TL;DR: The best CPU for LLM inference is the one attached to the most memory bandwidth, full stop. In August 2026 that means a used EPYC Genoa (460.8 GB/s, $1,925) for capacity, the AMD Ryzen 9 9950X ($485) if you’re staying dual-channel, and the Ryzen AI Max+ 395 if you want one box that just works.

	Ryzen 9 9950X desktop	Used EPYC 9554 server	Ryzen AI Max+ 395 mini PC
Best for	CPU inference on a PC you already need	100B+ MoE models no consumer GPU fits	Turnkey 128GB unified-memory box
Memory bandwidth	~89.6 GB/s (dual-channel DDR5-5600)	460.8 GB/s (12-channel DDR5-4800)	256 GB/s (LPDDR5X-8000)
30B-class MoE speed	~8–20 tok/s	~63 tok/s	~100 tok/s (iGPU-assisted)
Cost, August 2026	~$485 CPU + RAM you own	~$1,925 CPU + $5,200+ for 384GB RAM	~$3,499 (128GB, up from $1,999)
The catch	Dual-channel bandwidth wall	The RAM costs 2.7× the CPU	Price surged 75% mid-2026

Honest take: If a model fits in 24GB of VRAM, a used RTX 3090 (~$1,248) beats every CPU on this page by a mile. Buy a CPU for inference only when the model can’t fit in VRAM you can afford — and then buy memory channels, not cores.

Google’s own search data is why this article exists: people ask “what are the best processors for running LLM inference” expecting a core-count shootout, and every spec sheet nudges them toward the most expensive 16-core flagship. That’s the wrong axis. Token generation reads the model’s active weights out of RAM for every single token, so the CPU that wins is the one bolted to the widest memory bus. A 96-core EPYC on 4 populated channels loses to a 32-core EPYC on 12. Check what your target model needs with our VRAM calculator first — the same math (weights read per token) governs CPU inference, just with system RAM in place of VRAM.

We’ve covered whether CPU inference makes economic sense at all and which CPU to pair with a GPU. This page answers the narrower buying question: if you are going to generate tokens on a CPU in 2026, which ones are actually worth money?

The 60-second ranking

CPU	Channels × speed	Theoretical bandwidth	Measured inference	Street price (Aug 2026)
Used AMD EPYC 9554 (64c, Genoa)	12 × DDR5-4800	460.8 GB/s	~63 tok/s, 30B MoE Q4	~$1,925 used
Used AMD EPYC 9354 (32c, Genoa)	12 × DDR5-4800	460.8 GB/s	14.2 tok/s, Llama 3.1 8B Q4_K_M	~$1,999 used (9354P)
Intel Xeon 6 “Granite Rapids”	12 × MRDIMM-8800	~844.8 GB/s (spec math)	AMX: ~2× prefill vs no-AMX	new-server money
AMD Ryzen AI Max+ 395 (Strix Halo)	4 × LPDDR5X-8000	256 GB/s	~100 tok/s, 30B MoE (iGPU)	$3,499 in a mini PC
AMD Ryzen 9 9950X (16c, Zen 5)	2 × DDR5-5600	89.6 GB/s	~11–12 tok/s, 8B dense Q4	~$485
Threadripper PRO 7000	8 × DDR5-5200	~332.8 GB/s	(see below — skip it)	$2,500+ CPU alone

Every number in that table has a source at the bottom of the page. Now the reasoning.

Why bandwidth outranks cores (the napkin math)

Decoding one token means reading every active parameter once. A dense 8B model at Q4_K_M is roughly 5GB of weights, so a desktop with ~90 GB/s of real memory bandwidth tops out around 90 ÷ 5 ≈ 18 tok/s theoretical — and lands at 11–12 tok/s in practice on a Ryzen 9 9950X, per LocalScore-style community benchmarks. The same math says a dense 70B (~40GB at Q4) is capped near 2 tok/s on that desktop. No core count fixes this; the cores spend most of each token stalled on RAM.

That’s also why 2026’s sparse MoE models changed who should even consider CPU inference. A 30B-class MoE with ~3B active parameters reads ~2GB per token instead of 17GB, which is how an EPYC Genoa socket pushes ~63 tok/s on Qwen3.6-35B-A3B-class models — genuinely faster than most people read.

One warning before you extrapolate your own napkin math: it fails in both directions. The llama.cpp tracker for Qwen3-Coder-Next (80B-A3B) documents CPU decode landing about 5× slower than the bandwidth math predicts — expert routing overhead and memory layout matter, and a dense 32B measured just 3.54 tok/s on a 12-core desktop part where naive math promised more. Trust measured numbers over arithmetic, including ours.

Best value: used AMD EPYC Genoa (9354/9554)

The 4th-gen EPYC 9004 socket is the bandwidth king you can actually buy used: 12 channels of DDR5-4800, 460.8 GB/s per socket — about half an RTX 3090, attached to up to terabytes of RAM. eBay listings in August 2026 have the 64-core EPYC 9554 at ~$1,925 and the 32-core 9354P at ~$1,999 — the 64-core is currently cheaper than the 32-core, an artifact of datacenter fleets dumping 9554s. Take the deal.

