Skip to content

CUDA/HIP: Flash Attention tuning (gfx1201) - #28102

Open
pwilkin wants to merge 3 commits into
ggml-org:masterfrom
pwilkin:fattn-wmma-rdna4-256
Open

CUDA/HIP: Flash Attention tuning (gfx1201)#28102
pwilkin wants to merge 3 commits into
ggml-org:masterfrom
pwilkin:fattn-wmma-rdna4-256

Conversation

@pwilkin

@pwilkin pwilkin commented Aug 31, 2026

Copy link
Copy Markdown
Member

Overview

So, got my new R9700 PRO. It's nice and shiny and has 32GB VRAM, so I decided to try out Qwen3.8 27B. That was a mistake. The prefill performance at longer contexts was abysmal, so I decided to do something about it. Also managed to fix a HS=256 bug in the general CUDA FA code which was preventing the selection of 256 kernels before.

Additional information

Before:

model size params backend threads dev test t/s
qwen35 27B IQ4_XS - 4.25 bpw 13.26 GiB 27.32 B BLAS,CUDA,ROCm,Vulkan 8 ROCm0 pp512 1101.41 ± 221.33
qwen35 27B IQ4_XS - 4.25 bpw 13.26 GiB 27.32 B BLAS,CUDA,ROCm,Vulkan 8 ROCm0 tg128 29.81 ± 0.19
qwen35 27B IQ4_XS - 4.25 bpw 13.26 GiB 27.32 B BLAS,CUDA,ROCm,Vulkan 8 ROCm0 pp512 @ d40000 426.36 ± 37.60
qwen35 27B IQ4_XS - 4.25 bpw 13.26 GiB 27.32 B BLAS,CUDA,ROCm,Vulkan 8 ROCm0 tg128 @ d40000 26.75 ± 0.46
qwen35 27B IQ4_XS - 4.25 bpw 13.26 GiB 27.32 B BLAS,CUDA,ROCm,Vulkan 8 ROCm0 pp512 @ d150000 164.42 ± 5.79
qwen35 27B IQ4_XS - 4.25 bpw 13.26 GiB 27.32 B BLAS,CUDA,ROCm,Vulkan 8 ROCm0 tg128 @ d150000 19.70 ± 0.20

After:

model size params backend threads dev test t/s
qwen35 27B IQ4_XS - 4.25 bpw 13.26 GiB 27.32 B BLAS,CUDA,ROCm,Vulkan 8 ROCm0 pp512 1103.66 ± 254.83
qwen35 27B IQ4_XS - 4.25 bpw 13.26 GiB 27.32 B BLAS,CUDA,ROCm,Vulkan 8 ROCm0 tg128 29.87 ± 0.46
qwen35 27B IQ4_XS - 4.25 bpw 13.26 GiB 27.32 B BLAS,CUDA,ROCm,Vulkan 8 ROCm0 pp512 @ d40000 639.46 ± 185.36
qwen35 27B IQ4_XS - 4.25 bpw 13.26 GiB 27.32 B BLAS,CUDA,ROCm,Vulkan 8 ROCm0 tg128 @ d40000 26.94 ± 0.46
qwen35 27B IQ4_XS - 4.25 bpw 13.26 GiB 27.32 B BLAS,CUDA,ROCm,Vulkan 8 ROCm0 pp512 @ d150000 399.01 ± 30.87
qwen35 27B IQ4_XS - 4.25 bpw 13.26 GiB 27.32 B BLAS,CUDA,ROCm,Vulkan 8 ROCm0 tg128 @ d150000 19.67 ± 0.39

Requirements

@pwilkin
pwilkin requested review from a team and ggerganov as code owners August 31, 2026 16:54
@github-actions github-actions Bot added testing Everything test related ggml changes relating to the ggml tensor library for machine learning CUDA Related to the CUDA backend labels Aug 31, 2026
@Geramy

Geramy commented Aug 31, 2026

Copy link
Copy Markdown
Contributor

This looks really good! Such a simple change providing extra performance, good job on finding it!

@IMbackK IMbackK self-assigned this Aug 31, 2026
@pwilkin

pwilkin commented Sep 1, 2026

Copy link
Copy Markdown
Member Author

I'm not done yet!

model size params backend threads dev test t/s
qwen35 27B IQ4_XS - 4.25 bpw 13.26 GiB 27.32 B BLAS,CUDA,ROCm,Vulkan 8 ROCm0 pp512 1138.55 ± 224.34
qwen35 27B IQ4_XS - 4.25 bpw 13.26 GiB 27.32 B BLAS,CUDA,ROCm,Vulkan 8 ROCm0 tg128 29.48 ± 0.83
qwen35 27B IQ4_XS - 4.25 bpw 13.26 GiB 27.32 B BLAS,CUDA,ROCm,Vulkan 8 ROCm0 pp512 @ d40000 704.92 ± 239.12
qwen35 27B IQ4_XS - 4.25 bpw 13.26 GiB 27.32 B BLAS,CUDA,ROCm,Vulkan 8 ROCm0 tg128 @ d40000 27.00 ± 0.46
qwen35 27B IQ4_XS - 4.25 bpw 13.26 GiB 27.32 B BLAS,CUDA,ROCm,Vulkan 8 ROCm0 pp512 @ d150000 456.51 ± 38.04
qwen35 27B IQ4_XS - 4.25 bpw 13.26 GiB 27.32 B BLAS,CUDA,ROCm,Vulkan 8 ROCm0 tg128 @ d150000 19.67 ± 0.42

@pwilkin
pwilkin force-pushed the fattn-wmma-rdna4-256 branch from 6b24ee9 to d68f876 Compare September 1, 2026 07:54

@IMbackK IMbackK left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

nice, looks good

Details
GPU Model Microbatch size Test t/s master t/s fattn-wmma-rdna4-256 Speedup
AI PRO R9700 gemma4 26B.A4B Q6_K 8 pp2048@d32768 327.19 308.26 0.94
AI PRO R9700 gemma4 26B.A4B Q6_K 64 pp2048@d32768 662.22 668.99 1.01
AI PRO R9700 gemma4 26B.A4B Q6_K 512 pp2048@d32768 1263.12 1256.57 0.99
AI PRO R9700 gemma4 26B.A4B Q6_K 1024 pp2048@d32768 1366.90 1495.99 1.09
AI PRO R9700 gpt-oss 20B MXFP4 MoE 8 pp2048@d32768 523.46 525.46 1.00
AI PRO R9700 gpt-oss 20B MXFP4 MoE 64 pp2048@d32768 1408.82 1364.30 0.97
AI PRO R9700 gpt-oss 20B MXFP4 MoE 512 pp2048@d32768 2588.42 2838.27 1.10
AI PRO R9700 gpt-oss 20B MXFP4 MoE 1024 pp2048@d32768 2869.95 3281.25 1.14
AI PRO R9700 lfm2moe 8B.A1B Q8_0 8 pp2048@d32768 817.76 817.97 1.00
AI PRO R9700 lfm2moe 8B.A1B Q8_0 64 pp2048@d32768 2947.91 2947.64 1.00
AI PRO R9700 lfm2moe 8B.A1B Q8_0 512 pp2048@d32768 7976.07 7986.46 1.00
AI PRO R9700 lfm2moe 8B.A1B Q8_0 1024 pp2048@d32768 9433.99 8928.84 0.95
AI PRO R9700 llama 8B Q8_0 8 pp2048@d32768 312.91 312.85 1.00
AI PRO R9700 llama 8B Q8_0 64 pp2048@d32768 1209.11 1181.70 0.98
AI PRO R9700 llama 8B Q8_0 512 pp2048@d32768 1145.38 1705.73 1.49
AI PRO R9700 llama 8B Q8_0 1024 pp2048@d32768 1064.61 1587.85 1.49
AI PRO R9700 qwen35 27B Q5_K_M 8 pp2048@d32768 75.64 72.25 0.96
AI PRO R9700 qwen35 27B Q5_K_M 64 pp2048@d32768 372.91 479.10 1.28
AI PRO R9700 qwen35 27B Q5_K_M 512 pp2048@d32768 468.45 707.39 1.51
AI PRO R9700 qwen35 27B Q5_K_M 1024 pp2048@d32768 475.40 708.34 1.49

<\details>

Comment thread ggml/src/ggml-cuda/fattn-common.cuh Outdated
bool use_stream_k = cc >= GGML_CUDA_CC_ADA_LOVELACE || amd_wmma_available(cc) || tiles_efficiency_percent < 75;
if (amd_wmma_available(cc) && ntiles_dst >= 2*max_blocks && tiles_efficiency_percent >= 75) {
use_stream_k = false;
}

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

we might want to refactor this into a helper function with table at this point.

@sbstnh

sbstnh commented Sep 3, 2026

Copy link
Copy Markdown

can confirm the performance improvements on my R9700. Hope this PR will be merged soon.

@pwilkin

pwilkin commented Sep 5, 2026

Copy link
Copy Markdown
Member Author

@IMbackK added the helper.

@JohannesGaessler JohannesGaessler left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Please avoid piling on many unrelated changes into a single PR like this.

@JohannesGaessler JohannesGaessler self-assigned this Sep 5, 2026
@pwilkin

pwilkin commented Sep 5, 2026

Copy link
Copy Markdown
Member Author

Yeah, I just noticed the GDA changes got swept in as well, I'll move them to a separate branch.

@pwilkin

pwilkin commented Sep 5, 2026

Copy link
Copy Markdown
Member Author

@JohannesGaessler aight, cleaned it up to just the FATTN changes.

@JohannesGaessler

Copy link
Copy Markdown
Contributor
  • For this PR, please reduce it to just the configuration changes.
  • Make a follow-up with only the changes for the stream-k logic.
  • The other changes I'm not sure are logically correct so I would like to treat them separately.

It is a lot easier for me to check the performance with separate PRs. I'm available to work on this this weekend so I think we can get it done if you are as well.

@pwilkin

pwilkin commented Sep 5, 2026

Copy link
Copy Markdown
Member Author

Aight, will try that.

@JohannesGaessler

Copy link
Copy Markdown
Contributor

Actually, looking at the order the commits were done in it's not clear to me that the config change and the change to stream-k can be separated. But Claude is definitely wrong about the supposed race condition it fixed.

