Gemma 4 E2B
google/gemma-4-E2B-it
Google · 4 recorded configurations
Measured setup
Hardware
Runtime and tool
Quantisation
MTP draft depth
Every option opens a recorded setup. Where the corpus recorded nothing at this page's other settings, the option names the fields it switches with it. Allocation and runtime detail stay in the technical record below.
1× RTX 3090 · llama.cpp
UD-Q8_K_XL · MTP 3
PP · left axisTG · right axis
View exact measurements
| Context depth (tokens) | Fresh prefill PP (tok/s) | TG (tok/s) |
|---|---|---|
| 0 | 8,639.81 | 354.94 |
| 2,048 | 7,786.02 | 349.98 |
| 4,096 | 7,401.53 | 338.88 |
| 8,192 | 6,727.11 | 325.93 |
| 16,384 | 5,625.34 | 300.18 |
| 32,768 | 4,297.29 | 267.08 |
| 65,536 | 2,922.45 | 216.55 |
| 130,240 | 1,774.83 | 157.56 |
Concurrent-stream measurements
| Streams | Context depth (tokens) | Aggregate PP (tok/s) | Aggregate TG (tok/s) | TG / stream (tok/s) | Aggregate all (tok/s) |
|---|---|---|---|---|---|
| 2 | 8,192 | 6,323.39 | 478.75 | 239.38 | 1,227.71 |
| 3 | 8,192 | 5,831.64 | 538.88 | 179.63 | 1,432.69 |
| 4 | 8,192 | 5,425.97 | 587.84 | 146.96 | 1,589.12 |
| 5 | 8,192 | 5,336.50 | 623.61 | 124.72 | 1,692.45 |
| 6 | 8,192 | 5,222.31 | 638.58 | 106.43 | 1,774.03 |
| 7 | 8,192 | 5,128.62 | 660.21 | 94.32 | 1,846.09 |
| 8 | 8,192 | 5,011.65 | 670.85 | 83.86 | 1,886.58 |
Technical record and provenance
Source 1 · 2026-10-15
- Source run ID
- pub-20261016-trpro-gemma-e2b-q8-k-xl-context-2026-10-14-single-gpu-isolated-gpu1-co-scheduled-mtp-3-c1
- Page entry ID
- gemma-4-e2b-ud-q8-k-xl-single-gpu-20261016-mtp-3-5d78d5507b
- Entry sweep
- context
- Raw entry label
- Gemma 4 E2B UD-Q8_K_XL — Single GPU — no splitting
- Configuration label
- Single GPU — no splitting
- Measured date
- 2026-10-15
- Hardware
- 1× NVIDIA GeForce RTX 3090
- Runtime
- llamaCpp · version: 0.3.0-dev (build 50, commit fc6545d) built with GNU 15.2.0 for Linux x86_64
- Weight placement
- gpu-resident
- Observed devices
- CUDA0
- Requested devices
- CUDA0
- Split mode
- Single GPU — no splitting
- Placement detail
- Single GPU — no splitting
- KV-cache precision
- Q8_0
- Flash Attention
- Enabled
- Requested context per slot
- 131,072 tokens
- Actual context per slot
- 131,072 tokens
- Native model context
- 131,072 tokens
- Batch size
- 512
- Microbatch size
- 512
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- MTP · draft max 3
- Warm-ups
- 1
- Measured repetitions
- 3
- Timing method
- server durations aligned to client completion timestamps; aggregate-all uses client wall time
- Throughput spreads
- KV 0, 1 stream: PP mean 8,639.81 t/s (min 8,625.05, max 8,647.63, SD 10.44, n=3); TG mean 354.94 t/s (min 354.21, max 356.33, SD 0.98, n=3); All mean 943.63 t/s (min 940.39, max 947.67, SD 3.03, n=3) · KV 2,048, 1 stream: PP mean 7,786.02 t/s (min 7,730.99, max 7,847.71, SD 47.88, n=3); TG mean 349.98 t/s (min 347.58, max 351.21, SD 1.70, n=3); All mean 923.50 t/s (min 918.81, max 926.26, SD 3.34, n=3) · KV 4,096, 1 stream: PP mean 7,401.53 t/s (min 7,359.81, max 7,482.65, SD 57.36, n=3); TG mean 338.88 t/s (min 338.08, max 339.57, SD 0.61, n=3); All mean 893.41 t/s (min 892.05, max 895.25, SD 1.35, n=3) · KV 8,192, 1 stream: PP mean 6,727.11 t/s (min 6,708.24, max 6,761.93, SD 24.65, n=3); TG mean 325.93 t/s (min 325.18, max 326.49, SD 0.56, n=3); All mean 854.12 t/s (min 851.87, max 855.47, SD 1.60, n=3) · KV 16,384, 1 stream: PP mean 5,625.34 t/s (min 5,608.62, max 5,645.29, SD 15.14, n=3); TG mean 300.18 t/s (min 297.13, max 302.33, SD 2.22, n=3); All mean 780.89 t/s (min 773.74, max 786.22, SD 5.26, n=3) · KV 32,768, 1 stream: PP mean 