スクパネで入れていたので意訳のメモをそのままコピペしてもよかったのですが3イケ全員INT8なので意訳付けるのやめました
November 14, 2025 at 3:57 PM
スクパネで入れていたので意訳のメモをそのままコピペしてもよかったのですが3イケ全員INT8なので意訳付けるのやめました
그것도 그렇고 'RDNA3니까 FSR4는 불가능한가'라는 얘기도 국내 웹에 돌던데, RDNA3에서 FSR4 int8 돌리는 sdk 공개된거 가지고 지금 사람들 씹고뜯고맛보고 하는거 보면, 그것도 밸브가 미리 어느 정도 튜닝 해주지 않으려나...? 싶기도 하고...
November 13, 2025 at 1:24 AM
그것도 그렇고 'RDNA3니까 FSR4는 불가능한가'라는 얘기도 국내 웹에 돌던데, RDNA3에서 FSR4 int8 돌리는 sdk 공개된거 가지고 지금 사람들 씹고뜯고맛보고 하는거 보면, 그것도 밸브가 미리 어느 정도 튜닝 해주지 않으려나...? 싶기도 하고...
I mean, the files were there since the Open beta when I tried FSR4 int8 on it so i'm not really surprised
November 12, 2025 at 4:54 PM
I mean, the files were there since the Open beta when I tried FSR4 int8 on it so i'm not really surprised
Dynamic int8 quantization with Quanto offers the best trade-off for Whisper-small, reducing model size by 57% with minimal performance loss on LibriSpeech.
Quantizing Whisper-small: How design choices affect ASR performance
Arthur Söhler, Julian Irigoyen, Andreas Søeborg Kirkedal
arxiv.org
November 12, 2025 at 10:16 AM
Dynamic int8 quantization with Quanto offers the best trade-off for Whisper-small, reducing model size by 57% with minimal performance loss on LibriSpeech.
📰 AMD “Zen 6” Hadir dengan AVX-512 FP16, VNNI INT8, dan Instruksi Baru untuk Akselerasi AI & HPC
👉 Baca artikel lengkap di sini: https://ahmandonk.com/2025/11/11/amd-zen-6-avx512-fp16-vnni-int8/
#ai #amd ##aii#amdu#arsitektur##avx#fp16#cpuc#fp16t#hpci#instruksii#teknologin#6ef="/hashtag/vnni" class="hover:underline text-blue-600 dark:text-sky-400 no-card-link">#vnni href="/hashtag/zen" class="hover:underline text-blue-600 dark:text-sky-400 no-card-link">#zen
👉 Baca artikel lengkap di sini: https://ahmandonk.com/2025/11/11/amd-zen-6-avx512-fp16-vnni-int8/
#ai #amd ##aii#amdu#arsitektur##avx#fp16#cpuc#fp16t#hpci#instruksii#teknologin#6ef="/hashtag/vnni" class="hover:underline text-blue-600 dark:text-sky-400 no-card-link">#vnni href="/hashtag/zen" class="hover:underline text-blue-600 dark:text-sky-400 no-card-link">#zen
November 11, 2025 at 7:20 AM
📰 AMD “Zen 6” Hadir dengan AVX-512 FP16, VNNI INT8, dan Instruksi Baru untuk Akselerasi AI & HPC
👉 Baca artikel lengkap di sini: https://ahmandonk.com/2025/11/11/amd-zen-6-avx512-fp16-vnni-int8/
#ai #amd ##aii#amdu#arsitektur##avx#fp16#cpuc#fp16t#hpci#instruksii#teknologin#6ef="/hashtag/vnni" class="hover:underline text-blue-600 dark:text-sky-400 no-card-link">#vnni href="/hashtag/zen" class="hover:underline text-blue-600 dark:text-sky-400 no-card-link">#zen
👉 Baca artikel lengkap di sini: https://ahmandonk.com/2025/11/11/amd-zen-6-avx512-fp16-vnni-int8/
#ai #amd ##aii#amdu#arsitektur##avx#fp16#cpuc#fp16t#hpci#instruksii#teknologin#6ef="/hashtag/vnni" class="hover:underline text-blue-600 dark:text-sky-400 no-card-link">#vnni href="/hashtag/zen" class="hover:underline text-blue-600 dark:text-sky-400 no-card-link">#zen
the amount of accuracy you can get from fp32->int8 with QAT is kinda surprising, it's something like 1% error iirc (though a lot noisier outputs), 0.1% with int16
November 10, 2025 at 6:27 AM
the amount of accuracy you can get from fp32->int8 with QAT is kinda surprising, it's something like 1% error iirc (though a lot noisier outputs), 0.1% with int16
I'm like 90% confident most quantized LLMs have some kind of brain damage bc smaller networks get fully lobotomized, it's not even "oh it's worse but it's ok," like fully broken if you even try to quantize them for full int8 inference
November 10, 2025 at 6:17 AM
I'm like 90% confident most quantized LLMs have some kind of brain damage bc smaller networks get fully lobotomized, it's not even "oh it's worse but it's ok," like fully broken if you even try to quantize them for full int8 inference
actually if you int8 quatnize your vectors then the space partition tree literally is a trie
November 7, 2025 at 4:17 AM
actually if you int8 quatnize your vectors then the space partition tree literally is a trie
Oh nice! Curious how much RAM you have on your Mac. I'm using Whisper for Home Assistant, but it's on a NAS so I have to use base-int8 model to get decent processing speeds. Considering getting a dedicated Mac Mini just to run Whisper and Piper
November 6, 2025 at 4:12 PM
Oh nice! Curious how much RAM you have on your Mac. I'm using Whisper for Home Assistant, but it's on a NAS so I have to use base-int8 model to get decent processing speeds. Considering getting a dedicated Mac Mini just to run Whisper and Piper
Hay tantas unidades int8 en B200 que fp32 emulado va más rápido que fp32 nativo.