What it delivers, measured: ~63 tok/s on a 30B-class MoE at Q4, and 14.2 tok/s on Llama 3.1 8B Q4_K_M. And here’s the counterintuitive part of that second number — the Markaicode benchmark that produced it found the EPYC 9354 peaked at 12 threads, not 32 or 64. Past the point where memory bandwidth saturates, extra threads add contention and lose tokens per second. Benchmark your own thread count before trusting defaults:

$ llama-bench -m qwen3.6-35b-a3b-Q4_K_M.gguf -t 8,12,16,24,32
# expected shape of the result on a 12-channel EPYC:
# t=8 and t=12 land within a few percent of each other,
# t=24/32 come back SLOWER — bandwidth saturated, contention rising

If your -t $(nproc) run is slower than half your cores, that’s not a bug. It’s the bandwidth wall announcing itself.

The problem you’ll actually hit in 2026 isn’t the CPU — it’s the RAM. A 384GB DDR5 RDIMM kit that made these builds famous at ~$1,500 in 2025 now runs $5,200+, more than the CPU, motherboard, and chassis combined. Our CPU-vs-GPU cost breakdown works that math in full: ~$127 per tok/s of capex versus ~$12 for a used 3090. The EPYC build only wins when the model cannot fit in consumer VRAM — think 100B+ MoE weights — or when you already own the RDIMMs.

Best desktop CPU: AMD Ryzen 9 9950X

If CPU inference is a secondary use for a PC that mostly does other work, buy the Ryzen 9 9950X and spend the savings on RAM before prices climb further. Sixteen Zen 5 cores with full AVX-512 make it the fastest practical consumer chip for llama.cpp, and the street price has slid from a $649 launch MSRP to about $485–$511 as of late July 2026.

Temper expectations to the platform, not the chip: dual-channel DDR5-5600 is 89.6 GB/s theoretical, ~65–75 GB/s measured. That buys ~11–12 tok/s on an 8B dense model, 15–20 tok/s on a 13B Q4, and single digits on anything dense and large. The interesting ceiling is MoE: an 80B-A3B runs around 8 tok/s — slow for chat, fine for overnight batch jobs. There is no desktop CPU that meaningfully beats this, because Intel’s Core Ultra and AMD’s entire AM5 line share the same two-channel wall. Don’t pay flagship-plus prices chasing CPU-side tok/s on a desktop socket; it’s bandwidth-capped by design. (More RAM does help you load bigger MoE models — see how much system RAM you need.)

The turnkey pick: Ryzen AI Max+ 395 (Strix Halo)

Strictly speaking the Ryzen AI Max+ 395 wins by cheating: its 256 GB/s of quad-channel LPDDR5X-8000 feeds an iGPU that llama.cpp’s Vulkan backend uses instead of the CPU cores. But if your question is “what processor should my inference box have,” this is the honest 2026 answer for most people — ~100 tok/s on 30B-class MoE models and 35–40 tok/s on GPT-OSS 120B from a laptop-class part, with up to 128GB of unified memory and no discrete GPU.

The catch is what happened to pricing. The GMKtec EVO-X2 launched at $1,999 with 128GB; the same config now debuts at $3,499 after the mid-2026 memory squeeze. At $1,999 it was the obvious pick. At $3,499 it competes with a used EPYC on capacity and loses to a used 3090 on speed-per-dollar — buy it for the form factor and power draw, not the economics.

The prefill specialist: Intel Xeon 6 with AMX

Intel’s AMX matrix units are real, and the numbers are repeatable: ~57 tok/s with AMX versus ~28 tok/s without on small quantized models, and ~80 tok/s on 4-bit builds on a Xeon Gold 6530. Granite Rapids pairs that with MRDIMM-8800 support — 12 channels at 8800 MT/s is ~844.8 GB/s of theoretical socket bandwidth, closing in on discrete-GPU territory. Intel’s own comparison claims a Xeon 6767P delivers up to 1.4× the vLLM throughput of a 128-core EPYC 9755 — vendor benchmark, salt accordingly.

Where AMX genuinely shines is prefill (prompt ingestion), which is compute-bound where decode is bandwidth-bound. Feeding 30k-token prompts to a RAG pipeline? AMX pulls ahead of same-bandwidth EPYC. For chat-style decode, bandwidth still rules, and used Genoa costs a fraction of a new Granite Rapids platform. There’s no meaningful used market for Xeon 6 yet; check back in 2027. If you just want to see AMX-class CPU inference without buying anything, we ran Gemma 4 26B on a 13-year-old Xeon — the contrast is instructive.

Skip: Threadripper, and (for inference) Apple’s CPU cores

Threadripper PRO 7000 gives you 8 channels of DDR5-5200 — ~332.8 GB/s theoretical — for a CPU price starting around $2,500 before the workstation board and RDIMMs. That’s less bandwidth than a $1,925 used EPYC at more than twice the total platform cost, and the non-PRO Threadrippers drop to 4 channels. Threadripper is a fine workstation chip; as an inference purchase it’s dominated from both sides.

