llama-3.1-8B, recovering 97-98% of baseline long-context AUC at 128k and fully recovering aggregated AUC@1M. INT4 KV gives ~75% reduction but is research-stage; vLLM ships FP8 only. FP8 e4m3 KV (vLLM production, Apr 2026), INT4 KV (research: ~75% reduction), FP8 e5m2 (vLLM) link 2026-04
quantization Accuracy-constrained throughput on a neutral benchmark (MLPerf Inference) Under MLPerf’s accuracy- and latency-constrained server scenario (TTFT and TPS/user SLOs enforced at p99), MLPerf Inference v6.0 (2026-03-30) DeepSeek-R1 reaches 8,064 tokens/second/GPU on GB300 NVL72 (vs 9,821 tok/s/GPU unconstrained offline); DeepSeek-R1’s server SLO is 2s TTFT with a 12.5 tok/s/user p99 target while holding the required accuracy bar, and the GB300 server result is ~25% higher per-GPU than GB200 and up to 2.7x the GB300 debut via software. NVIDIA GB300 NVL72 (Blackwell Ultra, TensorRT-LLM/Dynamo), NVIDIA GB200 NVL72 (Blackwell), CoreWeave / Nebius GB300 NVL72 submissions link 2026-04-01
quantization Quantization format & low-precision datatype coverage Reference serving stacks now cover the full ladder: vLLM (2026) supports FP8 (W8A8), INT8 (W8A8 + W4A8), INT4 (W4A16), NVFP4 and MXFP4 (via Marlin), AWQ, GPTQ, BitsAndBytes (4/8-bit), AMD Quark, and online runtime quantization, across Turing->Hopper->Blackwell plus x86/Arm CPU for INT8. vLLM quantization (LLM Compressor, Marlin, AutoAWQ, GPTQModel, bitsandbytes, Quark), TensorRT Model Optimizer (FP8/INT4/NVFP4) link 2026-01
quantization Quantization granularity (per-tensor / per-channel / per-group / microscaling block size) Per-block (block-32) microscaling beats per-tensor/per-channel; measured SNR gains of 3.0-3.4 dB over group-128 and 9.2-9.4 dB over per-tensor for two-level microscaling. Common group sizes 32/64/128; group-128 is the de-facto 4-bit weight-only default. NVFP4’s block-16 + FP32 scale is finer than MXFP4’s block-32 power-of-two. MX formats (OCP MXFP4/MXFP8, block-32), NVFP4 (block-16, FP32 second-level), group-128 INT4 (AWQ/GPTQ), MOSS two-level microscaling link 2025-11

operability

See the full operability analysis for fak’s honest position on each dimension.

