8 min read
Engrams Embedding Entendre: Codesign for Efficient DRAM/SSD Offloading
New Model Architecture Implications for TAM of DRAM/NVMe, DeepSeek V4.1 Flash, AgentX, InferenceX, NVMe experiments
- inference
- agentx
- deepseek
- +3
InferenceX Research
Articles
Benchmark write-ups on agentic inference, AgentX results, and chip and serving-stack economics.
New to the terminology? Browse the AI inference glossary.
Tagged “inference”
34 articles
10 min read
Rubin NVL72 Agentic Inference: 67x better Performance per Dollar
Jensen Sandbagging Performance Again, 2x more Annual Profit Per GigaWatt, The More you Buy, The More you Earn, AgentX, InferenceX, Extreme Co-Design
- benchmark
- gpu
- inference
- +7
35 min read
TPU Inference Externalization Full Steam Ahead
InferenceX, Up to 50% Better Performance per Dollar, Rapid Externalization of TPU stack, Growing Customer Base, Ironwood, TPUv8i, Reducing CUDA Moat
- benchmark
- inference
- tpu
- +7
3 min read
DeepSeek V4 Pro on AgentX: B200 vs B300 and the KV Cache Working Set
50% more HBM squeezes out extra throughput, and the per-point telemetry shows exactly where it comes from
- agentx
- agentic
- benchmark
- +6
2 min read
DeepSeek V4 Pro on AgentX: GB200 vs GB300 Rack-Scale Disaggregation
Both lean on PD disagg, GB300 adds DEP32 wide-EP decode, and the gap shows up in first-token latency rather than token rate
- agentx
- agentic
- benchmark
- +8
3 min read
DeepSeek V4 Pro on AgentX: MI355X vs B200, and the August 21 Flip
AMD matched B200 vLLM on performance per dollar for end-to-end latency, then upstream vLLM work moved the line
- agentx
- agentic
- benchmark
- +6
2 min read
GLM 5.3 on AgentX: MI355X ATOM Beats GB300 NVL72 on Part of the Curve
Where AMD’s vendor engine wins on performance per dollar, and what E2E Normalized Interactivity actually measures
- agentx
- agentic
- benchmark
- +9
2 min read
GLM 5.3 on AgentX: NVIDIA Is Up to 5x Cheaper per Token at 150 tok/s/user
At this operating point, free AMD silicon would still not close the gap
- agentx
- agentic
- benchmark
- +8
2 min read
Kimi K3 on AgentX: MI355X ATOM Beats GB300 NVL72 on Part of the Curve
AMD’s vendor engine wins a real slice of the performance per dollar frontier, while Hopper struggles to serve K3 at all
- agentx
- agentic
- benchmark
- +8
3 min read
MiniMax M3 on AgentX: Why B200 and B300 Beat Their Rack-Scale GB200 NVL72 & GB300 NVL72 Counterparts
The Dynamo router becomes the bottleneck, no submission runs context parallelism, and AMD leaves KV offload on the table
- agentx
- agentic
- benchmark
- +10
2 min read
MiniMax M3 on AgentX: B300 TRT-LLM TP2 Owns the Crown
NVIDIA sweeps the 432B model, and the missing DP-attention points explain why cache locality became a routing constraint
- agentx
- agentic
- benchmark
- +8
23 min read
OpenAI Jalapeño: Better Than Nvidia Blackwell
OpenAI’s self-designed ASIC compared with Rubin, Jalapeño’s TCO, throughput per MW, and spicy deets
- inference
- benchmark
- gpu
- +4
2 min read
Qwen3.5 397B on AgentX: B300 FP4 Delivers 12x the Performance per Dollar of H100
What four years of hardware and a 4-bit format buy on a long-context agentic workload
- agentx
- agentic
- benchmark
- +9
2 min read
MI355X versus GB300 NVL72 Inference Performance: 20x Gap on Qwen3.5 SGLang
GatedDeltaNet, a 262k native context, and no AMD competition at all on the same engine
- agentx
- agentic
- benchmark
- +6
64 min read
AgentX - InferenceXv3: Does the CUDA Moat Hold Up in Agentic Inferencing?