4,297.29 t/s (min 4,260.60, max 4,325.79, SD 27.24, n=3); TG mean 267.08 t/s (min 266.38, max 267.54, SD 0.50, n=3); All mean 684.42 t/s (min 683.61, max 684.96, SD 0.58, n=3) · KV 65,536, 1 stream: PP mean 2,922.45 t/s (min 2,916.96, max 2,930.96, SD 6.10, n=3); TG mean 216.55 t/s (min 216.04, max 217.09, SD 0.43, n=3); All mean 543.82 t/s (min 542.99, max 544.40, SD 0.60, n=3) · KV 130,240, 1 stream: PP mean 1,774.83 t/s (min 1,772.03, max 1,778.43, SD 2.67, n=3); TG mean 157.56 t/s (min 157.34, max 157.93, SD 0.27, n=3); All mean 386.08 t/s (min 385.40, max 386.63, SD 0.51, n=3)
- Measurement outcomes
- 0 tokens · 1 stream: measured; 2,048 tokens · 1 stream: measured; 4,096 tokens · 1 stream: measured; 8,192 tokens · 1 stream: measured; 16,384 tokens · 1 stream: measured; 32,768 tokens · 1 stream: measured; 65,536 tokens · 1 stream: measured; 130,240 tokens · 1 stream: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-3, concurrency=1). Existing depth is contextTokens; PP=512, TG=256, warmups=1, repetitions=3, Q8_0 K/V. Arithmetic mean and population standard deviation of measured repetitions. PP/TG server durations are aligned to client completion timestamps for concurrent phase spans; aggregate-all includes SSH/HTTP wall time. Synthetic repeated-token prompts. Engine was reused as installed; no build was performed. Publication projection of reviewed source measurements; values and raw evidence unchanged. Single GPU — no splitting.
- Source article
- /benchmarks/
- Raw log path
research/benchmarks/logs/trpro/2026-10-15-independent-gpus/e2b-context-gpu1
Source 2 · 2026-10-15
- Source run ID
- pub-20261016-trpro-gemma-e2b-q8-k-xl-concurrency-2026-10-14-single-gpu-isolated-gpu1-co-scheduled-mtp-3-c2
- Page entry ID
- gemma-4-e2b-ud-q8-k-xl-single-gpu-20261016-mtp-3-5d78d5507b
- Entry sweep
- concurrency
- Raw entry label
- Gemma 4 E2B UD-Q8_K_XL — Single GPU — no splitting
- Configuration label
- Single GPU — no splitting
- Measured date
- 2026-10-15
- Hardware
- 1× NVIDIA GeForce RTX 3090
- Runtime
- llamaCpp · version: 0.3.0-dev (build 50, commit fc6545d) built with GNU 15.2.0 for Linux x86_64
- Weight placement
- gpu-resident
- Observed devices
- CUDA0
- Requested devices
- CUDA0
- Split mode
- Single GPU — no splitting
- Placement detail
- Single GPU — no splitting
- KV-cache precision
- Q8_0
- Flash Attention
- Enabled
- Requested context per slot
- 9,024 tokens
- Actual context per slot
- 9,216 tokens
- Native model context
- 131,072 tokens
- Batch size
- 512
- Microbatch size
- 512
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- MTP · draft max 3
- Warm-ups
- 1
- Measured repetitions
- 3
- Timing method
- server durations aligned to client completion timestamps; aggregate-all uses client wall time
- Throughput spreads
- KV 8,192, 2 streams: PP mean 6,323.39 t/s (min 6,108.24, max 6,470.94, SD 155.60, n=3); TG mean 478.75 t/s (min 468.72, max 484.58, SD 7.12, n=3); All mean 1,227.71 t/s (min 1,200.91, max 1,244.29, SD 19.13, n=3)
- Measurement outcomes
- 8,192 tokens · 2 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-3, concurrency=2). Existing depth is contextTokens; PP=512, TG=256, warmups=1, repetitions=3, Q8_0 K/V. Arithmetic mean and population standard deviation of measured repetitions. PP/TG server durations are aligned to client completion timestamps for concurrent phase spans; aggregate-all includes SSH/HTTP wall time. Synthetic repeated-token prompts. Engine was reused as installed; no build was performed. Publication projection of reviewed source measurements; values and raw evidence unchanged. Single GPU — no splitting.