Category Dimension SOTA bar Leading systems Source Date
operability Cold-start / autoscaling latency (scale-to-zero & scale-out) GPU-snapshotting cold start is now sub-200 ms: RunPod FlashBoot achieves sub-200 ms for ~48% of requests (<250 ms generally) and Modal GPU memory snapshots cut Ministral-3 3B median cold start ~10x (118 s -> 12 s) - the corroborable bar. InferX additionally claims an industry-leading 177 ms on H100 via GPU snapshotting, but that single number traces to one vendor source and is not independently confirmable, so the defensible figure is sub-200 ms. RunPod FlashBoot (sub-200 ms / <250 ms), Modal GPU memory snapshots (10x: 118 s -> 12 s), InferX (177 ms vendor claim, GPU snapshot, uncorroborated) link 2026
operability Fairness, priority, and tenant isolation in batch formation / scheduling No single audited cross-system number; SOTA is qualitative: FairBatching reformulates batch formation to bound per-request unfairness, and serving stacks add priority classes plus preemption (KV-page eviction with recompute-on-resume in SGLang) to isolate tenants. Buyers evaluate head-of-line-blocking behavior and priority-preemption support directly. FairBatching (fairness-aware batch formation), SGLang (FCFS + page-eviction preemption), hybrid real-time/best-effort schedulers (e.g. arXiv 2504.09590) link 2025-10
operability Multi-node fault tolerance & failure recovery (resilient serving) Research systems define the bar: KevlarFlow reports ~20x lower mean-time-to-recovery via decoupled parallel init, dynamic traffic rerouting, and background KV-cache replication; FailSafe sustains high-performance tensor-parallel serving under irregular GPU availability; ReviveMoE targets fast recovery from hardware failures specifically in large-scale MoE inference. Production stacks (Dynamo) add dynamic GPU scheduling and rerouting but full resilient serving is still maturing. FailSafe (research), KevlarFlow (research), ReviveMoE (research), NVIDIA Dynamo link 2026-01-01
operability Kubernetes-native deployment + LLM-aware autoscaling (metric-driven, scale-to-zero) KServe + KEDA autoscale vLLM on LLM-specific signals (e.g. vllm:num_requests_running) rather than CPU, supporting scale-from/to-zero with a default 5-min cooldown; the 2025-2026 production stack is vLLM + KServe/llm-d + Ray + Kueue + KEDA on K8s. Ray Serve and KServe provide canary, traffic-split, and multi-model endpoints. The SOTA bar is custom-LLM-metric autoscaling with scale-to-zero, not generic HPA. KServe + KEDA, Ray Serve, llm-d / vLLM on Kubernetes link 2025-09-23
operability Multi-tenant fairness, per-tenant quotas, and SLO attainment under contention VTC (the 2024 first fair scheduler) is now superseded by a 2025 family: Equinox (whole-system fair scheduling), FairBatching (fairness-aware batch formation), PROSERVE (multi-priority SLO-aware), and DLPM (locality-aware fair scheduling) – all addressing VTC’s gaps on prefix-locality, SLO, and diverse workloads. Equinox (2025), FairBatching (2025), PROSERVE (2025), DLPM locality-aware fair scheduling (2025), VTC (2024, prior bar), LiteLLM virtual-keys/budgets link 2025-08
operability Production observability: Prometheus metrics (TTFT/TPOT/queue) + OpenTelemetry tracing vLLM v1 exposes an extensive Prometheus /metrics surface (vllm:time_to_first_token_seconds, TPOT/inter-token, vllm:num_requests_running, vllm:kv_cache_usage_perc, prefix-cache hit) plus native OpenTelemetry distributed tracing via –otlp-traces-endpoint – the de facto serving-observability contract. vLLM v1 (Prometheus + OpenTelemetry), SGLang, NVIDIA Dynamo / GenAI-Perf link 2026-01
operability OpenAI-compatible API surface + reliable structured output / tool-calling Constrained/guided decoding now guarantees 100% schema-valid structured output by construction (XGrammar-2 is the default engine across vLLM/SGLang/TensorRT-LLM, <40 us/token), with the OpenAI-compatible /v1/chat/completions + response_format json_schema surface as the de facto contract. vLLM + XGrammar-2 (default), SGLang, TensorRT-LLM, OpenAI Structured Outputs (response_format) link 2026-01
operability Cross-instance / persistent KV cache sharing and coherence LMCache enables KV reuse across different vLLM engine instances plus multi-GPU peer-to-peer KV sharing and disaggregated prefill/decode; Mooncake Store exposes a global, deduplicated KVCache pool addressable across the cluster LMCache, Mooncake Store link 2025-10
operability Latency degradation under concurrency / load (saturation behavior) Under concurrency, single-GPU TTFT P99 explodes (e.g. ~8.6 s at C=180, ~34 s at C=420); pipeline parallelism (PP=2) holds it to ~3 s at C=180 and ~13 s at C=420 (~2.5-3x improvement). At 100 concurrent requests TensorRT-LLM P95 TTFT ~1,280 ms vs vLLM ~1,450 ms; vLLM sustains QoS to C=32 and stays usable to C=64. TensorRT-LLM (lowest P95 TTFT under load), SGLang, vLLM (most robust queueing/scheduling) link 2026-01

security

See the full security analysis for fak’s honest position on each dimension.

Category Dimension SOTA bar Leading systems Source Date
security Prompt-injection defense and agent security (attack-success-rate vs utility-under-attack) On AgentDojo, design-by-construction defenses now reach provable security: CaMeL drives successful indirect-prompt-injection attacks to 0 (vs 8 for the next-best tool-filter) while still solving 77% of tasks (vs 84% undefended); MELON (ICML’25) likewise beats prior SOTA defenses. Undefended best agents still see ASR <25%. CaMeL (Defeating Prompt Injections by Design, 2025), MELON (ICML’25), Meta SecAlign (secure foundation LLM), tool-filter / SecAlign baselines link 2025-03
security Tool/agent sandboxing, structural containment, and PII/exfil prevention Capability/IFC-based structural containment is the 2025 SOTA: CaMeL (Defeating Prompt Injections by Design) extracts control/data flow so untrusted data can never alter program flow and enforces capability policies at tool-call time (0 successful AgentDojo attacks); MELON adds masked re-execution + tool-comparison provable defense. CaMeL (capability + control/data-flow IFC, 2025), MELON (masked re-execution, ICML’25), container/microVM sandboxes, IFC / capability allow-lists link 2025-03

serving

See the full serving analysis for fak’s honest position on each dimension.