@IMbackK

IMbackK commented Sep 6, 2026

Copy link
Copy Markdown
Contributor

imo the state of this pr was fine at 2d55b5f

@IMbackK

IMbackK commented Sep 6, 2026

Copy link
Copy Markdown
Contributor

nvm, your right d6feff3 is not necessary its just reading its own rows in that case, got bamboozled by that one. up to 2d55b5f minus d6feff3 is good.

@pwilkin

pwilkin commented Sep 6, 2026

Copy link
Copy Markdown
Member Author

Yeah, shouldn't have trusted Claude on that one. Reverting that.

@IMbackK

IMbackK commented Sep 6, 2026

Copy link
Copy Markdown
Contributor

mind doing a git rebase -i instead? gets kinda confusing to review like this.

Comment thread ggml/src/ggml-cuda/fattn-mma-f16.cuh Outdated
Comment on lines +582 to +587
// swizzle the tile stride for K and V based on the batch size.
constexpr int stride_tile_K = ggml_cuda_fattn_smem_swizzle::tile_stride(nbatch_K2);
#if defined(AMD_WMMA_AVAILABLE)
constexpr int stride_tile_V = V_is_K_view ? stride_tile_K : nbatch_V2 + 6;
#else
constexpr int stride_tile_V = V_is_K_view ? stride_tile_K : ggml_cuda_fattn_smem_swizzle::tile_stride(nbatch_V2);
#endif // defined(AMD_WMMA_AVAILABLE)

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Again, please submit changes to the actual device code beyond changes to the config and the host-side orchestration separately. I would be surprised if this is actually the correct padding to minimize LDS bank conflicts.

@pwilkin
pwilkin force-pushed the fattn-wmma-rdna4-256 branch from f6264d8 to a3b28f9 Compare September 6, 2026 13:08
@pwilkin

pwilkin commented Sep 6, 2026

Copy link
Copy Markdown
Member Author

Rebased, squashed, hopefully it's good now.

Comment on lines +224 to +240
if (GGML_CUDA_CC_IS_RDNA4(cc)) {
if (use_gqa_opt && gqa_ratio % 8 == 0) {
ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 8>(ctx, dst);
return;
}

if (use_gqa_opt && gqa_ratio % 4 == 0) {
ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 4>(ctx, dst);
return;
}

if (use_gqa_opt && gqa_ratio % 2 == 0) {
ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 2>(ctx, dst);
return;
}
}

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I don't think these extra code branches are needed.

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Looking at the exact parameters for Qwen 3.5, what these seem to be doing is launch different kernels for GQA ratios that are not a power of 2, Qwen 3.5 has 6.

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yeah, is that wrong?

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I don't know yet.

@JohannesGaessler

JohannesGaessler commented Sep 6, 2026

Copy link
Copy Markdown
Contributor

I observed some performance regressions on my RX 9060 XT with this PR. I pushed revised logic for whether or not to use stream-k, can you check how that performs on your system? This is the data I got:

PW vs. master
GPU Model Microbatch size Test t/s master t/s a3b28f9 Speedup
RX 9060 XT gemma 2B Q4_0 1 pp1024 199.77 199.51 1.00
RX 9060 XT gemma 2B Q4_0 1 pp1024@d1024 195.02 194.96 1.00
RX 9060 XT gemma 2B Q4_0 1 pp1024@d2048 193.32 192.71 1.00
RX 9060 XT gemma 2B Q4_0 1 pp1024@d4096 185.70 185.87 1.00
RX 9060 XT gemma 2B Q4_0 1 pp1024@d8192 176.77 176.34 1.00
RX 9060 XT gemma 2B Q4_0 1 pp1024@d16384 159.14 158.58 1.00
RX 9060 XT gemma 2B Q4_0 1 pp1024@d32768 133.21 133.01 1.00
RX 9060 XT gemma 2B Q4_0 2 pp1024 251.97 244.16 0.97
RX 9060 XT gemma 2B Q4_0 2 pp1024@d1024 250.26 241.04 0.96
RX 9060 XT gemma 2B Q4_0 2 pp1024@d2048 247.19 235.81 0.95
RX 9060 XT gemma 2B Q4_0 2 pp1024@d4096 247.23 234.11 0.95
RX 9060 XT gemma 2B Q4_0 2 pp1024@d8192 239.84 227.26 0.95
RX 9060 XT gemma 2B Q4_0 2 pp1024@d16384 225.61 214.44 0.95
RX 9060 XT gemma 2B Q4_0 2 pp1024@d32768 201.53 191.32 0.95
RX 9060 XT gemma 2B Q4_0 4 pp1024 497.70 490.16 0.98
RX 9060 XT gemma 2B Q4_0 4 pp1024@d1024 480.05 473.93 0.99
RX 9060 XT gemma 2B Q4_0 4 pp1024@d2048 474.65 469.08 0.99
RX 9060 XT gemma 2B Q4_0 4 pp1024@d4096 462.00 455.97 0.99
RX 9060 XT gemma 2B Q4_0 4 pp1024@d8192 436.76 436.99 1.00
RX 9060 XT gemma 2B Q4_0 4 pp1024@d16384 398.36 406.74 1.02
RX 9060 XT gemma 2B Q4_0 4 pp1024@d32768 343.28 369.11 1.08
RX 9060 XT gemma 2B Q4_0 8 pp1024 813.96 807.04 0.99
RX 9060 XT gemma 2B Q4_0 8 pp1024@d1024 793.26 796.88 1.00
RX 9060 XT gemma 2B Q4_0 8 pp1024@d2048 776.17 789.05 1.02
RX 9060 XT gemma 2B Q4_0 8 pp1024@d4096 739.38 770.05 1.04
RX 9060 XT gemma 2B Q4_0 8 pp1024@d8192 677.58 745.12 1.10
RX 9060 XT gemma 2B Q4_0 8 pp1024@d16384 582.69 697.48 1.20
RX 9060 XT gemma 2B Q4_0 8 pp1024@d32768 447.92 618.61 1.38
RX 9060 XT gemma 2B Q4_0 16 pp1024 1534.08 1541.23 1.00
RX 9060 XT gemma 2B Q4_0 16 pp1024@d1024 1470.10 1514.17 1.03
RX 9060 XT gemma 2B Q4_0 16 pp1024@d2048 1404.48 1488.73 1.06
RX 9060 XT gemma 2B Q4_0 16 pp1024@d4096 1298.53 1431.61 1.10
RX 9060 XT gemma 2B Q4_0 16 pp1024@d8192 1113.28 1350.10 1.21
RX 9060 XT gemma 2B Q4_0 16 pp1024@d16384 882.45 1225.79 1.39
RX 9060 XT gemma 2B Q4_0 16 pp1024@d32768 608.97 1007.76 1.65
RX 9060 XT gemma 2B Q4_0 32 pp1024 2495.77 2598.26 1.04
RX 9060 XT gemma 2B Q4_0 32 pp1024@d1024 2308.08 2526.66 1.09
RX 9060 XT gemma 2B Q4_0 32 pp1024@d2048 2138.33 2463.82 1.15
RX 9060 XT gemma 2B Q4_0 32 pp1024@d4096 1891.24 2299.89 1.22
RX 9060 XT gemma 2B Q4_0 32 pp1024@d8192 1523.24 2092.87 1.37
RX 9060 XT gemma 2B Q4_0 32 pp1024@d16384 1099.01 1779.51 1.62
RX 9060 XT gemma 2B Q4_0 32 pp1024@d32768 708.51 1353.04 1.91
RX 9060 XT gemma 2B Q4_0 64 pp1024 3681.43 3768.68 1.02
RX 9060 XT gemma 2B Q4_0 64 pp1024@d1024 3286.00 3628.76 1.10
RX 9060 XT gemma 2B Q4_0 64 pp1024@d2048 2951.91 3462.08 1.17
RX 9060 XT gemma 2B Q4_0 64 pp1024@d4096 2493.95 3212.83 1.29
RX 9060 XT gemma 2B Q4_0 64 pp1024@d8192 1899.01 2844.27 1.50
RX 9060 XT gemma 2B Q4_0 64 pp1024@d16384 1291.77 2286.47 1.77
RX 9060 XT gemma 2B Q4_0 64 pp1024@d32768 784.82 1639.84 2.09
RX 9060 XT gemma 2B Q4_0 128 pp1024 4925.28 5037.19 1.02
RX 9060 XT gemma 2B Q4_0 128 pp1024@d1024 4321.66 4701.78 1.09
RX 9060 XT gemma 2B Q4_0 128 pp1024@d2048 3884.52 4465.72 1.15
RX 9060 XT gemma 2B Q4_0 128 pp1024@d4096 3231.15 4080.08 1.26
RX 9060 XT gemma 2B Q4_0 128 pp1024@d8192 2429.50 3465.77 1.43
RX 9060 XT gemma 2B Q4_0 128 pp1024@d16384 1637.36 2669.78 1.63
RX 9060 XT gemma 2B Q4_0 128 pp1024@d32768 966.82 1828.07 1.89
RX 9060 XT gemma 2B Q4_0 256 pp1024 7290.98 7670.59 1.05
RX 9060 XT gemma 2B Q4_0 256 pp1024@d1024 6013.49 7042.40 1.17
RX 9060 XT gemma 2B Q4_0 256 pp1024@d2048 5185.53 6697.28 1.29
RX 9060 XT gemma 2B Q4_0 256 pp1024@d4096 4099.34 6031.23 1.47
RX 9060 XT gemma 2B Q4_0 256 pp1024@d8192 2910.29 5102.77 1.75
RX 9060 XT gemma 2B Q4_0 256 pp1024@d16384 1862.82 3937.68 2.11
RX 9060 XT gemma 2B Q4_0 256 pp1024@d32768 1071.32 2686.63 2.51
RX 9060 XT gemma 2B Q4_0 512 pp1024 7538.15 8163.80 1.08
RX 9060 XT gemma 2B Q4_0 512 pp1024@d1024 6284.13 7601.69 1.21
RX 9060 XT gemma 2B Q4_0 512 pp1024@d2048 5439.49 7168.91 1.32
RX 9060 XT gemma 2B Q4_0 512 pp1024@d4096 4267.23 6480.79 1.52
RX 9060 XT gemma 2B Q4_0 512 pp1024@d8192 3024.31 5388.91 1.78
RX 9060 XT gemma 2B Q4_0 512 pp1024@d16384 1922.31 4049.56 2.11
RX 9060 XT gemma 2B Q4_0 512 pp1024@d32768 1099.73 2713.10 2.47
RX 9060 XT llama 1B Q4_0 1 pp1024 326.87 329.53 1.01
RX 9060 XT llama 1B Q4_0 1 pp1024@d1024 313.94 315.89 1.01
RX 9060 XT llama 1B Q4_0 1 pp1024@d2048 303.64 303.91 1.00
RX 9060 XT llama 1B Q4_0 1 pp1024@d4096 285.31 287.63 1.01
RX 9060 XT llama 1B Q4_0 1 pp1024@d8192 256.54 258.10 1.01
RX 9060 XT llama 1B Q4_0 1 pp1024@d16384 211.00 211.25 1.00
RX 9060 XT llama 1B Q4_0 1 pp1024@d32768 155.10 155.04 1.00
RX 9060 XT llama 1B Q4_0 2 pp1024 411.48 412.22 1.00
RX 9060 XT llama 1B Q4_0 2 pp1024@d1024 397.65 398.61 1.00
RX 9060 XT llama 1B Q4_0 2 pp1024@d2048 388.50 389.11 1.00
RX 9060 XT llama 1B Q4_0 2 pp1024@d4096 368.65 369.63 1.00
RX 9060 XT llama 1B Q4_0 2 pp1024@d8192 333.23 333.75 1.00
RX 9060 XT llama 1B Q4_0 2 pp1024@d16384 279.96 280.08 1.00
RX 9060 XT llama 1B Q4_0 2 pp1024@d32768 210.77 210.57 1.00
RX 9060 XT llama 1B Q4_0 4 pp1024 604.69 605.93 1.00
RX 9060 XT llama 1B Q4_0 4 pp1024@d1024 590.53 594.39 1.01
RX 9060 XT llama 1B Q4_0 4 pp1024@d2048 582.25 583.94 1.00
RX 9060 XT llama 1B Q4_0 4 pp1024@d4096 562.41 562.68 1.00
RX 9060 XT llama 1B Q4_0 4 pp1024@d8192 526.75 527.05 1.00
RX 9060 XT llama 1B Q4_0 4 pp1024@d16384 465.37 465.15 1.00
RX 9060 XT llama 1B Q4_0 4 pp1024@d32768 376.68 376.38 1.00
RX 9060 XT llama 1B Q4_0 8 pp1024 1208.72 1215.98 1.01
RX 9060 XT llama 1B Q4_0 8 pp1024@d1024 1183.24 1189.28 1.01
RX 9060 XT llama 1B Q4_0 8 pp1024@d2048 1158.59 1159.71 1.00
RX 9060 XT llama 1B Q4_0 8 pp1024@d4096 1116.13 1117.16 1.00
RX 9060 XT llama 1B Q4_0 8 pp1024@d8192 1040.70 1040.65 1.00
RX 9060 XT llama 1B Q4_0 8 pp1024@d16384 915.93 915.79 1.00
RX 9060 XT llama 1B Q4_0 8 pp1024@d32768 745.76 744.42 1.00
RX 9060 XT llama 1B Q4_0 16 pp1024 2603.47 2595.57 1.00
RX 9060 XT llama 1B Q4_0 16 pp1024@d1024 2517.27 2527.75 1.00
RX 9060 XT llama 1B Q4_0 16 pp1024@d2048 2465.07 2466.36 1.00
RX 9060 XT llama 1B Q4_0 16 pp1024@d4096 2375.96 2373.25 1.00
RX 9060 XT llama 1B Q4_0 16 pp1024@d8192 2201.16 2201.90 1.00
RX 9060 XT llama 1B Q4_0 16 pp1024@d16384 1911.19 1907.22 1.00
RX 9060 XT llama 1B Q4_0 16 pp1024@d32768 1514.74 1512.98 1.00
RX 9060 XT llama 1B Q4_0 32 pp1024 3843.24 3859.03 1.00
RX 9060 XT llama 1B Q4_0 32 pp1024@d1024 3737.16 3779.37 1.01
RX 9060 XT llama 1B Q4_0 32 pp1024@d2048 3663.59 3667.18 1.00
RX 9060 XT llama 1B Q4_0 32 pp1024@d4096 3525.29 3508.37 1.00
RX 9060 XT llama 1B Q4_0 32 pp1024@d8192 3306.39 3245.74 0.98
RX 9060 XT llama 1B Q4_0 32 pp1024@d16384 2777.85 2768.55 1.00
RX 9060 XT llama 1B Q4_0 32 pp1024@d32768 2121.97 2115.00 1.00
RX 9060 XT llama 1B Q4_0 64 pp1024 5335.02 5354.61 1.00
RX 9060 XT llama 1B Q4_0 64 pp1024@d1024 5219.23 5219.14 1.00
RX 9060 XT llama 1B Q4_0 64 pp1024@d2048 4979.65 4979.69 1.00
RX 9060 XT llama 1B Q4_0 64 pp1024@d4096 4681.70 4685.37 1.00
RX 9060 XT llama 1B Q4_0 64 pp1024@d8192 4132.44 4159.06 1.01
RX 9060 XT llama 1B Q4_0 64 pp1024@d16384 3372.08 3365.11 1.00
RX 9060 XT llama 1B Q4_0 64 pp1024@d32768 2413.12 2403.37 1.00
RX 9060 XT llama 1B Q4_0 128 pp1024 8225.01 8084.94 0.98
RX 9060 XT llama 1B Q4_0 128 pp1024@d1024 7756.62 7739.91 1.00
RX 9060 XT llama 1B Q4_0 128 pp1024@d2048 7189.73 7152.98 0.99
RX 9060 XT llama 1B Q4_0 128 pp1024@d4096 6542.08 6534.04 1.00
RX 9060 XT llama 1B Q4_0 128 pp1024@d8192 5507.57 5518.92 1.00
RX 9060 XT llama 1B Q4_0 128 pp1024@d16384 4186.53 4201.58 1.00