Apple Silicon is the reverse case: an M4 Max’s 546 GB/s of unified memory outruns every AMD socket here — but inference runs on the GPU cores via Metal/MLX, and you’re buying a whole Mac. If a Mac is on your shortlist anyway, that article has the numbers.

How to buy this in August 2026

The DRAM crisis is the buying condition, not a footnote. A 64GB DDR5 kit runs about $900 retail, entry 32GB kits start near $350, and TrendForce forecasts another 13–18% contract-price rise in Q3 with shortages projected into late 2027. Three consequences:

1. If you already own 128GB+ of DDR5, your desktop is a free MoE inference box. Set it up before deciding to spend anything.
2. If you’re buying RAM to do CPU inference, price the RAM first. It’s now the dominant line item in every build on this page, and it’s the reason the used-3090 GPU path wins most cost comparisons.
3. If you need big-model capacity occasionally, rent it. A cloud GPU hour on RunPod costs less than the monthly interest on a $5,200 RDIMM kit; our rent-vs-buy breakdown covers when owning flips back ahead. For the models that do fit in one card, start with the GPU buying guide instead — and if you’re eyeing 70B-class, the 32GB VRAM tier guide shows what a single card handles now.

For coding-assistant workloads specifically — where an agent hammers prefill with long contexts — our sister site aicoderscope.com covers which local models hold up in editors, and aifoss.dev tracks the FOSS serving stacks these CPUs run.

FAQ

What is the single best CPU for LLM inference in 2026? A used AMD EPYC 9554 (~$1,925) attached to all 12 memory channels is the best raw-value inference CPU — 460.8 GB/s and ~63 tok/s on 30B-class MoE models. It only makes sense if you can stomach current RDIMM prices or already own the RAM.

How many cores do I need? Fewer than you think. Decode saturates memory bandwidth at modest thread counts — an EPYC 9354 measured fastest at 12 threads on an 8B model, and more threads ran slower. Buy channels, not cores.

Is CPU inference fast enough to replace a GPU? For dense models, no — a used RTX 3090 is roughly 10× more tokens per hardware dollar. For sparse MoE models that don’t fit in 24GB of VRAM, a 12-channel EPYC is the cheapest way to run them at usable speed.

Do Intel CPUs beat AMD for LLM inference? On decode, whoever has more bandwidth wins, and used 12-channel EPYC is far cheaper per GB/s. Intel’s AMX gives Xeon 6 a real edge on prefill-heavy workloads (long prompts, RAG), at new-server prices.

Does RAM speed matter more than CPU model? On desktops, yes. Going from DDR5-4800 to DDR5-6400 on dual-channel raises the bandwidth ceiling ~33%, which flows almost directly into tok/s. That upgrade usually beats any CPU swap on the same socket.

Recommended Gear

• AMD Ryzen 9 9950X — the desktop ceiling for CPU inference (~$485, ~11–12 tok/s on 8B dense)
• AMD EPYC 9554 — 64 cores, 460.8 GB/s socket, ~$1,925 used; the capacity play

Sources

• AMD EPYC 9004 series: 12-channel DDR5-4800, 460.8 GB/s per socket — AMD
• Ollama CPU benchmark: EPYC 9354, Llama 3.1 8B Q4_K_M, 14.2 tok/s at 12 threads — Markaicode
• Best CPU for LLMs in 2026 (9950X ~11–12 tok/s CPU-only) — CodeX/Medium
• LLM inference benchmarks with llama.cpp on AMD EPYC 9554 — ahelpme.com
• Qwen3-Coder-Next CPU inference ~5× slower than bandwidth math predicts — llama.cpp issue #19480
• Used AMD EPYC 9554 listings, ~$1,925 — eBay
• Ryzen 9000 series price cuts: 9950X street pricing — VideoCardz
• Ryzen 9 9950X Amazon price history (~$500, late July 2026) — camelcamelcamel
• Running local LLMs on a Strix Halo laptop (30B MoE ~100 tok/s, GPT-OSS 120B 35–40 tok/s) — BogdanVarlamov.com
• Intel AMX CPU inference: ~57 tok/s with AMX vs ~28 without — OpenMetal
• Maximizing LLM performance on Intel CPUs (Xeon Gold 6530, ~80 tok/s at 4-bit) — Plain Concepts
• Accelerating vLLM inference: Xeon 6767P vs EPYC 9755 (vendor claim, up to 1.4×) — Intel
• Granite Rapids: MRDIMM-8800 support and AMX — Wikipedia
• RAM shortages until 2028: 64GB DDR5 at ~$900 retail — TweakTown
• Memory shortages to last until at least Q4 2027; entry 32GB DDR5 kits from ~$350 — Wccftech
• RAM price tracking 2026 (TrendForce Q3 +13–18% contract forecast) — Tom’s Hardware
• Threadripper 7000: up to 8-channel DDR5-5200 — AMD
• CPU-only inference thread-count and bandwidth benchmarks — llamafile discussion #450

Last updated August 14, 2026. Prices and specs change; verify current rates before purchasing.