$3 Million USD dataset open sourced, 1 Mil+ Context Length, Multiturn, Sub Agents 95%+ KVCache HitRate, GB300 NVL72, MI355X, B200
- agentx
- agentic
- benchmark
- +5
10 min read
Agentic Benchmark for LLM Inference: Metrics and Methodology
How an agent benchmark replays long-context, multi-turn workloads to measure latency, throughput, cache behavior, and serving cost
- benchmark
- agents
- agentic
- +3
6 min read
A Brief Overview of Agentic Workloads
Multi-turn sessions, long contexts, and near-total prefix reuse make agentic inference a systems problem, and change what a benchmark has to measure
- agentic
- agents
- agentx
- +2
17 min read
Ultra-High Interactivity on NVIDIA GPUs? TileRT on InferenceX
Can TileRT software on NVIDIA GPUs compete with Cerebras, Groq LPU, and SambaNova? Batch size 1, disaggregated engine, high-throughput prefill engine, high-interactivity decode engine
- benchmark
- gpu
- inference
- +8
23 min read
Kimi K3: The Manos, The Mythos, The Legendos
Kimi K3's architecture: compressed memory, attention across depth, latent expert routing, and serving performance
- inference
- benchmark
- gpu
- +6
20 min read
Vera Rubin NVL72 vs GB200 NVL72? Inference TCO & Architecture Analysis
Rubin LUT Based Tensor Core, Feynman, Rack Scale, Perf Per MegaWatt, Perf Per Dollar, Software Improvements, Public Rubin Software, PyTorch, vLLM, OpenAI Triton
- benchmark
- gpu
- inference
- +7
29 min read
DeepSeekV4 1.6T Day 0 to Day 43 Performance Over Time — Huawei, GB300 NVL72, MI355X, B200
Day 0 Inference Performance, InferenceX, 100x performance improvement in 26 Days, Cost per Million Tokens, Huawei 950DT Inference Trace Analysis
- benchmark
- gpu
- inference
- +13
10 min read
GB300 NVL72 vs GB200 NVL72 Inference Performance & Perf per Dollar - on DeepSeek-V4-Pro 1.6T: Up to 2.83x Throughput
DSv4-Pro FP4 8K/1K, Dynamo+vLLM, disaggregated on both racks. GB300's 50% extra HBM (288 vs 192 GB/GPU) unlocks a wider prefill+decode recipe GB200 can't fit — lifting middle-of-curve perf/$ by 2.31x despite a 20% per-GPU TCO premium.
- benchmark
- gpu
- inference
- +9
10 min read
B200 NVFP4 vs H200 FP8 on GLM-5: Up to 3.65x Better Performance per Dollar with SGLang MTP
Both SKUs run SGLang EAGLE MTP; the Blackwell generation lifts perf/$ by ~1.2x at the peak and the NVIDIA GLM-5-NVFP4 checkpoint on FlashInfer TRT-LLM sparse MLA stacks another ~2.4–3.0x on 8K/1K
- benchmark
- gpu
- inference
- +6
12 min read
B200 NVFP4 vs H100 FP8 on MiniMax-M2.5: Up to 8.2x Better Performance per Dollar with vLLM
vLLM PR #36307 unlocks the trtllm-gen FP8 MoE kernel for MiniMax on B200; combined with NVFP4, perf/$ scales from 4.0x at 22 tok/s/user to 8.2x at 110 on 8K/1K
- benchmark
- gpu
- inference
- +6
11 min read
B200 NVFP4 vs H200 INT4 on Kimi K2.5/K2.6: Up to 2.95x Better Performance per Dollar
On vLLM 8K/1K the NVFP4 path on B200 is 2.71x–2.95x cheaper per million tokens than H200 INT4 across the entire 30–90 tok/s/user serving band, and 2.45x–2.74x cheaper than B200 INT4 on the same silicon. Both factors decompose cleanly into B200's HBM bandwidth, HBM capacity, and NVFP4 tensor cores