RX 9060 XT llama 1B Q4_0 128 pp1024@d32768 2789.97 2783.58 1.00
RX 9060 XT llama 1B Q4_0 256 pp1024 12965.20 13001.92 1.00
RX 9060 XT llama 1B Q4_0 256 pp1024@d1024 11452.63 11462.87 1.00
RX 9060 XT llama 1B Q4_0 256 pp1024@d2048 10539.07 10416.98 0.99
RX 9060 XT llama 1B Q4_0 256 pp1024@d4096 9192.08 9153.15 1.00
RX 9060 XT llama 1B Q4_0 256 pp1024@d8192 7259.93 7211.04 0.99
RX 9060 XT llama 1B Q4_0 256 pp1024@d16384 5078.09 5086.47 1.00
RX 9060 XT llama 1B Q4_0 256 pp1024@d32768 3162.75 3153.39 1.00
RX 9060 XT llama 1B Q4_0 512 pp1024 12388.58 12692.93 1.02
RX 9060 XT llama 1B Q4_0 512 pp1024@d1024 11988.93 12100.80 1.01
RX 9060 XT llama 1B Q4_0 512 pp1024@d2048 11252.73 11171.83 0.99
RX 9060 XT llama 1B Q4_0 512 pp1024@d4096 9744.95 9590.98 0.98
RX 9060 XT llama 1B Q4_0 512 pp1024@d8192 7656.92 7280.01 0.95
RX 9060 XT llama 1B Q4_0 512 pp1024@d16384 5247.30 4949.73 0.94
RX 9060 XT llama 1B Q4_0 512 pp1024@d32768 3195.56 2993.52 0.94
RX 9060 XT llama 8B Q4_0 1 pp1024 68.21 68.24 1.00
RX 9060 XT llama 8B Q4_0 1 pp1024@d1024 66.15 66.18 1.00
RX 9060 XT llama 8B Q4_0 1 pp1024@d2048 64.34 64.37 1.00
RX 9060 XT llama 8B Q4_0 1 pp1024@d4096 61.09 61.06 1.00
RX 9060 XT llama 8B Q4_0 1 pp1024@d8192 55.44 55.45 1.00
RX 9060 XT llama 8B Q4_0 1 pp1024@d16384 46.69 46.69 1.00
RX 9060 XT llama 8B Q4_0 1 pp1024@d32768 35.56 35.56 1.00
RX 9060 XT llama 8B Q4_0 2 pp1024 100.66 100.58 1.00
RX 9060 XT llama 8B Q4_0 2 pp1024@d1024 98.28 98.23 1.00
RX 9060 XT llama 8B Q4_0 2 pp1024@d2048 96.34 96.27 1.00
RX 9060 XT llama 8B Q4_0 2 pp1024@d4096 92.76 92.62 1.00
RX 9060 XT llama 8B Q4_0 2 pp1024@d8192 86.07 85.87 1.00
RX 9060 XT llama 8B Q4_0 2 pp1024@d16384 74.97 75.00 1.00
RX 9060 XT llama 8B Q4_0 2 pp1024@d32768 59.84 59.81 1.00
RX 9060 XT llama 8B Q4_0 4 pp1024 182.37 182.12 1.00
RX 9060 XT llama 8B Q4_0 4 pp1024@d1024 177.02 176.78 1.00
RX 9060 XT llama 8B Q4_0 4 pp1024@d2048 173.13 173.07 1.00
RX 9060 XT llama 8B Q4_0 4 pp1024@d4096 165.30 165.23 1.00
RX 9060 XT llama 8B Q4_0 4 pp1024@d8192 151.11 151.10 1.00
RX 9060 XT llama 8B Q4_0 4 pp1024@d16384 130.26 130.25 1.00
RX 9060 XT llama 8B Q4_0 4 pp1024@d32768 100.95 100.85 1.00
RX 9060 XT llama 8B Q4_0 8 pp1024 199.43 198.86 1.00
RX 9060 XT llama 8B Q4_0 8 pp1024@d1024 196.56 196.26 1.00
RX 9060 XT llama 8B Q4_0 8 pp1024@d2048 194.44 194.77 1.00
RX 9060 XT llama 8B Q4_0 8 pp1024@d4096 190.47 190.18 1.00
RX 9060 XT llama 8B Q4_0 8 pp1024@d8192 182.34 182.38 1.00
RX 9060 XT llama 8B Q4_0 8 pp1024@d16384 169.30 169.17 1.00
RX 9060 XT llama 8B Q4_0 8 pp1024@d32768 147.88 147.82 1.00
RX 9060 XT llama 8B Q4_0 16 pp1024 660.73 661.43 1.00
RX 9060 XT llama 8B Q4_0 16 pp1024@d1024 641.12 638.73 1.00
RX 9060 XT llama 8B Q4_0 16 pp1024@d2048 624.75 624.71 1.00
RX 9060 XT llama 8B Q4_0 16 pp1024@d4096 595.31 594.47 1.00
RX 9060 XT llama 8B Q4_0 16 pp1024@d8192 542.94 542.96 1.00
RX 9060 XT llama 8B Q4_0 16 pp1024@d16384 462.64 462.56 1.00
RX 9060 XT llama 8B Q4_0 16 pp1024@d32768 357.99 357.78 1.00
RX 9060 XT llama 8B Q4_0 32 pp1024 990.51 991.28 1.00
RX 9060 XT llama 8B Q4_0 32 pp1024@d1024 953.80 948.60 0.99
RX 9060 XT llama 8B Q4_0 32 pp1024@d2048 921.67 920.75 1.00
RX 9060 XT llama 8B Q4_0 32 pp1024@d4096 866.31 864.65 1.00
RX 9060 XT llama 8B Q4_0 32 pp1024@d8192 769.47 769.72 1.00
RX 9060 XT llama 8B Q4_0 32 pp1024@d16384 628.25 628.00 1.00
RX 9060 XT llama 8B Q4_0 32 pp1024@d32768 464.45 461.98 0.99
RX 9060 XT llama 8B Q4_0 64 pp1024 1610.44 1604.77 1.00
RX 9060 XT llama 8B Q4_0 64 pp1024@d1024 1507.52 1504.59 1.00
RX 9060 XT llama 8B Q4_0 64 pp1024@d2048 1447.53 1444.09 1.00
RX 9060 XT llama 8B Q4_0 64 pp1024@d4096 1320.89 1318.74 1.00
RX 9060 XT llama 8B Q4_0 64 pp1024@d8192 1087.31 1087.43 1.00
RX 9060 XT llama 8B Q4_0 64 pp1024@d16384 832.08 832.58 1.00
RX 9060 XT llama 8B Q4_0 64 pp1024@d32768 569.23 566.67 1.00
RX 9060 XT llama 8B Q4_0 128 pp1024 2359.64 2340.84 0.99
RX 9060 XT llama 8B Q4_0 128 pp1024@d1024 2142.74 2142.86 1.00
RX 9060 XT llama 8B Q4_0 128 pp1024@d2048 2005.35 2005.38 1.00
RX 9060 XT llama 8B Q4_0 128 pp1024@d4096 1771.69 1775.94 1.00
RX 9060 XT llama 8B Q4_0 128 pp1024@d8192 1374.73 1377.59 1.00
RX 9060 XT llama 8B Q4_0 128 pp1024@d16384 1003.99 1000.71 1.00
RX 9060 XT llama 8B Q4_0 128 pp1024@d32768 648.49 646.06 1.00
RX 9060 XT llama 8B Q4_0 256 pp1024 2612.44 2595.07 0.99
RX 9060 XT llama 8B Q4_0 256 pp1024@d1024 2347.26 2372.22 1.01
RX 9060 XT llama 8B Q4_0 256 pp1024@d2048 2185.63 2206.63 1.01
RX 9060 XT llama 8B Q4_0 256 pp1024@d4096 1934.63 1967.04 1.02
RX 9060 XT llama 8B Q4_0 256 pp1024@d8192 1477.02 1607.39 1.09
RX 9060 XT llama 8B Q4_0 256 pp1024@d16384 1056.31 1186.13 1.12
RX 9060 XT llama 8B Q4_0 256 pp1024@d32768 674.72 785.23 1.16
RX 9060 XT llama 8B Q4_0 512 pp1024 2700.30 2701.59 1.00
RX 9060 XT llama 8B Q4_0 512 pp1024@d1024 2458.57 2490.01 1.01
RX 9060 XT llama 8B Q4_0 512 pp1024@d2048 2287.08 2301.61 1.01
RX 9060 XT llama 8B Q4_0 512 pp1024@d4096 2009.40 2031.68 1.01
RX 9060 XT llama 8B Q4_0 512 pp1024@d8192 1502.39 1641.36 1.09
RX 9060 XT llama 8B Q4_0 512 pp1024@d16384 1056.29 1196.08 1.13
RX 9060 XT llama 8B Q4_0 512 pp1024@d32768 674.64 787.39 1.17
JG vs. master
GPU Model Microbatch size Test t/s master t/s 0ce2938 Speedup
RX 9060 XT gemma 2B Q4_0 1 pp1024 199.77 198.92 1.00
RX 9060 XT gemma 2B Q4_0 1 pp1024@d1024 195.02 195.75 1.00
RX 9060 XT gemma 2B Q4_0 1 pp1024@d2048 193.32 192.97 1.00
RX 9060 XT gemma 2B Q4_0 1 pp1024@d4096 185.70 185.69 1.00
RX 9060 XT gemma 2B Q4_0 1 pp1024@d8192 176.77 176.69 1.00
RX 9060 XT gemma 2B Q4_0 1 pp1024@d16384 159.14 158.96 1.00
RX 9060 XT gemma 2B Q4_0 1 pp1024@d32768 133.21 133.54 1.00
RX 9060 XT gemma 2B Q4_0 2 pp1024 251.97 245.62 0.97
RX 9060 XT gemma 2B Q4_0 2 pp1024@d1024 250.26 237.92 0.95
RX 9060 XT gemma 2B Q4_0 2 pp1024@d2048 247.19 231.55 0.94
RX 9060 XT gemma 2B Q4_0 2 pp1024@d4096 247.23 234.05 0.95
RX 9060 XT gemma 2B Q4_0 2 pp1024@d8192 239.84 227.32 0.95
RX 9060 XT gemma 2B Q4_0 2 pp1024@d16384 225.61 214.02 0.95
RX 9060 XT gemma 2B Q4_0 2 pp1024@d32768 201.53 191.30 0.95
RX 9060 XT gemma 2B Q4_0 4 pp1024 497.70 488.44 0.98
RX 9060 XT gemma 2B Q4_0 4 pp1024@d1024 480.05 473.32 0.99
RX 9060 XT gemma 2B Q4_0 4 pp1024@d2048 474.65 469.23 0.99
RX 9060 XT gemma 2B Q4_0 4 pp1024@d4096 462.00 455.94 0.99
RX 9060 XT gemma 2B Q4_0 4 pp1024@d8192 436.76 437.16 1.00
RX 9060 XT gemma 2B Q4_0 4 pp1024@d16384 398.36 405.73 1.02
RX 9060 XT gemma 2B Q4_0 4 pp1024@d32768 343.28 368.71 1.07
RX 9060 XT gemma 2B Q4_0 8 pp1024 813.96 804.03 0.99
RX 9060 XT gemma 2B Q4_0 8 pp1024@d1024 793.26 799.46 1.01
RX 9060 XT gemma 2B Q4_0 8 pp1024@d2048 776.17 788.50 1.02
RX 9060 XT gemma 2B Q4_0 8 pp1024@d4096 739.38 771.84 1.04
RX 9060 XT gemma 2B Q4_0 8 pp1024@d8192 677.58 742.04 1.10
RX 9060 XT gemma 2B Q4_0 8 pp1024@d16384 582.69 695.22 1.19
RX 9060 XT gemma 2B Q4_0 8 pp1024@d32768 447.92 621.36 1.39
RX 9060 XT gemma 2B Q4_0 16 pp1024 1534.08 1517.28 0.99
RX 9060 XT gemma 2B Q4_0 16 pp1024@d1024 1470.10 1513.59 1.03
RX 9060 XT gemma 2B Q4_0 16 pp1024@d2048 1404.48 1491.95 1.06
RX 9060 XT gemma 2B Q4_0 16 pp1024@d4096 1298.53 1435.81 1.11
RX 9060 XT gemma 2B Q4_0 16 pp1024@d8192 1113.28 1352.22 1.21
RX 9060 XT gemma 2B Q4_0 16 pp1024@d16384 882.45 1221.14 1.38
RX 9060 XT gemma 2B Q4_0 16 pp1024@d32768 608.97 1013.05 1.66
RX 9060 XT gemma 2B Q4_0 32 pp1024 2495.77 2556.41 1.02
RX 9060 XT gemma 2B Q4_0 32 pp1024@d1024 2308.08 2513.54 1.09
RX 9060 XT gemma 2B Q4_0 32 pp1024@d2048 2138.33 2447.43 1.14
RX 9060 XT gemma 2B Q4_0 32 pp1024@d4096 1891.24 2313.25 1.22
RX 9060 XT gemma 2B Q4_0 32 pp1024@d8192 1523.24 2094.07 1.37
RX 9060 XT gemma 2B Q4_0 32 pp1024@d16384 1099.01 1770.91 1.61
RX 9060 XT gemma 2B Q4_0 32 pp1024@d32768 708.51 1353.49 1.91