- Source article
- /benchmarks/
- Raw log path
research/benchmarks/logs/trpro/2026-10-15-independent-gpus/gemma-e2b-q8-k-xl-concurrency-physical1
Source 3 · 2026-10-15
- Source run ID
- pub-20261016-trpro-gemma-e2b-q8-k-xl-concurrency-2026-10-14-single-gpu-isolated-gpu1-co-scheduled-mtp-3-c3
- Page entry ID
- gemma-4-e2b-ud-q8-k-xl-single-gpu-20261016-mtp-3-5d78d5507b
- Entry sweep
- concurrency
- Raw entry label
- Gemma 4 E2B UD-Q8_K_XL — Single GPU — no splitting
- Configuration label
- Single GPU — no splitting
- Measured date
- 2026-10-15
- Hardware
- 1× NVIDIA GeForce RTX 3090
- Runtime
- llamaCpp · version: 0.3.0-dev (build 50, commit fc6545d) built with GNU 15.2.0 for Linux x86_64
- Weight placement
- gpu-resident
- Observed devices
- CUDA0
- Requested devices
- CUDA0
- Split mode
- Single GPU — no splitting
- Placement detail
- Single GPU — no splitting
- KV-cache precision
- Q8_0
- Flash Attention
- Enabled
- Requested context per slot
- 9,024 tokens
- Actual context per slot
- 9,216 tokens
- Native model context
- 131,072 tokens
- Batch size
- 512
- Microbatch size
- 512
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- MTP · draft max 3
- Warm-ups
- 1
- Measured repetitions
- 3
- Timing method
- server durations aligned to client completion timestamps; aggregate-all uses client wall time
- Throughput spreads
- KV 8,192, 3 streams: PP mean 5,831.64 t/s (min 5,721.93, max 6,024.72, SD 136.95, n=3); TG mean 538.88 t/s (min 534.62, max 544.16, SD 3.96, n=3); All mean 1,432.69 t/s (min 1,427.22, max 1,437.42, SD 4.20, n=3)
- Measurement outcomes
- 8,192 tokens · 3 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-3, concurrency=3). Existing depth is contextTokens; PP=512, TG=256, warmups=1, repetitions=3, Q8_0 K/V. Arithmetic mean and population standard deviation of measured repetitions. PP/TG server durations are aligned to client completion timestamps for concurrent phase spans; aggregate-all includes SSH/HTTP wall time. Synthetic repeated-token prompts. Engine was reused as installed; no build was performed. Publication projection of reviewed source measurements; values and raw evidence unchanged. Single GPU — no splitting.
