The best CPU for AI and local LLMs is the one that feeds your memory fastest, not the one with the highest gaming score. For a typical desktop with a GPU, the AMD Ryzen 9 9950X is the best all-round pick thanks to 16 Zen 5 cores, full 512-bit AVX-512 and the AM5 upgrade path. If you want to run large models mostly on the CPU, Threadripper 9000 doubles memory bandwidth with four DDR5 channels. And if you want one chip that runs 100B-plus models without a graphics card, the Ryzen AI Max+ 395 and its 192GB successor are in a class of their own. Here is how each option stacks up, and when the CPU barely matters at all.
Our Top Picks
Top 3 picks at a glance
Verdict at a glance
- Best overall: AMD Ryzen 9 9950X. Strong llama.cpp results, AVX-512, AM5 socket that AMD says will take Zen 6.
- Best for CPU-heavy inference and MoE offload: AMD Threadripper 9960X or 9970X. Four DDR5-6400 channels for about 205 GB/s theoretical.
- Best for maximum CPU bandwidth: Threadripper PRO 9000WX on WRX90, with eight memory channels.
- Best all-in-one AI chip: AMD Ryzen AI Max+ 395 (128GB) or Ryzen AI Max+ PRO 495 (192GB).
- Best value: Intel Core Ultra 7 270K Plus, praised by Puget Systems and Gamers Nexus for performance per dollar.
- Best if you also game hard: Ryzen 9 9950X3D. Gaming chops of an X3D chip plus 16 cores.
- Skip for AI: Ryzen 9 9950X3D2. Its extra cache helps some workloads, but not token generation.
Does the CPU matter for local AI?
It depends where the model lives.
If the whole model fits in your graphics card’s VRAM, the CPU mostly handles tokenizing, scheduling and the operating system. Almost any modern six-core chip will do, and spending more on the CPU will not raise tokens per second in any meaningful way.
The CPU starts to matter in three cases:
- CPU-only inference. No GPU, or a GPU too small to help. Here, memory bandwidth decides generation speed and core count plus vector instructions decide prompt processing speed.
- Partial offload. The model is too big for VRAM, so some layers run on the CPU. Speed drops sharply, and the CPU’s memory system becomes the bottleneck.
- MoE expert offload. Mixture-of-experts models like gpt-oss-120b or Qwen 30B-A3B variants can keep attention layers on the GPU and expert weights in system RAM, using llama.cpp’s n-cpu-moe option. This is the most practical way to run big MoE models on a gaming PC, and it is very sensitive to RAM bandwidth.
There is also the APU route: chips like the Ryzen AI Max+ 395 put a large GPU and a large pool of fast memory on the same package, which blurs the line between CPU and GPU entirely.
Memory bandwidth is the real spec
Token generation streams the model’s active weights from memory for every token. That makes bandwidth the first number to check.
| Platform | Memory channels | Theoretical bandwidth | Max memory |
|---|---|---|---|
| AM5 (Ryzen 9000) or LGA 1851 (Core Ultra 200S) | 2 x DDR5 | About 102 GB/s at DDR5-6400; about 90 GB/s at DDR5-5600 | Up to 192GB to 256GB depending on board |
| TRX50 (Threadripper 9000) | 4 x DDR5-6400 RDIMM | About 205 GB/s | Up to 1TB |
| WRX90 (Threadripper PRO 9000WX) | 8 x DDR5-6400 RDIMM | About 410 GB/s | Up to 2TB |
| Ryzen AI Max+ 395 (Strix Halo) | 256-bit LPDDR5X-8000 | 256 GB/s | 128GB soldered |
| Ryzen AI Max+ PRO 495 (Gorgon Halo) | 256-bit LPDDR5X-8533 | About 273 GB/s | 192GB soldered |
Bandwidth figures for desktop sockets are our own arithmetic from channel count and transfer rate; real-world results are lower. For context, an RTX 4090 has about 1,008 GB/s, so even an eight-channel workstation is well short of a single high-end GPU.