RX 9060 XT gemma 2B Q4_0 64 pp1024 3681.43 3722.82 1.01
RX 9060 XT gemma 2B Q4_0 64 pp1024@d1024 3286.00 3605.65 1.10
RX 9060 XT gemma 2B Q4_0 64 pp1024@d2048 2951.91 3477.13 1.18
RX 9060 XT gemma 2B Q4_0 64 pp1024@d4096 2493.95 3217.50 1.29
RX 9060 XT gemma 2B Q4_0 64 pp1024@d8192 1899.01 2837.30 1.49
RX 9060 XT gemma 2B Q4_0 64 pp1024@d16384 1291.77 2286.57 1.77
RX 9060 XT gemma 2B Q4_0 64 pp1024@d32768 784.82 1646.99 2.10
RX 9060 XT gemma 2B Q4_0 128 pp1024 4925.28 4979.60 1.01
RX 9060 XT gemma 2B Q4_0 128 pp1024@d1024 4321.66 4739.79 1.10
RX 9060 XT gemma 2B Q4_0 128 pp1024@d2048 3884.52 4467.60 1.15
RX 9060 XT gemma 2B Q4_0 128 pp1024@d4096 3231.15 4077.59 1.26
RX 9060 XT gemma 2B Q4_0 128 pp1024@d8192 2429.50 3451.47 1.42
RX 9060 XT gemma 2B Q4_0 128 pp1024@d16384 1637.36 2663.55 1.63
RX 9060 XT gemma 2B Q4_0 128 pp1024@d32768 966.82 1834.82 1.90
RX 9060 XT gemma 2B Q4_0 256 pp1024 7290.98 7674.55 1.05
RX 9060 XT gemma 2B Q4_0 256 pp1024@d1024 6013.49 7099.76 1.18
RX 9060 XT gemma 2B Q4_0 256 pp1024@d2048 5185.53 6663.30 1.28
RX 9060 XT gemma 2B Q4_0 256 pp1024@d4096 4099.34 6042.98 1.47
RX 9060 XT gemma 2B Q4_0 256 pp1024@d8192 2910.29 5091.22 1.75
RX 9060 XT gemma 2B Q4_0 256 pp1024@d16384 1862.82 3942.74 2.12
RX 9060 XT gemma 2B Q4_0 256 pp1024@d32768 1071.32 2708.68 2.53
RX 9060 XT gemma 2B Q4_0 512 pp1024 7538.15 8075.33 1.07
RX 9060 XT gemma 2B Q4_0 512 pp1024@d1024 6284.13 7581.18 1.21
RX 9060 XT gemma 2B Q4_0 512 pp1024@d2048 5439.49 7169.73 1.32
RX 9060 XT gemma 2B Q4_0 512 pp1024@d4096 4267.23 6458.02 1.51
RX 9060 XT gemma 2B Q4_0 512 pp1024@d8192 3024.31 5373.58 1.78
RX 9060 XT gemma 2B Q4_0 512 pp1024@d16384 1922.31 4017.62 2.09
RX 9060 XT gemma 2B Q4_0 512 pp1024@d32768 1099.73 2729.53 2.48
RX 9060 XT llama 1B Q4_0 1 pp1024 326.87 328.13 1.00
RX 9060 XT llama 1B Q4_0 1 pp1024@d1024 313.94 316.89 1.01
RX 9060 XT llama 1B Q4_0 1 pp1024@d2048 303.64 306.31 1.01
RX 9060 XT llama 1B Q4_0 1 pp1024@d4096 285.31 285.68 1.00
RX 9060 XT llama 1B Q4_0 1 pp1024@d8192 256.54 256.90 1.00
RX 9060 XT llama 1B Q4_0 1 pp1024@d16384 211.00 211.71 1.00
RX 9060 XT llama 1B Q4_0 1 pp1024@d32768 155.10 155.09 1.00
RX 9060 XT llama 1B Q4_0 2 pp1024 411.48 413.35 1.00
RX 9060 XT llama 1B Q4_0 2 pp1024@d1024 397.65 397.89 1.00
RX 9060 XT llama 1B Q4_0 2 pp1024@d2048 388.50 389.31 1.00
RX 9060 XT llama 1B Q4_0 2 pp1024@d4096 368.65 369.20 1.00
RX 9060 XT llama 1B Q4_0 2 pp1024@d8192 333.23 334.24 1.00
RX 9060 XT llama 1B Q4_0 2 pp1024@d16384 279.96 280.07 1.00
RX 9060 XT llama 1B Q4_0 2 pp1024@d32768 210.77 210.79 1.00
RX 9060 XT llama 1B Q4_0 4 pp1024 604.69 607.44 1.00
RX 9060 XT llama 1B Q4_0 4 pp1024@d1024 590.53 592.13 1.00
RX 9060 XT llama 1B Q4_0 4 pp1024@d2048 582.25 583.17 1.00
RX 9060 XT llama 1B Q4_0 4 pp1024@d4096 562.41 563.12 1.00
RX 9060 XT llama 1B Q4_0 4 pp1024@d8192 526.75 527.40 1.00
RX 9060 XT llama 1B Q4_0 4 pp1024@d16384 465.37 465.15 1.00
RX 9060 XT llama 1B Q4_0 4 pp1024@d32768 376.68 376.29 1.00
RX 9060 XT llama 1B Q4_0 8 pp1024 1208.72 1215.55 1.01
RX 9060 XT llama 1B Q4_0 8 pp1024@d1024 1183.24 1185.52 1.00
RX 9060 XT llama 1B Q4_0 8 pp1024@d2048 1158.59 1162.43 1.00
RX 9060 XT llama 1B Q4_0 8 pp1024@d4096 1116.13 1116.34 1.00
RX 9060 XT llama 1B Q4_0 8 pp1024@d8192 1040.70 1042.10 1.00
RX 9060 XT llama 1B Q4_0 8 pp1024@d16384 915.93 915.60 1.00
RX 9060 XT llama 1B Q4_0 8 pp1024@d32768 745.76 753.23 1.01
RX 9060 XT llama 1B Q4_0 16 pp1024 2603.47 2609.29 1.00
RX 9060 XT llama 1B Q4_0 16 pp1024@d1024 2517.27 2522.66 1.00
RX 9060 XT llama 1B Q4_0 16 pp1024@d2048 2465.07 2478.37 1.01
RX 9060 XT llama 1B Q4_0 16 pp1024@d4096 2375.96 2370.15 1.00
RX 9060 XT llama 1B Q4_0 16 pp1024@d8192 2201.16 2200.86 1.00
RX 9060 XT llama 1B Q4_0 16 pp1024@d16384 1911.19 1905.40 1.00
RX 9060 XT llama 1B Q4_0 16 pp1024@d32768 1514.74 1511.23 1.00
RX 9060 XT llama 1B Q4_0 32 pp1024 3843.24 3892.46 1.01
RX 9060 XT llama 1B Q4_0 32 pp1024@d1024 3737.16 3757.39 1.01
RX 9060 XT llama 1B Q4_0 32 pp1024@d2048 3663.59 3690.24 1.01
RX 9060 XT llama 1B Q4_0 32 pp1024@d4096 3525.29 3509.35 1.00
RX 9060 XT llama 1B Q4_0 32 pp1024@d8192 3306.39 3241.91 0.98
RX 9060 XT llama 1B Q4_0 32 pp1024@d16384 2777.85 2773.07 1.00
RX 9060 XT llama 1B Q4_0 32 pp1024@d32768 2121.97 2119.66 1.00
RX 9060 XT llama 1B Q4_0 64 pp1024 5335.02 5431.01 1.02
RX 9060 XT llama 1B Q4_0 64 pp1024@d1024 5219.23 5179.40 0.99
RX 9060 XT llama 1B Q4_0 64 pp1024@d2048 4979.65 5029.23 1.01
RX 9060 XT llama 1B Q4_0 64 pp1024@d4096 4681.70 4680.36 1.00
RX 9060 XT llama 1B Q4_0 64 pp1024@d8192 4132.44 4162.66 1.01
RX 9060 XT llama 1B Q4_0 64 pp1024@d16384 3372.08 3364.45 1.00
RX 9060 XT llama 1B Q4_0 64 pp1024@d32768 2413.12 2416.83 1.00
RX 9060 XT llama 1B Q4_0 128 pp1024 8225.01 8210.65 1.00
RX 9060 XT llama 1B Q4_0 128 pp1024@d1024 7756.62 7654.44 0.99
RX 9060 XT llama 1B Q4_0 128 pp1024@d2048 7189.73 7166.29 1.00
RX 9060 XT llama 1B Q4_0 128 pp1024@d4096 6542.08 6511.14 1.00
RX 9060 XT llama 1B Q4_0 128 pp1024@d8192 5507.57 5528.07 1.00
RX 9060 XT llama 1B Q4_0 128 pp1024@d16384 4186.53 4174.03 1.00
RX 9060 XT llama 1B Q4_0 128 pp1024@d32768 2789.97 2798.74 1.00
RX 9060 XT llama 1B Q4_0 256 pp1024 12965.20 13003.61 1.00
RX 9060 XT llama 1B Q4_0 256 pp1024@d1024 11452.63 11463.06 1.00
RX 9060 XT llama 1B Q4_0 256 pp1024@d2048 10539.07 10549.90 1.00
RX 9060 XT llama 1B Q4_0 256 pp1024@d4096 9192.08 9144.06 0.99
RX 9060 XT llama 1B Q4_0 256 pp1024@d8192 7259.93 7274.18 1.00
RX 9060 XT llama 1B Q4_0 256 pp1024@d16384 5078.09 5060.10 1.00
RX 9060 XT llama 1B Q4_0 256 pp1024@d32768 3162.75 3156.67 1.00
RX 9060 XT llama 1B Q4_0 512 pp1024 12388.58 12430.03 1.00
RX 9060 XT llama 1B Q4_0 512 pp1024@d1024 11988.93 11833.29 0.99
RX 9060 XT llama 1B Q4_0 512 pp1024@d2048 11252.73 11289.88 1.00
RX 9060 XT llama 1B Q4_0 512 pp1024@d4096 9744.95 9743.92 1.00
RX 9060 XT llama 1B Q4_0 512 pp1024@d8192 7656.92 7671.60 1.00
RX 9060 XT llama 1B Q4_0 512 pp1024@d16384 5247.30 5200.23 0.99
RX 9060 XT llama 1B Q4_0 512 pp1024@d32768 3195.56 3195.23 1.00
RX 9060 XT llama 8B Q4_0 1 pp1024 68.21 68.24 1.00
RX 9060 XT llama 8B Q4_0 1 pp1024@d1024 66.15 66.19 1.00
RX 9060 XT llama 8B Q4_0 1 pp1024@d2048 64.34 64.41 1.00
RX 9060 XT llama 8B Q4_0 1 pp1024@d4096 61.09 61.01 1.00
RX 9060 XT llama 8B Q4_0 1 pp1024@d8192 55.44 55.53 1.00
RX 9060 XT llama 8B Q4_0 1 pp1024@d16384 46.69 46.72 1.00
RX 9060 XT llama 8B Q4_0 1 pp1024@d32768 35.56 35.56 1.00
RX 9060 XT llama 8B Q4_0 2 pp1024 100.66 100.52 1.00
RX 9060 XT llama 8B Q4_0 2 pp1024@d1024 98.28 98.20 1.00
RX 9060 XT llama 8B Q4_0 2 pp1024@d2048 96.34 96.35 1.00
RX 9060 XT llama 8B Q4_0 2 pp1024@d4096 92.76 92.67 1.00
RX 9060 XT llama 8B Q4_0 2 pp1024@d8192 86.07 86.04 1.00
RX 9060 XT llama 8B Q4_0 2 pp1024@d16384 74.97 75.06 1.00
RX 9060 XT llama 8B Q4_0 2 pp1024@d32768 59.84 59.80 1.00
RX 9060 XT llama 8B Q4_0 4 pp1024 182.37 181.86 1.00
RX 9060 XT llama 8B Q4_0 4 pp1024@d1024 177.02 176.88 1.00
RX 9060 XT llama 8B Q4_0 4 pp1024@d2048 173.13 172.84 1.00
RX 9060 XT llama 8B Q4_0 4 pp1024@d4096 165.30 165.19 1.00
RX 9060 XT llama 8B Q4_0 4 pp1024@d8192 151.11 151.20 1.00
RX 9060 XT llama 8B Q4_0 4 pp1024@d16384 130.26 130.37 1.00
RX 9060 XT llama 8B Q4_0 4 pp1024@d32768 100.95 100.74 1.00
RX 9060 XT llama 8B Q4_0 8 pp1024 199.43 199.17 1.00
RX 9060 XT llama 8B Q4_0 8 pp1024@d1024 196.56 196.15 1.00
RX 9060 XT llama 8B Q4_0 8 pp1024@d2048 194.44 194.35 1.00
RX 9060 XT llama 8B Q4_0 8 pp1024@d4096 190.47 189.79 1.00
RX 9060 XT llama 8B Q4_0 8 pp1024@d8192 182.34 182.97 1.00
RX 9060 XT llama 8B Q4_0 8 pp1024@d16384 169.30 169.50 1.00
RX 9060 XT llama 8B Q4_0 8 pp1024@d32768 147.88 147.57 1.00
