This page is for a local-LLM shopper picking an R9700 or RX 7900-class card versus an RTX 5090, 4090, 3090, 5080, 5070 Ti, or 5060 Ti 16GB. Plain language. No fanboying. Specs cite the vendor page. Speed that is ours says measured. Speed that is not ours stays empty.
How we test
— Measured cells cite official amdai-bench/run@1 cards. Estimates stay labeled. Q3 is estimate-only. Q1/Q2 stay off.
Measured peak system RAM on the local desktop: 42.9 GB (Linux GiB labeled GB). Measured, pending re-verify after RAM fix.
Breakdown
Offloaded weights ·
0 GiB
Separable embed in RAM ·
0 GiB
KV in RAM at selected ·
0 GiB
KV in RAM at full context ·
0 GiB
Runtime / backend ·
0.85 GiB
OS / desktop baseline ·
8.00 GiB
Headroom ·
16.0 GiB
Where each camp wins
Where Radeon wins
VRAM class, 32 GB, two cards, our cards
Radeon wins the VRAM-per-dollar class we actually shop: 32 GB on the R9700, 24 / 20 / 16 GB on the RX 7900 / 7800 sheets, and two-card layouts we have measured on the local desktop. Buttons say Check price — we do not print a figure.
NVIDIA wins the software day-one: CUDA is the default for vLLM and a long tail of engines, and more models and apps land there first. That is ecosystem maturity, not a tok/s we measured on an RTX card.
VRAM, memory bandwidth, and board power. Each cell links the vendor spec page.
TBP is AMD’s label; TGP is NVIDIA’s. Buttons say Check price — we do not show a live or stale figure.
GPU directory:
/gpus.
Cited vendor VRAM, bandwidth, and board power for the Radeon and RTX cards on this page.
Newegg's R9700 offers have been marketplace sellers (checked 2026-10-30), so that seller's return policy applies, not Newegg's. Open the listing's Warranty & Returns tab before you buy.
Return terms vary by seller and item. Newegg's Hassle-Free page covers only Sold-and-Shipped-by-Newegg listings; marketplace items use the seller's policy. See the listing's Warranty & Returns tab.
Return terms vary by seller and item. Newegg's Hassle-Free page covers only Sold-and-Shipped-by-Newegg listings; marketplace items use the seller's policy. See the listing's Warranty & Returns tab.
Return terms vary by seller and item. Newegg's Hassle-Free page covers only Sold-and-Shipped-by-Newegg listings; marketplace items use the seller's policy. See the listing's Warranty & Returns tab.
Check price · no live or invented prices
NVIDIA GeForce RTX 5090
Official NVIDIA product page plus retailer search. No live or invented price.
Return terms vary by seller and item. Newegg's Hassle-Free page covers only Sold-and-Shipped-by-Newegg listings; marketplace items use the seller's policy. See the listing's Warranty & Returns tab.
Check price · no live or invented prices
NVIDIA GeForce RTX 4090
Official NVIDIA product page plus retailer search. No live or invented price.
Return terms vary by seller and item. Newegg's Hassle-Free page covers only Sold-and-Shipped-by-Newegg listings; marketplace items use the seller's policy. See the listing's Warranty & Returns tab.
Check price · no live or invented prices
NVIDIA GeForce RTX 3090
Official NVIDIA compare page plus retailer search. No live or invented price.
Return terms vary by seller and item. Newegg's Hassle-Free page covers only Sold-and-Shipped-by-Newegg listings; marketplace items use the seller's policy. See the listing's Warranty & Returns tab.
Check price · no live or invented prices
NVIDIA GeForce RTX 5080
Official NVIDIA 50-series page plus retailer search. No live or invented price.
Return terms vary by seller and item. Newegg's Hassle-Free page covers only Sold-and-Shipped-by-Newegg listings; marketplace items use the seller's policy. See the listing's Warranty & Returns tab.
Check price · no live or invented prices
NVIDIA GeForce RTX 5070 Ti
Official NVIDIA 50-series page plus retailer search. No live or invented price.
Return terms vary by seller and item. Newegg's Hassle-Free page covers only Sold-and-Shipped-by-Newegg listings; marketplace items use the seller's policy. See the listing's Warranty & Returns tab.
Check price · no live or invented prices
NVIDIA GeForce RTX 5060 Ti 16GB
16 GB SKU. NVIDIA also lists an 8 GB 5060 Ti — confirm the listing. No live or invented price.
Return terms vary by seller and item. Newegg's Hassle-Free page covers only Sold-and-Shipped-by-Newegg listings; marketplace items use the seller's policy. See the listing's Warranty & Returns tab.