Comparison table: the best CPUs for local LLMs
| CPU | Cores / threads | AVX-512 | Memory | Best for | Price tier |
|---|---|---|---|---|---|
| Ryzen 9 9950X | 16 / 32 | Yes, full 512-bit | Dual-channel DDR5 | GPU builds plus capable CPU inference | Upper mid-range |
| Ryzen 9 9950X3D | 16 / 32 | Yes | Dual-channel DDR5 | Gaming and AI on one PC | Premium |
| Threadripper 9960X | 24 / 48 | Yes | Quad-channel DDR5-6400 | Best-value bandwidth upgrade | Workstation |
| Threadripper 9970X | 32 / 64 | Yes | Quad-channel DDR5-6400 | Balanced HEDT for AI and rendering | Workstation |
| Threadripper 9980X | 64 / 128 | Yes | Quad-channel DDR5-6400 | Fastest prompt processing on CPU | High-end workstation |
| Ryzen AI Max+ 395 | 16 / 32 plus 40-CU GPU | Yes | 128GB LPDDR5X unified | Large models without a discrete GPU | Sold in complete systems |
| Core Ultra 7 270K Plus | Hybrid P-core and E-core | No | Dual-channel DDR5 | Value builds focused on a GPU | Mid-range |
| Core Ultra 9 285K | 24 / 24 | No | Dual-channel DDR5 | Existing Intel users | Upper mid-range |
1. AMD Ryzen 9 9950X: best overall
The 9950X is the sensible center of gravity for a local AI desktop. Its 16 Zen 5 cores handle prompt processing well on the CPU, and Zen 5’s full 512-bit AVX-512 data path is something Phoronix singled out as a real advantage in llama.cpp workloads. Broad Linux testing has shown AMD’s Zen 5 chips consistently ahead of the Core Ultra 9 285K in llama.cpp, partly because Intel’s current desktop chips lack AVX-512.
Set expectations for CPU-only work, though. LocalScore results cited by PopularAI put the 9950X at about 11 to 12 tokens per second on an 8B model running purely on the CPU, which is usable for small assistants but not for 70B models. Where the 9950X really earns its place is as the partner to a big GPU, with 64GB to 128GB of fast DDR5 for MoE expert offload.
Reasons to pick: AVX-512 at full width; plenty of cores for prefill; AM5 should accept Zen 6 chips when AMD launches them in 2027.
Reasons to skip: dual-channel memory caps CPU generation speed; AMD officially rates it for DDR5-5600, so higher speeds are overclocking.
The catch: with a GPU that holds your whole model, a cheaper Ryzen 7 will deliver the same tokens per second.
2. Threadripper 9960X and 9970X: best for CPU-heavy inference
AMD’s Threadripper 9000 chips, launched in July 2025, run on TRX50 boards with four channels of DDR5-6400 RDIMM, up to 1TB of memory and up to 80 PCIe 5.0 lanes. That doubles the theoretical bandwidth of any desktop socket to about 205 GB/s and gives you room for several GPUs.
Puget Systems’ llama.cpp results contain the most useful detail for buyers: in token generation, the 64-core 9980X was fastest, but the 24-core 9960X came in ahead of the 32-core 9970X. That shows generation depends far more on memory bandwidth than core count, so the cheapest Threadripper gets you most of the benefit.
Reasons to pick: double the memory bandwidth of AM5; huge memory capacity for big MoE models; lanes for multiple GPUs.
Reasons to skip: expensive platform with RDIMM memory, which has been hit hard by the 2026 memory shortage; high power draw, with LanOC measuring around 500W at the wall under heavy load on the 9970X system.
The catch: about 205 GB/s is still roughly a fifth of a single RTX 4090, so a Threadripper alone will not match even a mid-range GPU for models that fit in VRAM.
3. Threadripper 9980X: fastest prompt processing on a CPU
If you process long prompts or batch jobs on the CPU, core count matters again. Puget Systems found the 9980X the fastest CPU for llama.cpp prompt processing, ahead of the 7980X by 20 percent and the 9970X by 48 percent. Phoronix’s launch review showed it topping LocalScore runs on 8B, 14B and 32B models, while the 9970X sometimes won on very small models.
Reasons to pick: 64 cores for prefill and mixed workloads like rendering and compiling; same four-channel platform.
Reasons to skip: generation speed barely beats the cheaper 9960X; flagship price.
The catch: for local chat, the money is better spent on a GPU with more VRAM.
4. Threadripper PRO 9000WX: maximum CPU bandwidth
The PRO 9000WX series on WRX90 boards supports eight memory channels and up to 2TB of RAM. At DDR5-6400 that works out to about 410 GB/s theoretical, approaching the bandwidth of an entry-level GPU and enough for very large MoE models to run at usable speeds on the CPU side. Puget Systems notes the PRO chips cost considerably more than non-PRO Threadripper at the same core count.
Reasons to pick: the most CPU memory bandwidth outside of servers; terabytes of capacity.
Reasons to skip: filling eight channels with RDIMMs in 2026 costs a fortune; overkill for most home users.
The catch: at this tier, a Mac Studio M5 Ultra or a pair of 32GB GPUs is often faster per token for less.