RX 9060 XT llama 8B Q4_0 16 pp1024 660.73 660.13 1.00
RX 9060 XT llama 8B Q4_0 16 pp1024@d1024 641.12 638.67 1.00
RX 9060 XT llama 8B Q4_0 16 pp1024@d2048 624.75 624.18 1.00
RX 9060 XT llama 8B Q4_0 16 pp1024@d4096 595.31 593.88 1.00
RX 9060 XT llama 8B Q4_0 16 pp1024@d8192 542.94 543.14 1.00
RX 9060 XT llama 8B Q4_0 16 pp1024@d16384 462.64 463.15 1.00
RX 9060 XT llama 8B Q4_0 16 pp1024@d32768 357.99 357.17 1.00
RX 9060 XT llama 8B Q4_0 32 pp1024 990.51 989.06 1.00
RX 9060 XT llama 8B Q4_0 32 pp1024@d1024 953.80 949.24 1.00
RX 9060 XT llama 8B Q4_0 32 pp1024@d2048 921.67 919.02 1.00
RX 9060 XT llama 8B Q4_0 32 pp1024@d4096 866.31 863.64 1.00
RX 9060 XT llama 8B Q4_0 32 pp1024@d8192 769.47 769.91 1.00
RX 9060 XT llama 8B Q4_0 32 pp1024@d16384 628.25 628.19 1.00
RX 9060 XT llama 8B Q4_0 32 pp1024@d32768 464.45 462.46 1.00
RX 9060 XT llama 8B Q4_0 64 pp1024 1610.44 1609.12 1.00
RX 9060 XT llama 8B Q4_0 64 pp1024@d1024 1507.52 1492.75 0.99
RX 9060 XT llama 8B Q4_0 64 pp1024@d2048 1447.53 1435.53 0.99
RX 9060 XT llama 8B Q4_0 64 pp1024@d4096 1320.89 1314.26 0.99
RX 9060 XT llama 8B Q4_0 64 pp1024@d8192 1087.31 1089.54 1.00
RX 9060 XT llama 8B Q4_0 64 pp1024@d16384 832.08 831.96 1.00
RX 9060 XT llama 8B Q4_0 64 pp1024@d32768 569.23 567.28 1.00
RX 9060 XT llama 8B Q4_0 128 pp1024 2359.64 2355.97 1.00
RX 9060 XT llama 8B Q4_0 128 pp1024@d1024 2142.74 2125.51 0.99
RX 9060 XT llama 8B Q4_0 128 pp1024@d2048 2005.35 1992.39 0.99
RX 9060 XT llama 8B Q4_0 128 pp1024@d4096 1771.69 1759.99 0.99
RX 9060 XT llama 8B Q4_0 128 pp1024@d8192 1374.73 1380.62 1.00
RX 9060 XT llama 8B Q4_0 128 pp1024@d16384 1003.99 1003.00 1.00
RX 9060 XT llama 8B Q4_0 128 pp1024@d32768 648.49 646.66 1.00
RX 9060 XT llama 8B Q4_0 256 pp1024 2612.44 2617.89 1.00
RX 9060 XT llama 8B Q4_0 256 pp1024@d1024 2347.26 2361.80 1.01
RX 9060 XT llama 8B Q4_0 256 pp1024@d2048 2185.63 2204.24 1.01
RX 9060 XT llama 8B Q4_0 256 pp1024@d4096 1934.63 1953.74 1.01
RX 9060 XT llama 8B Q4_0 256 pp1024@d8192 1477.02 1614.86 1.09
RX 9060 XT llama 8B Q4_0 256 pp1024@d16384 1056.31 1187.05 1.12
RX 9060 XT llama 8B Q4_0 256 pp1024@d32768 674.72 780.75 1.16
RX 9060 XT llama 8B Q4_0 512 pp1024 2700.30 2722.46 1.01
RX 9060 XT llama 8B Q4_0 512 pp1024@d1024 2458.57 2483.76 1.01
RX 9060 XT llama 8B Q4_0 512 pp1024@d2048 2287.08 2308.23 1.01
RX 9060 XT llama 8B Q4_0 512 pp1024@d4096 2009.40 2019.50 1.01
RX 9060 XT llama 8B Q4_0 512 pp1024@d8192 1502.39 1651.06 1.10
RX 9060 XT llama 8B Q4_0 512 pp1024@d16384 1056.29 1197.76 1.13
RX 9060 XT llama 8B Q4_0 512 pp1024@d32768 674.64 783.05 1.16
JG vs. PW
GPU Model Microbatch size Test t/s a3b28f9 t/s 0ce2938 Speedup
RX 9060 XT gemma 2B Q4_0 1 pp1024 199.51 198.92 1.00
RX 9060 XT gemma 2B Q4_0 1 pp1024@d1024 194.96 195.75 1.00
RX 9060 XT gemma 2B Q4_0 1 pp1024@d2048 192.71 192.97 1.00
RX 9060 XT gemma 2B Q4_0 1 pp1024@d4096 185.87 185.69 1.00
RX 9060 XT gemma 2B Q4_0 1 pp1024@d8192 176.34 176.69 1.00
RX 9060 XT gemma 2B Q4_0 1 pp1024@d16384 158.58 158.96 1.00
RX 9060 XT gemma 2B Q4_0 1 pp1024@d32768 133.01 133.54 1.00
RX 9060 XT gemma 2B Q4_0 2 pp1024 244.16 245.62 1.01
RX 9060 XT gemma 2B Q4_0 2 pp1024@d1024 241.04 237.92 0.99
RX 9060 XT gemma 2B Q4_0 2 pp1024@d2048 235.81 231.55 0.98
RX 9060 XT gemma 2B Q4_0 2 pp1024@d4096 234.11 234.05 1.00
RX 9060 XT gemma 2B Q4_0 2 pp1024@d8192 227.26 227.32 1.00
RX 9060 XT gemma 2B Q4_0 2 pp1024@d16384 214.44 214.02 1.00
RX 9060 XT gemma 2B Q4_0 2 pp1024@d32768 191.32 191.30 1.00
RX 9060 XT gemma 2B Q4_0 4 pp1024 490.16 488.44 1.00
RX 9060 XT gemma 2B Q4_0 4 pp1024@d1024 473.93 473.32 1.00
RX 9060 XT gemma 2B Q4_0 4 pp1024@d2048 469.08 469.23 1.00
RX 9060 XT gemma 2B Q4_0 4 pp1024@d4096 455.97 455.94 1.00
RX 9060 XT gemma 2B Q4_0 4 pp1024@d8192 436.99 437.16 1.00
RX 9060 XT gemma 2B Q4_0 4 pp1024@d16384 406.74 405.73 1.00
RX 9060 XT gemma 2B Q4_0 4 pp1024@d32768 369.11 368.71 1.00
RX 9060 XT gemma 2B Q4_0 8 pp1024 807.04 804.03 1.00
RX 9060 XT gemma 2B Q4_0 8 pp1024@d1024 796.88 799.46 1.00
RX 9060 XT gemma 2B Q4_0 8 pp1024@d2048 789.05 788.50 1.00
RX 9060 XT gemma 2B Q4_0 8 pp1024@d4096 770.05 771.84 1.00
RX 9060 XT gemma 2B Q4_0 8 pp1024@d8192 745.12 742.04 1.00
RX 9060 XT gemma 2B Q4_0 8 pp1024@d16384 697.48 695.22 1.00
RX 9060 XT gemma 2B Q4_0 8 pp1024@d32768 618.61 621.36 1.00
RX 9060 XT gemma 2B Q4_0 16 pp1024 1541.23 1517.28 0.98
RX 9060 XT gemma 2B Q4_0 16 pp1024@d1024 1514.17 1513.59 1.00
RX 9060 XT gemma 2B Q4_0 16 pp1024@d2048 1488.73 1491.95 1.00
RX 9060 XT gemma 2B Q4_0 16 pp1024@d4096 1431.61 1435.81 1.00
RX 9060 XT gemma 2B Q4_0 16 pp1024@d8192 1350.10 1352.22 1.00
RX 9060 XT gemma 2B Q4_0 16 pp1024@d16384 1225.79 1221.14 1.00
RX 9060 XT gemma 2B Q4_0 16 pp1024@d32768 1007.76 1013.05 1.01
RX 9060 XT gemma 2B Q4_0 32 pp1024 2598.26 2556.41 0.98
RX 9060 XT gemma 2B Q4_0 32 pp1024@d1024 2526.66 2513.54 0.99
RX 9060 XT gemma 2B Q4_0 32 pp1024@d2048 2463.82 2447.43 0.99
RX 9060 XT gemma 2B Q4_0 32 pp1024@d4096 2299.89 2313.25 1.01
RX 9060 XT gemma 2B Q4_0 32 pp1024@d8192 2092.87 2094.07 1.00
RX 9060 XT gemma 2B Q4_0 32 pp1024@d16384 1779.51 1770.91 1.00
RX 9060 XT gemma 2B Q4_0 32 pp1024@d32768 1353.04 1353.49 1.00
RX 9060 XT gemma 2B Q4_0 64 pp1024 3768.68 3722.82 0.99
RX 9060 XT gemma 2B Q4_0 64 pp1024@d1024 3628.76 3605.65 0.99
RX 9060 XT gemma 2B Q4_0 64 pp1024@d2048 3462.08 3477.13 1.00
RX 9060 XT gemma 2B Q4_0 64 pp1024@d4096 3212.83 3217.50 1.00
RX 9060 XT gemma 2B Q4_0 64 pp1024@d8192 2844.27 2837.30 1.00
RX 9060 XT gemma 2B Q4_0 64 pp1024@d16384 2286.47 2286.57 1.00
RX 9060 XT gemma 2B Q4_0 64 pp1024@d32768 1639.84 1646.99 1.00
RX 9060 XT gemma 2B Q4_0 128 pp1024 5037.19 4979.60 0.99
RX 9060 XT gemma 2B Q4_0 128 pp1024@d1024 4701.78 4739.79 1.01
RX 9060 XT gemma 2B Q4_0 128 pp1024@d2048 4465.72 4467.60 1.00
RX 9060 XT gemma 2B Q4_0 128 pp1024@d4096 4080.08 4077.59 1.00
RX 9060 XT gemma 2B Q4_0 128 pp1024@d8192 3465.77 3451.47 1.00
RX 9060 XT gemma 2B Q4_0 128 pp1024@d16384 2669.78 2663.55 1.00
RX 9060 XT gemma 2B Q4_0 128 pp1024@d32768 1828.07 1834.82 1.00
RX 9060 XT gemma 2B Q4_0 256 pp1024 7670.59 7674.55 1.00
RX 9060 XT gemma 2B Q4_0 256 pp1024@d1024 7042.40 7099.76 1.01
RX 9060 XT gemma 2B Q4_0 256 pp1024@d2048 6697.28 6663.30 0.99
RX 9060 XT gemma 2B Q4_0 256 pp1024@d4096 6031.23 6042.98 1.00
RX 9060 XT gemma 2B Q4_0 256 pp1024@d8192 5102.77 5091.22 1.00
RX 9060 XT gemma 2B Q4_0 256 pp1024@d16384 3937.68 3942.74 1.00
RX 9060 XT gemma 2B Q4_0 256 pp1024@d32768 2686.63 2708.68 1.01
RX 9060 XT gemma 2B Q4_0 512 pp1024 8163.80 8075.33 0.99
RX 9060 XT gemma 2B Q4_0 512 pp1024@d1024 7601.69 7581.18 1.00
RX 9060 XT gemma 2B Q4_0 512 pp1024@d2048 7168.91 7169.73 1.00
RX 9060 XT gemma 2B Q4_0 512 pp1024@d4096 6480.79 6458.02 1.00
RX 9060 XT gemma 2B Q4_0 512 pp1024@d8192 5388.91 5373.58 1.00
RX 9060 XT gemma 2B Q4_0 512 pp1024@d16384 4049.56 4017.62 0.99
RX 9060 XT gemma 2B Q4_0 512 pp1024@d32768 2713.10 2729.53 1.01
RX 9060 XT llama 1B Q4_0 1 pp1024 329.53 328.13 1.00
RX 9060 XT llama 1B Q4_0 1 pp1024@d1024 315.89 316.89 1.00
RX 9060 XT llama 1B Q4_0 1 pp1024@d2048 303.91 306.31 1.01
RX 9060 XT llama 1B Q4_0 1 pp1024@d4096 287.63 285.68 0.99
RX 9060 XT llama 1B Q4_0 1 pp1024@d8192 258.10 256.90 1.00
RX 9060 XT llama 1B Q4_0 1 pp1024@d16384 211.25 211.71 1.00
RX 9060 XT llama 1B Q4_0 1 pp1024@d32768 155.04 155.09 1.00
RX 9060 XT llama 1B Q4_0 2 pp1024 412.22 413.35 1.00
RX 9060 XT llama 1B Q4_0 2 pp1024@d1024 398.61 397.89 1.00
RX 9060 XT llama 1B Q4_0 2 pp1024@d2048 389.11 389.31 1.00
RX 9060 XT llama 1B Q4_0 2 pp1024@d4096 369.63 369.20 1.00
RX 9060 XT llama 1B Q4_0 2 pp1024@d8192 333.75 334.24 1.00
RX 9060 XT llama 1B Q4_0 2 pp1024@d16384 280.08 280.07 1.00
RX 9060 XT llama 1B Q4_0 2 pp1024@d32768 210.57 210.79 1.00
