Builds / AMD Ryzen AI Max+ 395 (Radeon 8060S iGPU, gfx1151)
Strix Halo (Ryzen AI Max+ 395, 128 GB): Nemotron 3.5 Bolt E4B at 55.5 tok/s on llama.cpp, 96.4 on strix-halo-engine-lab EXL3 with MTP drafter
Last updated:
- GPU(s)
- AMD Ryzen AI Max+ 395 (Radeon 8060S iGPU, gfx1151)
Amazon Β· eBay (used) Β· Newegg - VRAM / unified memory (GB)
- 128
- CPU
- AMD Ryzen AI Max+ 395
Amazon Β· eBay (used) Β· Newegg - Motherboard
- not stated
- RAM
- 128 GB unified
- PSU
- not stated
- Other
- strix-halo-engine-lab (ROCm/HIP, gfx1151, gemma4 family) vs llama.cpp 74a4fb00 reference, same machine and session; greedy decoding, 256-token English chat
- Reported total cost
- not stated
- Power
- not stated
Measured speed (generation tokens/sec)
| Model | Quant | Engine | tok/s | Prompt processing (tok/s) | Details |
|---|---|---|---|---|---|
| Nemotron 3.5 Bolt E4B legacy | Q4_K_M (NVIDIA official 4GB GGUF, 6.18 GB) | llama.cpp 74a4fb00 | 55.5 | 1,881 @ 8K-token prompt est. TTFT @8k β 4.4 s (calculated from measured prefill) | context: 256-token chat decode, 1 requestπ Reference row measured in the same turn as the engine rows; 4 requests 193.3 total, 8 requests 279.9; prefill 2,077 @512, 1,376 @32K |
| Nemotron 3.5 Bolt E4B legacy | Q4_K_M (NVIDIA official 4GB GGUF, 6.18 GB) | strix-halo-engine-lab, profile default | 55.8 | 509 @ 8K-token prompt est. TTFT @8k β 16.1 s (calculated from measured prefill) | context: 256-token chat decode, 1 requestπ Drafting off; 4 requests 169.8 total, 8 requests 237.9 |
| Nemotron 3.5 Bolt E4B legacy | Q4_K_M (NVIDIA official 4GB GGUF, 6.18 GB) + MTP drafter | strix-halo-engine-lab, profile default | 91.9 | not stated | β‘ spec-decode (MTP + prompt lookup)context: 256-token chat decode, 1 requestπ Shipped C1 setting (MTP drafter + prompt lookup), token ids identical to drafting off; the int8-activation 'fast' profile reaches 99.5 (54.0 without drafting) |
| Nemotron 3.5 Bolt E4B legacy | EXL3 M4 4.65 bpw (hyukudan, 9.71 GB download, 2.70 GB read per token) | strix-halo-engine-lab, profile quality | 50.2 | 746 @ 8K-token prompt est. TTFT @8k β 11.0 s (calculated from measured prefill) | context: 256-token chat decode, 1 requestπ Drafting off; 4 requests 184.8 total, 8 requests 303.0; prefill 794 @512, 533 @32K |
| Nemotron 3.5 Bolt E4B legacy | EXL3 M4 4.65 bpw (hyukudan) + MTP drafter | strix-halo-engine-lab, profile quality | 96.4 | not stated | β‘ spec-decode (MTP + prompt lookup)context: 256-token chat decode, 1 requestπ Shipped C1 setting (MTP drafter + prompt lookup), token ids identical to drafting off |
Prompt processing = prefill speed (how fast the prompt is read) as reported by the source, with the prompt length it was measured at; blank = the source does not report it. est. TTFT @8k is calculated from the measured prefill speed (8,192 Γ· prefill tok/s), shown only when prefill was measured on a prompt of 512+ tokens. It is not a measurement and ignores load time and caching. measured TTFT is the source's own time-to-first-token. Why prefill matters for coding assistants.
Quant publisher's own measurements on the Hugging Face model card (hyukudan, card updated 2 Nov 2026), one measuring session of one engine build. The EXL3 M4 pack loads only in strix-halo-engine-lab. Quality on the same card: EXL3 M4 92.94% top-1 / mean KL 0.0367 vs BF16, official Q4_K_M 87.0% / 0.141. llama.cpp prefills about 2.5x faster than the engine today (the engine's prefill is being optimised).
Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.
Speculative-decoding rows on this page are not like-for-like with plain decode. See how much MTP and DFlash really speed things up (measured pairs).