5. Ryzen AI Max+ 395 and PRO 495: best all-in-one AI chip
Strix Halo is technically a CPU, but it is better thought of as a whole AI computer on one package. The Ryzen AI Max+ 395 combines 16 Zen 5 cores with a 40-CU Radeon 8060S GPU and up to 128GB of LPDDR5X-8000 at 256 GB/s. Up to 96GB can be assigned to the GPU in BIOS, and AMD documents a Linux setting that raises this to about 120GB on a 128GB system.
Community llama.cpp results put gpt-oss-120b at roughly 49 to 54.5 tokens per second on this chip, close to NVIDIA’s DGX Spark. The 2026 Ryzen AI Max+ PRO 495, part of the Gorgon Halo refresh, raises the ceiling to 192GB with up to 160GB usable as VRAM, according to AMD.
Reasons to pick: runs 100B-plus MoE models with no discrete GPU; Windows and Linux; compact, efficient systems.
Reasons to skip: only sold soldered into complete systems; prompt processing is slow next to NVIDIA hardware; 128GB systems have roughly doubled in price since launch due to memory costs.
The catch: dense 70B models still crawl at 256 GB/s, so this is a MoE machine first.
6. Intel Core Ultra 7 270K Plus: best value
Intel’s Arrow Lake Refresh, sold as Core Ultra 200S Plus, brought faster interconnects and memory support, and launch reviews found its memory latency now roughly in line with Ryzen 9000. Puget Systems found the 270K Plus excellent in content creation, and Gamers Nexus highlighted its value: in one comparison the far pricier 9950X3D2 led it by just 2.6 percent.
For AI, the 270K Plus is a fine host for a GPU-centric build. It lacks AVX-512, which hurts CPU-only llama.cpp throughput versus Zen 5, so we would not choose it for heavy CPU inference.
Reasons to pick: strong performance per dollar; good all-round desktop chip.
Reasons to skip: no AVX-512; LGA 1851 is a dead-end socket because Nova Lake moves to LGA 1954.
The catch: great value today, but no upgrade path tomorrow.
7. Ryzen 9 9950X3D: best if you also game
If one PC has to be both a top gaming rig and a local AI workstation, the 9950X3D combines 16 Zen 5 cores with 3D V-Cache on one of its two chiplets. It performs like the 9950X in AI workloads and close to the best X3D chips in games. OpenBenchmarking prompt-processing results show it well ahead of the regular 9950X in one CPU BLAS test, though results vary by workload.
Reasons to pick: top-tier gaming plus 16 cores.
Reasons to skip: costs more than the 9950X for similar AI performance.
The catch: for pure gaming, the 8-core Ryzen 7 9850X3D is faster per dollar.
What about the Ryzen 9 9950X3D2 and Core Ultra 9 285K?
The 9950X3D2 Dual Edition stacks 3D V-Cache on both chiplets for 192MB of cache. Tom’s Hardware found it matches the Ryzen 7 9800X3D in games and is about 4 percent faster than the 9950X3D in multithreaded work, and Phoronix found real gains in some cache-sensitive workloads. LLM token generation is bandwidth-bound and does not benefit much from extra cache, so we would not pay the halo premium for AI.
The Core Ultra 9 285K is a capable chip, but without AVX-512 it trails Zen 5 in llama.cpp. If you already own one, keep it and spend on a GPU.
Wait for Zen 6 or Nova Lake?
Neither is coming in 2026. Tom’s Hardware reports both AMD’s Zen 6 desktop chips and Intel’s Nova Lake-S have slipped into 2027, with announcements expected around CES 2027. If you buy now, AM5 is the safer bet because AMD has said it will support Zen 6, while Nova Lake needs a new socket.
How to choose a CPU for local AI
- Decide where the model lives. If it fits in VRAM, buy a modest CPU and the biggest GPU you can. If not, prioritize memory bandwidth.
- Count memory channels before cores. Four channels beat more cores on two channels for generation speed.
- Prefer AVX-512 for CPU inference. Zen 5 chips run llama.cpp’s CPU paths faster.
- Buy enough RAM for MoE offload. 64GB is a good floor for a GPU build; 128GB lets you offload gpt-oss-120b-class experts.
- Run the fastest stable RAM. Public results show 17 to 18 percent CPU inference gains from DDR5-4800 to DDR5-6000, and one owner report showed MoE offload speed roughly tripling after enabling the DDR5-6000 XMP profile versus a very slow default.
Setup tips
- Match threads to physical cores. In llama.cpp, using all logical threads can be slower than using physical cores only. Try both.
- Use n-cpu-moe for big MoE models. Keep attention on the GPU and push expert weights to the CPU a few layers at a time until the model fits.