RX 9060 XT llama 1B Q4_0 4 pp1024 605.93 607.44 1.00
RX 9060 XT llama 1B Q4_0 4 pp1024@d1024 594.39 592.13 1.00
RX 9060 XT llama 1B Q4_0 4 pp1024@d2048 583.94 583.17 1.00
RX 9060 XT llama 1B Q4_0 4 pp1024@d4096 562.68 563.12 1.00
RX 9060 XT llama 1B Q4_0 4 pp1024@d8192 527.05 527.40 1.00
RX 9060 XT llama 1B Q4_0 4 pp1024@d16384 465.15 465.15 1.00
RX 9060 XT llama 1B Q4_0 4 pp1024@d32768 376.38 376.29 1.00
RX 9060 XT llama 1B Q4_0 8 pp1024 1215.98 1215.55 1.00
RX 9060 XT llama 1B Q4_0 8 pp1024@d1024 1189.28 1185.52 1.00
RX 9060 XT llama 1B Q4_0 8 pp1024@d2048 1159.71 1162.43 1.00
RX 9060 XT llama 1B Q4_0 8 pp1024@d4096 1117.16 1116.34 1.00
RX 9060 XT llama 1B Q4_0 8 pp1024@d8192 1040.65 1042.10 1.00
RX 9060 XT llama 1B Q4_0 8 pp1024@d16384 915.79 915.60 1.00
RX 9060 XT llama 1B Q4_0 8 pp1024@d32768 744.42 753.23 1.01
RX 9060 XT llama 1B Q4_0 16 pp1024 2595.57 2609.29 1.01
RX 9060 XT llama 1B Q4_0 16 pp1024@d1024 2527.75 2522.66 1.00
RX 9060 XT llama 1B Q4_0 16 pp1024@d2048 2466.36 2478.37 1.00
RX 9060 XT llama 1B Q4_0 16 pp1024@d4096 2373.25 2370.15 1.00
RX 9060 XT llama 1B Q4_0 16 pp1024@d8192 2201.90 2200.86 1.00
RX 9060 XT llama 1B Q4_0 16 pp1024@d16384 1907.22 1905.40 1.00
RX 9060 XT llama 1B Q4_0 16 pp1024@d32768 1512.98 1511.23 1.00
RX 9060 XT llama 1B Q4_0 32 pp1024 3859.03 3892.46 1.01
RX 9060 XT llama 1B Q4_0 32 pp1024@d1024 3779.37 3757.39 0.99
RX 9060 XT llama 1B Q4_0 32 pp1024@d2048 3667.18 3690.24 1.01
RX 9060 XT llama 1B Q4_0 32 pp1024@d4096 3508.37 3509.35 1.00
RX 9060 XT llama 1B Q4_0 32 pp1024@d8192 3245.74 3241.91 1.00
RX 9060 XT llama 1B Q4_0 32 pp1024@d16384 2768.55 2773.07 1.00
RX 9060 XT llama 1B Q4_0 32 pp1024@d32768 2115.00 2119.66 1.00
RX 9060 XT llama 1B Q4_0 64 pp1024 5354.61 5431.01 1.01
RX 9060 XT llama 1B Q4_0 64 pp1024@d1024 5219.14 5179.40 0.99
RX 9060 XT llama 1B Q4_0 64 pp1024@d2048 4979.69 5029.23 1.01
RX 9060 XT llama 1B Q4_0 64 pp1024@d4096 4685.37 4680.36 1.00
RX 9060 XT llama 1B Q4_0 64 pp1024@d8192 4159.06 4162.66 1.00
RX 9060 XT llama 1B Q4_0 64 pp1024@d16384 3365.11 3364.45 1.00
RX 9060 XT llama 1B Q4_0 64 pp1024@d32768 2403.37 2416.83 1.01
RX 9060 XT llama 1B Q4_0 128 pp1024 8084.94 8210.65 1.02
RX 9060 XT llama 1B Q4_0 128 pp1024@d1024 7739.91 7654.44 0.99
RX 9060 XT llama 1B Q4_0 128 pp1024@d2048 7152.98 7166.29 1.00
RX 9060 XT llama 1B Q4_0 128 pp1024@d4096 6534.04 6511.14 1.00
RX 9060 XT llama 1B Q4_0 128 pp1024@d8192 5518.92 5528.07 1.00
RX 9060 XT llama 1B Q4_0 128 pp1024@d16384 4201.58 4174.03 0.99
RX 9060 XT llama 1B Q4_0 128 pp1024@d32768 2783.58 2798.74 1.01
RX 9060 XT llama 1B Q4_0 256 pp1024 13001.92 13003.61 1.00
RX 9060 XT llama 1B Q4_0 256 pp1024@d1024 11462.87 11463.06 1.00
RX 9060 XT llama 1B Q4_0 256 pp1024@d2048 10416.98 10549.90 1.01
RX 9060 XT llama 1B Q4_0 256 pp1024@d4096 9153.15 9144.06 1.00
RX 9060 XT llama 1B Q4_0 256 pp1024@d8192 7211.04 7274.18 1.01
RX 9060 XT llama 1B Q4_0 256 pp1024@d16384 5086.47 5060.10 0.99
RX 9060 XT llama 1B Q4_0 256 pp1024@d32768 3153.39 3156.67 1.00
RX 9060 XT llama 1B Q4_0 512 pp1024 12692.93 12430.03 0.98
RX 9060 XT llama 1B Q4_0 512 pp1024@d1024 12100.80 11833.29 0.98
RX 9060 XT llama 1B Q4_0 512 pp1024@d2048 11171.83 11289.88 1.01
RX 9060 XT llama 1B Q4_0 512 pp1024@d4096 9590.98 9743.92 1.02
RX 9060 XT llama 1B Q4_0 512 pp1024@d8192 7280.01 7671.60 1.05
RX 9060 XT llama 1B Q4_0 512 pp1024@d16384 4949.73 5200.23 1.05
RX 9060 XT llama 1B Q4_0 512 pp1024@d32768 2993.52 3195.23 1.07
RX 9060 XT llama 8B Q4_0 1 pp1024 68.24 68.24 1.00
RX 9060 XT llama 8B Q4_0 1 pp1024@d1024 66.18 66.19 1.00
RX 9060 XT llama 8B Q4_0 1 pp1024@d2048 64.37 64.41 1.00
RX 9060 XT llama 8B Q4_0 1 pp1024@d4096 61.06 61.01 1.00
RX 9060 XT llama 8B Q4_0 1 pp1024@d8192 55.45 55.53 1.00
RX 9060 XT llama 8B Q4_0 1 pp1024@d16384 46.69 46.72 1.00
RX 9060 XT llama 8B Q4_0 1 pp1024@d32768 35.56 35.56 1.00
RX 9060 XT llama 8B Q4_0 2 pp1024 100.58 100.52 1.00
RX 9060 XT llama 8B Q4_0 2 pp1024@d1024 98.23 98.20 1.00
RX 9060 XT llama 8B Q4_0 2 pp1024@d2048 96.27 96.35 1.00
RX 9060 XT llama 8B Q4_0 2 pp1024@d4096 92.62 92.67 1.00
RX 9060 XT llama 8B Q4_0 2 pp1024@d8192 85.87 86.04 1.00
RX 9060 XT llama 8B Q4_0 2 pp1024@d16384 75.00 75.06 1.00
RX 9060 XT llama 8B Q4_0 2 pp1024@d32768 59.81 59.80 1.00
RX 9060 XT llama 8B Q4_0 4 pp1024 182.12 181.86 1.00
RX 9060 XT llama 8B Q4_0 4 pp1024@d1024 176.78 176.88 1.00
RX 9060 XT llama 8B Q4_0 4 pp1024@d2048 173.07 172.84 1.00
RX 9060 XT llama 8B Q4_0 4 pp1024@d4096 165.23 165.19 1.00
RX 9060 XT llama 8B Q4_0 4 pp1024@d8192 151.10 151.20 1.00
RX 9060 XT llama 8B Q4_0 4 pp1024@d16384 130.25 130.37 1.00
RX 9060 XT llama 8B Q4_0 4 pp1024@d32768 100.85 100.74 1.00
RX 9060 XT llama 8B Q4_0 8 pp1024 198.86 199.17 1.00
RX 9060 XT llama 8B Q4_0 8 pp1024@d1024 196.26 196.15 1.00
RX 9060 XT llama 8B Q4_0 8 pp1024@d2048 194.77 194.35 1.00
RX 9060 XT llama 8B Q4_0 8 pp1024@d4096 190.18 189.79 1.00
RX 9060 XT llama 8B Q4_0 8 pp1024@d8192 182.38 182.97 1.00
RX 9060 XT llama 8B Q4_0 8 pp1024@d16384 169.17 169.50 1.00
RX 9060 XT llama 8B Q4_0 8 pp1024@d32768 147.82 147.57 1.00
RX 9060 XT llama 8B Q4_0 16 pp1024 661.43 660.13 1.00
RX 9060 XT llama 8B Q4_0 16 pp1024@d1024 638.73 638.67 1.00
RX 9060 XT llama 8B Q4_0 16 pp1024@d2048 624.71 624.18 1.00
RX 9060 XT llama 8B Q4_0 16 pp1024@d4096 594.47 593.88 1.00
RX 9060 XT llama 8B Q4_0 16 pp1024@d8192 542.96 543.14 1.00
RX 9060 XT llama 8B Q4_0 16 pp1024@d16384 462.56 463.15 1.00
RX 9060 XT llama 8B Q4_0 16 pp1024@d32768 357.78 357.17 1.00
RX 9060 XT llama 8B Q4_0 32 pp1024 991.28 989.06 1.00
RX 9060 XT llama 8B Q4_0 32 pp1024@d1024 948.60 949.24 1.00
RX 9060 XT llama 8B Q4_0 32 pp1024@d2048 920.75 919.02 1.00
RX 9060 XT llama 8B Q4_0 32 pp1024@d4096 864.65 863.64 1.00
RX 9060 XT llama 8B Q4_0 32 pp1024@d8192 769.72 769.91 1.00
RX 9060 XT llama 8B Q4_0 32 pp1024@d16384 628.00 628.19 1.00
RX 9060 XT llama 8B Q4_0 32 pp1024@d32768 461.98 462.46 1.00
RX 9060 XT llama 8B Q4_0 64 pp1024 1604.77 1609.12 1.00
RX 9060 XT llama 8B Q4_0 64 pp1024@d1024 1504.59 1492.75 0.99
RX 9060 XT llama 8B Q4_0 64 pp1024@d2048 1444.09 1435.53 0.99
RX 9060 XT llama 8B Q4_0 64 pp1024@d4096 1318.74 1314.26 1.00
RX 9060 XT llama 8B Q4_0 64 pp1024@d8192 1087.43 1089.54 1.00
RX 9060 XT llama 8B Q4_0 64 pp1024@d16384 832.58 831.96 1.00
RX 9060 XT llama 8B Q4_0 64 pp1024@d32768 566.67 567.28 1.00
RX 9060 XT llama 8B Q4_0 128 pp1024 2340.84 2355.97 1.01
RX 9060 XT llama 8B Q4_0 128 pp1024@d1024 2142.86 2125.51 0.99
RX 9060 XT llama 8B Q4_0 128 pp1024@d2048 2005.38 1992.39 0.99
RX 9060 XT llama 8B Q4_0 128 pp1024@d4096 1775.94 1759.99 0.99
RX 9060 XT llama 8B Q4_0 128 pp1024@d8192 1377.59 1380.62 1.00
RX 9060 XT llama 8B Q4_0 128 pp1024@d16384 1000.71 1003.00 1.00
RX 9060 XT llama 8B Q4_0 128 pp1024@d32768 646.06 646.66 1.00
RX 9060 XT llama 8B Q4_0 256 pp1024 2595.07 2617.89 1.01
RX 9060 XT llama 8B Q4_0 256 pp1024@d1024 2372.22 2361.80 1.00
RX 9060 XT llama 8B Q4_0 256 pp1024@d2048 2206.63 2204.24 1.00
RX 9060 XT llama 8B Q4_0 256 pp1024@d4096 1967.04 1953.74 0.99
RX 9060 XT llama 8B Q4_0 256 pp1024@d8192 1607.39 1614.86 1.00
RX 9060 XT llama 8B Q4_0 256 pp1024@d16384 1186.13 1187.05 1.00
RX 9060 XT llama 8B Q4_0 256 pp1024@d32768 785.23 780.75 0.99
RX 9060 XT llama 8B Q4_0 512 pp1024 2701.59 2722.46 1.01
RX 9060 XT llama 8B Q4_0 512 pp1024@d1024 2490.01 2483.76 1.00
RX 9060 XT llama 8B Q4_0 512 pp1024@d2048 2301.61 2308.23 1.00
RX 9060 XT llama 8B Q4_0 512 pp1024@d4096 2031.68 2019.50 0.99
RX 9060 XT llama 8B Q4_0 512 pp1024@d8192 1641.36 1651.06 1.01
RX 9060 XT llama 8B Q4_0 512 pp1024@d16384 1196.08 1197.76 1.00
RX 9060 XT llama 8B Q4_0 512 pp1024@d32768 787.39 783.05 0.99