Category Dimension SOTA bar Leading systems Source Date
latency End-to-end request latency (E2EL) End-to-end request latency (TTFT + total decode) is reported per request by the standard tools (NVIDIA GenAI-Perf, vLLM bench serve, LLMPerf) with P99 as the default percentile; there is no single SOTA number - it is a tunable composite governed by the per-model MLPerf v6.0 TTFT+TPOT SLOs (e.g. a 1000-token DeepSeek-R1 interactive response budgets ~1.5 s TTFT + ~15 s decode). NVIDIA GenAI-Perf, vLLM bench serve (–metric-percentiles, default p99), LLMPerf link 2026-01
latency Inter-token latency / time-per-output-token (ITL/TPOT) MLPerf Inference v6.0 reasoning-model interactive scenarios now require TPOT <= 15 ms P99 (GPT-OSS 120B and DeepSeek-R1), down from the prior 80 ms TPOT bar; the legacy Llama-2-70B interactive bound is 40 ms TPOT (~25 tok/s/user) and server is 80-200 ms. MLPerf Inference v6.0 (GPT-OSS 120B, DeepSeek-R1 interactive), MLPerf Inference v5.1 link 2026-03-24
latency Interactive latency SLO attainment: TTFT and TPOT/ITL tail under load MLPerf Inference v6.0 (Apr 2026) tightened the interactive SLO floor: the most stringent latency-aware bounds are now TPOT <= 15 ms with TTFT <= 1.5-2.0 s P99 (GPT-OSS 120B: TTFT <= 2.0 s / TPOT <= 15 ms; DeepSeek-R1: TTFT <= 1.5 s / TPOT <= 15 ms), superseding the v5.1 30 ms-TPOT / 0.5 s-TTFT interactive bar. MLPerf Inference v6.0 (NVIDIA Blackwell Ultra GB300, AMD MI355X), MLPerf Inference v5.1 link 2026-03-24
latency Prefill-decode interference & chunked-prefill TTFT/TPOT tension Chunked prefill cuts P95 inter-token latency by ~68% under mixed long/short workloads (e.g. 2,800 ms -> 890 ms P95 ITL at 32K inputs) by prioritizing decode and chunking prefill into the token budget. vLLM chunked prefill (decode-prioritized scheduling), SGLang chunked prefill link 2026-06
latency Tail latency distribution (P50/P95/P99/P99.9) P99 (99th percentile) is the canonical tail SLO across MLPerf Inference v6.0 (all TTFT/TPOT constraints are ‘99th percentile <= X’), NVIDIA GenAI-Perf/NIM, and vLLM bench (default –metric-percentiles p99). P99 remains the industry tail metric; some leaderboards also surface P95/P50. MLPerf Inference v6.0 (99th-percentile TTFT/TPOT constraints), vLLM bench (p99 default), NVIDIA GenAI-Perf / NIM link 2026-03-24
latency Time-to-first-token (TTFT) under realistic prefill MLPerf Inference v6.0 defines per-model P99 TTFT SLOs: GPT-OSS 120B interactive TTFT <= 2.0 s (server <= 3.0 s), DeepSeek-R1 interactive TTFT <= 1.5 s (server <= 2.0 s), Llama-3.1-405B interactive TTFT <= 4.5 s (server <= 6.0 s). The 4.5 s P99 405B bar from v5.1 still stands for the largest model. MLPerf Inference v6.0 (DeepSeek-R1 interactive), MLPerf Inference v5.1 (Llama-3.1-405B interactive) link 2026-03-24
scheduling Request scheduling and cache-aware routing (FCFS vs priority vs prefix/KV-cache-aware) SGLang’s cache-aware load balancer still holds the cited bar: up to 1.9x throughput and 3.8x higher prefix-cache hit rate in multi-node deployments, via an approximate radix tree mirroring each worker’s cache. SGLang router / cache-aware load balancer (v0.4+, current v0.5.x) link 2024-12
scheduling Chunked / piggyback prefill and prefill-decode interference (stall-free batching) Proactive intra-GPU prefill/decode disaggregation (Nexus) now beats chunked-prefill, delivering up to 2.2x higher throughput, 20x lower TTFT, and 2.5x lower TBT vs vLLM (and up to 2x vs SGLang), superseding Sarathi-Serve’s 5.6x chunked-prefill capacity claim. Nexus (proactive intra-GPU PD disaggregation, arXiv 2507.06608), DuetServe (adaptive GPU multiplexing, arXiv 2511.04791), Sarathi-Serve (chunked prefill, original bar) link 2025-08