- Fill every memory channel. On Threadripper, populating only two of four channels halves your bandwidth.
- Enable EXPO or XMP. Default JEDEC speeds can leave a lot of bandwidth on the table.
Common mistakes
- Buying a flagship CPU for a GPU-only workload. If the model is in VRAM, the CPU barely matters.
- Running one memory stick. Single-channel memory halves bandwidth and token generation speed.
- Offloading dense layers instead of MoE experts. Dense layer offload is far slower; expert offload is the efficient path.
- Judging CPUs by gaming charts. Cache-heavy gaming chips do not necessarily generate tokens faster.
Products Mentioned in This Guide
These are the specific CPUs, the all-in-one Strix Halo system type and the memory kit size discussed in the picks above.
AMD Ryzen 9 9950X 16-Core Desktop Processor
As an Amazon Associate we earn from qualifying purchases at no extra cost to you.
Our best overall pick: 16 Zen 5 cores, full 512-bit AVX-512 for llama.cpp and an AM5 socket that AMD says will take Zen 6.
AMD Ryzen Threadripper 9960X
The best-value way to double memory bandwidth with four DDR5-6400 channels. In Puget Systems’ results it generated tokens faster than the 32-core 9970X.
MINISFORUM MS-S1 Max Mini Workstation, Ryzen AI Max+ 395, 128GB
A complete system built on the Ryzen AI Max+ 395, our best all-in-one AI chip, which runs 100B-plus MoE models without a discrete GPU.
Intel Core Ultra 7 270K Plus
The best value pick and a fine host for a GPU-centric build, though it lacks AVX-512 and sits on a socket with no upgrade path.
AMD Ryzen 9 9950X3D 16-Core Processor
The pick if one PC has to be both a top gaming rig and a local AI workstation, performing like the 9950X in AI workloads.
G.SKILL Trident Z5 Neo DDR5 64GB (2x32GB) 6000MT/s CL30
A 64GB dual-channel DDR5-6000 kit, the comfortable floor for MoE expert offload in a GPU build. Enable EXPO so it runs at its rated speed.
How we compared
We compared these using manufacturer specifications, independent lab measurements from the sources named above, and owner reports. We did not bench-test these units ourselves. Theoretical bandwidth numbers are calculated from channel count and memory speed and are labeled as such.
Sources
- Phoronix, AMD Threadripper 9980X and 9970X Linux benchmarks and Ryzen 9 9950X3D2 review
- Puget Systems, Threadripper 9000 content creation review, Threadripper 9000 vs PRO 9000WX, and Ryzen 9 9950X3D2 review
- OpenBenchmarking.org, llama.cpp test profile results
- PopularAI, best CPU for running local LLMs and RAM speed analysis
- Tom’s Hardware, Ryzen 9 9950X3D2 review, Threadripper 9000 launch, and Zen 6 and Nova Lake delay report
- Gamers Nexus, Ryzen 9 9950X3D2 review
- LanOC Reviews, Threadripper 9980X and 9970X review
- AMD, Ryzen AI Max+ 395 technical articles and Ryzen AI Max 400 announcement
- AIMultiple, DGX Spark alternatives
- llama.cpp GitHub discussions on MoE offload
Frequently Asked Questions
Is CPU or GPU more important for local LLMs?
The GPU, by a wide margin, if the model fits in VRAM. The CPU becomes important for CPU-only inference and for running models too big for your graphics card.
How fast is CPU-only LLM inference?
On a Ryzen 9 9950X, LocalScore results cited by PopularAI show about 11 to 12 tokens per second for an 8B model. Larger dense models get much slower on dual-channel memory.
Does AVX-512 matter for AI?
Yes, for CPU inference. llama.cpp uses wide vector instructions, and Phoronix credits Zen 5’s full 512-bit data path for strong results.
Is Threadripper worth it for local AI?
Only if you plan to run large models partly or fully on the CPU, or need lanes for several GPUs. For a single-GPU setup, a desktop chip plus a bigger GPU is better value.
Is the Ryzen AI Max+ 395 a CPU or a GPU?
Both. It is an APU with 16 CPU cores and a 40-CU GPU sharing up to 128GB of fast memory, which is why it can run large models without a discrete card.
How much RAM do I need for local AI?
32GB is a minimum for a GPU build, 64GB is a comfortable target, and 128GB or more makes sense if you want to offload large MoE models to system memory.
Should I wait for Zen 6 or Nova Lake?
Both are expected in 2027, not 2026. If you need a machine now, AM5 gives you an upgrade path to Zen 6.