@JohannesGaessler

Copy link
Copy Markdown
Contributor

Sorry, the columns were swapped in the "PW vs. master" table, this should be the correct data.

@eugene-su

Copy link
Copy Markdown

Just leaving this here.

  1. Without the patch.

Device 0: AMD Radeon RX 9070 XT, gfx1201 (0x1201), VMM: no, Wave Size: 32, VRAM: 16304 MiB

model size test t/s
qwen35 27B IQ4_XS - 4.25 bpw 15.50 GiB pp2048 1282.84 ± 0.00
qwen35 27B IQ4_XS - 4.25 bpw 15.50 GiB tg64 29.82 ± 0.00
qwen35 27B IQ4_XS - 4.25 bpw 15.50 GiB pp2048 @ d16384 352.02 ± 0.00
qwen35 27B IQ4_XS - 4.25 bpw 15.50 GiB tg64 @ d16384 27.52 ± 0.00
qwen35 27B IQ4_XS - 4.25 bpw 15.50 GiB pp2048 @ d32768 262.59 ± 0.00
qwen35 27B IQ4_XS - 4.25 bpw 15.50 GiB tg64 @ d32768 19.83 ± 0.00
qwen35 27B IQ4_XS - 4.25 bpw 15.50 GiB pp2048 @ d65535 357.37 ± 0.00
qwen35 27B IQ4_XS - 4.25 bpw 15.50 GiB tg64 @ d65535 12.25 ± 0.00
  1. After the patch (a huge boost!)
    Device 0: AMD Radeon RX 9070 XT, gfx1201 (0x1201), VMM: no, Wave Size: 32, VRAM: 16304 MiB

Device 0: AMD Radeon RX 9070 XT, gfx1201 (0x1201), VMM: no, Wave Size: 32, VRAM: 16304 MiB

model size test t/s
qwen35 27B IQ4_XS - 4.25 bpw 15.50 GiB pp2048 1332.62 ± 0.00
qwen35 27B IQ4_XS - 4.25 bpw 15.50 GiB tg64 29.86 ± 0.00
qwen35 27B IQ4_XS - 4.25 bpw 15.50 GiB pp2048 @ d16384 428.81 ± 0.00
qwen35 27B IQ4_XS - 4.25 bpw 15.50 GiB tg64 @ d16384 23.86 ± 0.00
qwen35 27B IQ4_XS - 4.25 bpw 15.50 GiB pp2048 @ d32768 936.58 ± 0.00
qwen35 27B IQ4_XS - 4.25 bpw 15.50 GiB tg64 @ d32768 19.90 ± 0.00
qwen35 27B IQ4_XS - 4.25 bpw 15.50 GiB pp2048 @ d65535 727.25 ± 0.00
qwen35 27B IQ4_XS - 4.25 bpw 15.50 GiB tg64 @ d65535 12.16 ± 0.00

But I'm noticing a strange slowdown on pp2048 @ d16384.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

CUDA Related to the CUDA backend ggml changes relating to the ggml tensor library for machine learning testing Everything test related

Projects

None yet

Development

Successfully merging this pull request may close these issues.

6 participants