Ordered by predicted speed. Bandwidth sets decode rate; memory decides whether it runs at all.
Every figure here is computed, not measured. Fit is weights at each quantisation against device memory, with a 25% allowance for the KV cache, activations and the OS. Decode rate is the memory-bandwidth roofline at 70% efficiency. Nobody has run this model on these devices.
Device
Memory
Bandwidth
Best quality
Weights
tok/s
Fastest
NVIDIA GB200 Grace Blackwell Superchip
372.0 GB
16000.0 GB/s
bf16
141.11 GB
~79.4
~317.5 q4
AMD Instinct MI355X
288.0 GB
8000.0 GB/s
bf16
141.11 GB
~39.7
~158.7 q4
NVIDIA B200
180.0 GB
7700.0 GB/s
fp8
70.55 GB
~76.4
~152.8 q4
Google TPU7x (Ironwood)
192.0 GB
7380.0 GB/s
bf16
141.11 GB
~36.6
~146.4 q4
AMD Instinct MI325X
256.0 GB
6000.0 GB/s
bf16
141.11 GB
~29.8
~119.1 q4
AMD Instinct MI300X
192.0 GB
5300.0 GB/s
bf16
141.11 GB
~26.3
~105.2 q4
NVIDIA H200 SXM
141.0 GB
4800.0 GB/s
fp8
70.55 GB
~47.6
~95.2 q4
NVIDIA H100 NVL
94.0 GB
3900.0 GB/s
q6
52.92 GB
~51.6
~77.4 q4
NVIDIA H100 SXM
80.0 GB
3350.0 GB/s
q6
52.92 GB
~44.3
~66.5 q4
AMD Instinct MI250X
128.0 GB
3200.0 GB/s
fp8
70.55 GB
~31.7
~63.5 q4
Google TPU v5p
95.0 GB
2765.0 GB/s
fp8
70.55 GB
~27.4
~54.9 q4
NVIDIA A100 80GB SXM
80.0 GB
2039.0 GB/s
q6
52.92 GB
~27.0
~40.5 q4
NVIDIA H100 PCIe
80.0 GB
2000.0 GB/s
q6
52.92 GB
~26.5
~39.7 q4
AMD Instinct MI210
64.0 GB
1600.0 GB/s
q5
44.1 GB
~25.4
~31.7 q4
NVIDIA L40S
48.0 GB
864.0 GB/s
q4
35.28 GB
~17.1
~17.1 q4
Apple M3 Ultra
512.0 GB
819.0 GB/s
bf16
141.11 GB
~4.1
~16.3 q4
Apple M2 Ultra
192.0 GB
800.0 GB/s
bf16
141.11 GB
~4.0
~15.9 q4
Apple M4 Max
128.0 GB
546.0 GB/s
fp8
70.55 GB
~5.4
~10.8 q4
Apple M1 Max
64.0 GB
400.0 GB/s
q5
44.1 GB
~6.3
~7.9 q4
Apple M2 Max
96.0 GB
400.0 GB/s
fp8
70.55 GB
~4.0
~7.9 q4
Apple M3 Max
128.0 GB
400.0 GB/s
fp8
70.55 GB
~4.0
~7.9 q4
Apple M4 Pro
64.0 GB
273.0 GB/s
q5
44.1 GB
~4.3
~5.4 q4
Where it runs
Taken from the card's availability section.
Platform
Id on that platform
Hugging Face
NousResearch/Hermes-3-Llama-3.1-70B
Verified benchmark evidence
Each row was checked against its source by a reviewer, and carries the model identifier as actually evaluated — which is not always the same as this card's.