- benchmark
- gpu
- inference
- +6
13 min read
MI355X DeepSeek-V4-Pro on SGLang: 110.5x Throughput per GPU in 26 Days
The amd/deepseek_v4 side branch shipped TileLang attention indexer, Triton sparse MLA, fused RoPE/Hadamard, FlyDSL MoE, and FP4 weights across 31 performance optimizations PRs — lifting first-light 20 tok/s/GPU at 2.4 tok/s/user into 2,256 tok/s/GPU at 9.4 tok/s/user on 8K/1K, with both throughput and interactivity climbing together
- benchmark
- gpu
- inference
- +6
7 min read
AMD MI355X GLM-5 Inference: Up to 40% Cheaper per Million Tokens than B200 on SGLang FP8
14 weeks after GLM-5 launched, AMD landed both MTP and non-MTP SGLang FP8 recipes on MI355X — fused MLA + FP8 KV cache via TileLang flips the single-node FP8 cost curve in AMD favor across most of the performance Pareto
- benchmark
- gpu
- inference
- +7
6 min read
AMD MI355X Qwen3.5 397B-A17B Inference: Up to 19x Throughput per GPU in 3 Months on SGLang FP8
From v0.5.8 (Feb) → v0.5.10rc0 (Apr) → v0.5.12 (May), three AITER kernel landings on MI355X plus a TP=8 → TP=2/TP=4 retune push Qwen3.5 8k/1k peak from 1.3k to 6.4k tok/s/GPU and extend the curve out to 75 tok/s/user
- benchmark
- gpu
- inference
- +5
8 min read
GB200 NVL72 vs B200 on DeepSeek R1 670B: Up to 4.4x Throughput per GPU at 125 tok/s/user
DeepSeek R1 FP4 1k/1k. NVL72's 72-GPU NVLink scale-up fabric lets decode run wide EP up to EP=32, where B200's 8-GPU NVLink island caps out at EP=8 over RoCEv2
- benchmark
- gpu
- inference
- +9
5 min read
SGLang 0.5.6 on B200 DeepSeek R1 FP4: Up to 1.8x at Low Concurrency
Piecewise CUDA graphs for DeepSeek V3, a unified event loop, and JIT kernels push 8k/1k throughput from 508 to 907 tok/s/GPU on the same 16 GPU B200 pool
- benchmark
- inference
- gpu
- +5
6 min read
GB200 NVL72 vs B200 on Kimi K2.5: 3.1x from Wide EP vLLM
Rack scale NVLink on NVL72 lets Dynamo vLLM run Kimi K2.5 wide EP up to Decode EP 16, taking peak throughput from 4,021 to 12,587 tok/s/GPU on 8k/1k NVFP4
- benchmark
- gpu
- inference
- +7
6 min read
AMD MI355X Kimi K2.5 Inference: 7.7x Throughput, Up To 15x Interactivity in 25 Days on vLLM
vLLM PR #35850 Fixed AITER MLA Dispatch on MI355X CDNA4, Unlocking Kimi K2.5 Inference Performance at TP=8, Shipped in vLLM 0.18
- benchmark
- gpu
- inference
- +5
47 min read
InferenceX v2: NVIDIA Blackwell Vs AMD vs Hopper - Formerly InferenceMAX
GB300 NVL72, MI355X, B200, H100, Disaggregated Serving, Wide Expert Parallelism, Large Mixture of Experts, SGLang, vLLM, TRTLLM
- benchmark
- gpu
- inference
- +1
37 min read
InferenceMAX: Open Source Inference Benchmarking
NVIDIA GB200 NVL72, AMD MI355X, Throughput Token per GPU, Latency Tok/s/user, Perf per Dollar, Cost per Million Tokens, Tokens per Provisioned Megawatt, DeepSeek R1 670B, GPTOSS 120B, Llama3 70B
- benchmark
- gpu
- inference
- +1