- Source article
- /benchmarks/
- Raw log path
research/benchmarks/logs/trpro/2026-10-15-independent-gpus/gemma-e2b-q8-k-xl-concurrency-physical1
Source 4 · 2026-10-15
- Source run ID
- pub-20261016-trpro-gemma-e2b-q8-k-xl-concurrency-2026-10-14-single-gpu-isolated-gpu1-co-scheduled-mtp-3-c4
- Page entry ID
- gemma-4-e2b-ud-q8-k-xl-single-gpu-20261016-mtp-3-5d78d5507b
- Entry sweep
- concurrency
- Raw entry label
- Gemma 4 E2B UD-Q8_K_XL — Single GPU — no splitting
- Configuration label
- Single GPU — no splitting
- Measured date
- 2026-10-15
- Hardware
- 1× NVIDIA GeForce RTX 3090
- Runtime
- llamaCpp · version: 0.3.0-dev (build 50, commit fc6545d) built with GNU 15.2.0 for Linux x86_64
- Weight placement
- gpu-resident
- Observed devices
- CUDA0
- Requested devices
- CUDA0
- Split mode
- Single GPU — no splitting
- Placement detail
- Single GPU — no splitting
- KV-cache precision
- Q8_0
- Flash Attention
- Enabled
- Requested context per slot
- 9,024 tokens
- Actual context per slot
- 9,216 tokens
- Native model context
- 131,072 tokens
- Batch size
- 512
- Microbatch size
- 512
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- MTP · draft max 3
- Warm-ups
- 1
- Measured repetitions
- 3
- Timing method
- server durations aligned to client completion timestamps; aggregate-all uses client wall time
- Throughput spreads
- KV 8,192, 4 streams: PP mean 5,425.97 t/s (min 5,411.33, max 5,439.17, SD 11.41, n=3); TG mean 587.84 t/s (min 587.52, max 588.18, SD 0.27, n=3); All mean 1,589.12 t/s (min 1,588.02, max 1,589.83, SD 0.79, n=3)
- Measurement outcomes
- 8,192 tokens · 4 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-3, concurrency=4). Existing depth is contextTokens; PP=512, TG=256, warmups=1, repetitions=3, Q8_0 K/V. Arithmetic mean and population standard deviation of measured repetitions. PP/TG server durations are aligned to client completion timestamps for concurrent phase spans; aggregate-all includes SSH/HTTP wall time. Synthetic repeated-token prompts. Engine was reused as installed; no build was performed. Publication projection of reviewed source measurements; values and raw evidence unchanged. Single GPU — no splitting.
- Source article
- /benchmarks/
- Raw log path
research/benchmarks/logs/trpro/2026-10-15-independent-gpus/gemma-e2b-q8-k-xl-concurrency-physical1
Source 5 · 2026-10-15
- Source run ID
- pub-20261016-trpro-gemma-e2b-q8-k-xl-concurrency-2026-10-14-single-gpu-isolated-gpu1-co-scheduled-mtp-3-c5
- Page entry ID
- gemma-4-e2b-ud-q8-k-xl-single-gpu-20261016-mtp-3-5d78d5507b
- Entry sweep
- concurrency
- Raw entry label
- Gemma 4 E2B UD-Q8_K_XL — Single GPU — no splitting
- Configuration label
- Single GPU — no splitting
- Measured date
- 2026-10-15
- Hardware
- 1× NVIDIA GeForce RTX 3090
- Runtime
- llamaCpp · version: 0.3.0-dev (build 50, commit fc6545d) built with GNU 15.2.0 for Linux x86_64
- Weight placement
- gpu-resident
- Observed devices
- CUDA0
- Requested devices
- CUDA0
- Split mode
- Single GPU — no splitting
- Placement detail
- Single GPU — no splitting
- KV-cache precision
- Q8_0
- Flash Attention
- Enabled
- Requested context per slot
- 9,024 tokens
- Actual context per slot
- 9,216 tokens
- Native model context
- 131,072 tokens
- Batch size
- 512
- Microbatch size
- 512
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- MTP · draft max 3
- Warm-ups
- 1
- Measured repetitions
- 3
- Timing method
- server durations aligned to client completion timestamps; aggregate-all uses client wall time
- Throughput spreads
- KV 8,192, 5 streams: PP mean 5,336.50 t/s (min 5,313.07, max 5,372.56, SD 25.88, n=3); TG mean 623.61 t/s (min 611.47, max 644.37, SD 14.75, n=3); All mean 1,692.45 t/s (min 1,688.53, max 1,699.29, SD 4.85, n=3)
- Measurement outcomes
- 8,192 tokens · 5 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-3, concurrency=5). Existing depth is contextTokens; PP=512, TG=256, warmups=1, repetitions=3, Q8_0 K/V. Arithmetic mean and population standard deviation of measured repetitions. PP/TG server durations are aligned to client completion timestamps for concurrent phase spans; aggregate-all includes SSH/HTTP wall time. Synthetic repeated-token prompts. Engine was reused as installed; no build was performed. Publication projection of reviewed source measurements; values and raw evidence unchanged. Single GPU — no splitting.