As an Amazon Associate I earn from qualifying purchases. Some Newegg links use Rakuten Advertising and some B&H links use B&H’s affiliate program when those IDs are configured. Micro Center and AMD Store links are plain “also available at” links — no tracking. Buttons say “Check price” and never invent or live-quote a price.
Full disclosure.
Speed
R9700 decode is measured on the local desktop (amdai-bench/run@1, llama.cpp Vulkan). Any NVIDIA speed figure on this site must be third-party, linked to the exact source, and labeled not apples-to-apples (different engine, quant, or context). We have no ingested cited NVIDIA 35B number for these SKUs, so those cells say “no cited number yet.” We will not estimate one.
HIP is the ROCm compile when rocminfo lists the gfx. gfx1201 “no binary for GPU” reports live on the problems tracker. Try Vulkan first if HIP configure is a mess.
NVIDIA
HIP is not an NVIDIA path. CUDA is the matching compile.
First-party Flash try on the house Pro rig — cited pair ~32.6 tok/s versus llama.cpp ~16 tok/s at 262K on Qwen 3.8 Flash UD-Q6_K_XL. Not the 35B overnight matrix.
NVIDIA
No house Strata figure on an RTX card. We will not invent one.
Advanced / experimental here. Default pip install is CUDA. ROCm is a pin (AMD 10.0 supported combo is still 0.27 + PyTorch 2.13; community 0.31 is a try). No invented local desktop 0.31 tok/s.
NVIDIA
This is where NVIDIA wins on day-one maturity: CUDA vLLM is the upstream default. Broader engine and model coverage on launch week. Still not our measured 35B path.
Convenience wrappers. gfx1201 is still cited pain in Ollama. LM Studio may download a CUDA llama.cpp on a Radeon if you are not watching the runtime picker. Prefer a Vulkan llama.cpp you built.
NVIDIA
Day-one CUDA runtimes are the usual path in both apps. We have no house tok/s. Confirm a GPU meter moves before you trust the wrapper.
Use case first, then VRAM, then software. Size vs speed still changes the pick on
setup by budget
and the
calculator.
FAQ:
/faq.
Daily-driver 35B at long context
One R9700, if you want the path we measured.
House daily driver is Qwen3.6-35B-A3B Unsloth UD-Q5_K_XL on one R9700, llama.cpp Vulkan, 262144 context, f16 KV. It stays in 32 GB VRAM on the overnight cards. An RTX 5090 is the same 32 GB class on paper; we have no cited 35B llama.cpp tok/s for it. A 16 GB RTX 5080 / 5070 Ti / 5060 Ti 16GB or RX 7800 XT does not hold that Q5 at full context in VRAM — expect RAM offload and a slower decode.
32 GB (R9700 or RTX 5090) is the class that holds the house 35B Q5 at 262K without offload on our calculator / cards. 24 GB (7900 XTX, RTX 4090, RTX 3090) is the next bracket — Q4 may fit; Q5 at long context needs the calculator. 16 GB cards (7800 XT, RTX 5080, 5070 Ti, 5060 Ti 16GB) are a different job: smaller listed files or RAM offload. We do not invent a “fits” claim past the calculator and the vendor VRAM label.
Minimum system RAM needed
32 GB kit
Estimate · 29.2 GiB raw · 32K context
RAM needed at full context
48 GB kit
Estimate · 33.6 GiB raw · 256K full context
Estimate. Offloaded weights + KV in RAM + runtime + OS/desktop + headroom, rounded up to a 32 GB DIMM kit.
A second R9700 is measured. A second RTX card is not, on this site.
Two R9700s on the local desktop: a split of the same 35B is often slower than one card at short context (UD-Q5 4K is 102.39 tok/s on one vs 82.78 on the split). Two copies serve two requests — that is concurrency, not 2× one chat. Desktop 2 (dual R9700 + 192 GB DDR5) is planned, not measured. We have no cited two-RTX 35B llama.cpp figure. Do not assume tensor parallel is free speed.
Linux if you want our measured Radeon path. Windows CUDA if you need day-one NVIDIA apps.
Ubuntu LTS on AMD’s current ROCm matrix is the path we write. llama.cpp Vulkan is the honest Windows path for tokens on Radeon. Full ROCm userspace is a Linux product. ROCm 10.1 lists WSL2 as a technical preview — not a measured local desktop daily driver. NVIDIA’s CUDA stack is the broader Windows day-one story (vLLM, more app binaries). We do not have Windows tok/s on either camp.
Email only — no account. Double opt-in. About once a week, and only when run cards, guides, ROCm problems, or Stack Lab actually changed. Not a daily drip.
Issues may include one labeled Sponsored mention.
Slots and rates:
Sponsor.