single-stream Single-stream (one chat) decode throughput on CPU llama.cpp Metal/CPU leads single-stream decode on the same box (e.g. Qwen2.5-7B Q8 ~17.3 tok/s Metal, M3 Pro). llama.cpp, Ollama, MLC-LLM link 2026-06
single-stream Single-stream decode throughput on a consumer GPU llama.cpp Q8_0 ~120 ± 15 tok/s on an RTX 4070 (-ngl 99). llama.cpp (CUDA) link 2026-06
single-stream Single-stream prefill throughput, apples-to-apples CPU-vs-CPU llama.cpp CPU sets the prefill bar on the same box + Q8 weights (1× reference); fak arm64 is ~0.12× until the int8 tile lands. llama.cpp link 2026-06
throughput Absolute aggregate tokens/s at scale on standardized hardware (MLPerf Inference) MLPerf Inference v6.0 (results 2026-03-30 / published 2026-04-01) supersedes v5.0 and v5.1, and the headline model is now DeepSeek-R1 (671B MoE reasoning), not Llama 3.1 405B: NVIDIA’s largest-ever submission of 288 Blackwell Ultra GPUs (4x GB300 NVL72 over Quantum-X800 InfiniBand) hit ~2.494 million tokens/second aggregate in the offline scenario, at 9,821 tok/s/GPU offline (8,064 tok/s/GPU server) — up to 2.7x higher than the GB300 debut six months prior via TensorRT-LLM/Dynamo software. NVIDIA GB300 NVL72 x4 (288 Blackwell Ultra GPUs, TensorRT-LLM/Dynamo), NVIDIA GB300 NVL72 single rack, CoreWeave / Nebius GB300 NVL72 link 2026-04-01
throughput Continuous / in-flight (iteration-level) batching: max aggregate token throughput at high concurrency Continuous (in-flight) batching is now table stakes across vLLM/SGLang/TensorRT-LLM; the live competitive number is no longer ‘vLLM 24x vs Orca/TGI’ but cross-engine: SGLang’s RadixAttention holds roughly a 29% aggregate-throughput edge over vLLM on H100 (~16,200 vs ~12,500 tok/s, Llama 3.1 8B), widening to up to ~6.4x on prefix-heavy RAG/multi-turn and shrinking to near-zero on unique-prompt batches; continuous batching itself still delivers ~3-4x over naive/static batching. SGLang (RadixAttention), vLLM v1 (continuous batching + automatic prefix caching), TensorRT-LLM (in-flight batching) link 2026
throughput Latency-vs-throughput Pareto frontier The latency-throughput Pareto frontier is now framed around goodput (max request rate meeting both TTFT and TBT/TPOT SLOs), and the dominant lever has moved from Sarathi-Serve chunked prefill (2024) to prefill-decode (PD) disaggregation: DistServe reports up to 7.4x more goodput or 12.6x stricter SLOs than prior systems while keeping >90% of requests in budget, and 2025-2026 work (DynaServe, MuxWise, DuetServe, RAPID-Serve) pushes intra-GPU PD multiplexing to recover disaggregation’s isolation without dedicating GPU pools. DistServe (PD disaggregation), Mooncake / Splitwise (PD disaggregation), Sarathi-Serve (chunked prefill, now a baseline point), DuetServe / MuxWise / DynaServe (intra-GPU PD multiplexing, 2025-2026) link 2024-2026
throughput Peak served throughput per accelerator (model-served-by-engine, MLPerf-grounded) MLPerf Inference v6.0 (2026-03-30) is the current round and the audited per-accelerator peak: DeepSeek-R1 671B on GB300 NVL72 reaches 9,821 tok/s/GPU offline (8,064 tok/s/GPU accuracy+latency-constrained server); the full 288-GPU (4x GB300 NVL72) submission aggregates ~2.494M tok/s (see absolute-tokens-per-second-mlperf). The audited single-H200 Llama-2-70B peak of ~33,000 tok/s (40% over H100) still stands for that older/smaller workload. The prior v5.1/v5.0 ~1.1M tok/s/rack Azure GB300 figure was an unverified submission and is superseded. NVIDIA GB300 NVL72 (DeepSeek-R1 671B, MLPerf v6.0, TensorRT-LLM/Dynamo), NVIDIA H200 (Llama-2-70B audited), MangoBoost/AMD MI300X link 2026-04-01