- Source article
- /benchmarks/
- Raw log path
research/benchmarks/logs/trpro/2026-10-15-independent-gpus/gemma-e2b-q8-k-xl-concurrency-physical1
Source 6 · 2026-10-15
- Source run ID
- pub-20261016-trpro-gemma-e2b-q8-k-xl-concurrency-2026-10-14-single-gpu-isolated-gpu1-co-scheduled-mtp-3-c6
- Page entry ID
- gemma-4-e2b-ud-q8-k-xl-single-gpu-20261016-mtp-3-5d78d5507b
- Entry sweep
- concurrency
- Raw entry label
- Gemma 4 E2B UD-Q8_K_XL — Single GPU — no splitting
- Configuration label
- Single GPU — no splitting
- Measured date
- 2026-10-15
- Hardware
- 1× NVIDIA GeForce RTX 3090
- Runtime
- llamaCpp · version: 0.3.0-dev (build 50, commit fc6545d) built with GNU 15.2.0 for Linux x86_64
- Weight placement
- gpu-resident
- Observed devices
- CUDA0
- Requested devices
- CUDA0
- Split mode
- Single GPU — no splitting
- Placement detail
- Single GPU — no splitting
- KV-cache precision
- Q8_0
- Flash Attention
- Enabled
- Requested context per slot
- 9,024 tokens
- Actual context per slot
- 9,216 tokens
- Native model context
- 131,072 tokens
- Batch size
- 512
- Microbatch size
- 512
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- MTP · draft max 3
- Warm-ups
- 1
- Measured repetitions
- 3
- Timing method
- server durations aligned to client completion timestamps; aggregate-all uses client wall time
- Throughput spreads
- KV 8,192, 6 streams: PP mean 5,222.31 t/s (min 5,200.98, max 5,259.05, SD 26.09, n=3); TG mean 638.58 t/s (min 637.08, max 639.51, SD 1.07, n=3); All mean 1,774.03 t/s (min 1,767.15, max 1,780.62, SD 5.50, n=3)
- Measurement outcomes
- 8,192 tokens · 6 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-3, concurrency=6). Existing depth is contextTokens; PP=512, TG=256, warmups=1, repetitions=3, Q8_0 K/V. Arithmetic mean and population standard deviation of measured repetitions. PP/TG server durations are aligned to client completion timestamps for concurrent phase spans; aggregate-all includes SSH/HTTP wall time. Synthetic repeated-token prompts. Engine was reused as installed; no build was performed. Publication projection of reviewed source measurements; values and raw evidence unchanged. Single GPU — no splitting.
- Source article
- /benchmarks/
- Raw log path
research/benchmarks/logs/trpro/2026-10-15-independent-gpus/gemma-e2b-q8-k-xl-concurrency-physical1
Source 7 · 2026-10-15
- Source run ID
- pub-20261016-trpro-gemma-e2b-q8-k-xl-concurrency-2026-10-14-single-gpu-isolated-gpu1-co-scheduled-mtp-3-c7
- Page entry ID
- gemma-4-e2b-ud-q8-k-xl-single-gpu-20261016-mtp-3-5d78d5507b
- Entry sweep
- concurrency
- Raw entry label
- Gemma 4 E2B UD-Q8_K_XL — Single GPU — no splitting
- Configuration label
- Single GPU — no splitting
- Measured date
- 2026-10-15
- Hardware
- 1× NVIDIA GeForce RTX 3090
- Runtime
- llamaCpp · version: 0.3.0-dev (build 50, commit fc6545d) built with GNU 15.2.0 for Linux x86_64
- Weight placement
- gpu-resident
- Observed devices
- CUDA0
- Requested devices
- CUDA0
- Split mode
- Single GPU — no splitting
- Placement detail
- Single GPU — no splitting
- KV-cache precision
- Q8_0
- Flash Attention
- Enabled
- Requested context per slot
- 9,024 tokens
- Actual context per slot
- 9,216 tokens
- Native model context
- 131,072 tokens
- Batch size
- 512
- Microbatch size
- 512
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- MTP · draft max 3
- Warm-ups
- 1
- Measured repetitions
- 3
- Timing method
- server durations aligned to client completion timestamps; aggregate-all uses client wall time
- Throughput spreads
- KV 8,192, 7 streams: PP mean 5,128.62 t/s (min 5,116.59, max 5,150.08, SD 15.22, n=3); TG mean 660.21 t/s (min 659.17, max 660.99, SD 0.76, n=3); All mean 1,846.09 t/s (min 1,842.84, max 1,848.81, SD 2.47, n=3)
- Measurement outcomes
- 8,192 tokens · 7 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-3, concurrency=7). Existing depth is contextTokens; PP=512, TG=256, warmups=1, repetitions=3, Q8_0 K/V. Arithmetic mean and population standard deviation of measured repetitions. PP/TG server durations are aligned to client completion timestamps for concurrent phase spans; aggregate-all includes SSH/HTTP wall time. Synthetic repeated-token prompts. Engine was reused as installed; no build was performed. Publication projection of reviewed source measurements; values and raw evidence unchanged. Single GPU — no splitting.
- Source article
- /benchmarks/
- Raw log path
research/benchmarks/logs/trpro/2026-10-15-independent-gpus/gemma-e2b-q8-k-xl-concurrency-physical1
Source 8 · 2026-10-15
- Source run ID
- pub-20261016-trpro-gemma-e2b-q8-k-xl-concurrency-2026-10-14-single-gpu-isolated-gpu1-co-scheduled-mtp-3-c8
- Page entry ID
- gemma-4-e2b-ud-q8-k-xl-single-gpu-20261016-mtp-3-5d78d5507b
- Entry sweep
- concurrency
- Raw entry label
- Gemma 4 E2B UD-Q8_K_XL — Single GPU — no splitting
- Configuration label
- Single GPU — no splitting
- Measured date
- 2026-10-15
- Hardware
- 1× NVIDIA GeForce RTX 3090
- Runtime
- llamaCpp · version: 0.3.0-dev (build 50, commit fc6545d) built with GNU 15.2.0 for Linux x86_64
- Weight placement
- gpu-resident
- Observed devices
- CUDA0
- Requested devices
- CUDA0
- Split mode
- Single GPU — no splitting
- Placement detail
- Single GPU — no splitting
- KV-cache precision
- Q8_0
- Flash Attention
- Enabled
- Requested context per slot
- 9,024 tokens
- Actual context per slot
- 9,216 tokens
- Native model context
- 131,072 tokens
- Batch size
- 512
- Microbatch size
- 512
- Prompt tokens
- 512
- Generated tokens
- 256
- Speculative decoding
- MTP · draft max 3
- Warm-ups
- 1
- Measured repetitions
- 3
- Timing method
- server durations aligned to client completion timestamps; aggregate-all uses client wall time
- Throughput spreads
- KV 8,192, 8 streams: PP mean 5,011.65 t/s (min 5,005.09, max 5,023.38, SD 8.31, n=3); TG mean 670.85 t/s (min 666.53, max 673.32, SD 3.07, n=3); All mean 1,886.58 t/s (min 1,872.66, max 1,893.89, SD 9.85, n=3)
- Measurement outcomes
- 8,192 tokens · 8 streams: measured
- Notes
- Automated llama-server KV-depth sweep (mtp-3, concurrency=8). Existing depth is contextTokens; PP=512, TG=256, warmups=1, repetitions=3, Q8_0 K/V. Arithmetic mean and population standard deviation of measured repetitions. PP/TG server durations are aligned to client completion timestamps for concurrent phase spans; aggregate-all includes SSH/HTTP wall time. Synthetic repeated-token prompts. Engine was reused as installed; no build was performed. Publication projection of reviewed source measurements; values and raw evidence unchanged. Single GPU — no splitting.
- Source article
- /benchmarks/
- Raw log path
research/benchmarks/logs/trpro/2026-10-15-independent-gpus/gemma-e2b-q8-k-xl-concurrency-physical1