=>
"Unlocking Tensor Core Performance with Floating Point Emulation in cuBLAS", Oct 24, 2025 developer.nvidia.com/blog/unlocki...
NVIDIA CUDA Toolkit 13.0 Update 2
FP64 emulation with Blackwell INT8 tensor cores
Sep 16 bsky.app/profile/ogaw...
Jul 1 bsky.app/profile/ogaw...
K. Ozaki, Jul 2
"Unlocking Tensor Core Performance with Floating Point Emulation in cuBLAS", Oct 24, 2025 developer.nvidia.com/blog/unlocki...
NVIDIA CUDA Toolkit 13.0 Update 2
FP64 emulation with Blackwell INT8 tensor cores
Sep 16 bsky.app/profile/ogaw...
Jul 1 bsky.app/profile/ogaw...
K. Ozaki, Jul 2
November 5, 2025 at 1:45 PM
Hay tantas unidades int8 en B200 que fp32 emulado va más rápido que fp32 nativo.
AMD dropping RDNA1/2 in the driver following the leaked INT8 FSR4 dll that they were pretending didn't exist is certainly a choice for a company that's on the backfoot
fuck off, guys.
fuck off, guys.
October 30, 2025 at 5:16 PM
AMD dropping RDNA1/2 in the driver following the leaked INT8 FSR4 dll that they were pretending didn't exist is certainly a choice for a company that's on the backfoot
fuck off, guys.
fuck off, guys.
=>
The Convergence of HPC, AI, & Quantum: Towards FugakuNEXT “made with Japan”, S. Matsuoka, Riken, HPC UF Paris, Oct 7, 2025 www.hpcuserforum.com/wp-content/u...
NVIDIA, Oct 24 bsky.app/profile/ogaw...
FP64 w/ INT8
近藤 正章、R-CCS、10月6日 bsky.app/profile/ogaw...
富士通とNVIDIA、10月3日
MONAXA-X (1.4nm)、8月22日
The Convergence of HPC, AI, & Quantum: Towards FugakuNEXT “made with Japan”, S. Matsuoka, Riken, HPC UF Paris, Oct 7, 2025 www.hpcuserforum.com/wp-content/u...
NVIDIA, Oct 24 bsky.app/profile/ogaw...
FP64 w/ INT8
近藤 正章、R-CCS、10月6日 bsky.app/profile/ogaw...
富士通とNVIDIA、10月3日
MONAXA-X (1.4nm)、8月22日
October 30, 2025 at 8:51 AM
=>
The Convergence of HPC, AI, & Quantum: Towards FugakuNEXT “made with Japan”, S. Matsuoka, Riken, HPC UF Paris, Oct 7, 2025 www.hpcuserforum.com/wp-content/u...
NVIDIA, Oct 24 bsky.app/profile/ogaw...
FP64 w/ INT8
近藤 正章、R-CCS、10月6日 bsky.app/profile/ogaw...
富士通とNVIDIA、10月3日
MONAXA-X (1.4nm)、8月22日
The Convergence of HPC, AI, & Quantum: Towards FugakuNEXT “made with Japan”, S. Matsuoka, Riken, HPC UF Paris, Oct 7, 2025 www.hpcuserforum.com/wp-content/u...
NVIDIA, Oct 24 bsky.app/profile/ogaw...
FP64 w/ INT8
近藤 正章、R-CCS、10月6日 bsky.app/profile/ogaw...
富士通とNVIDIA、10月3日
MONAXA-X (1.4nm)、8月22日
=>
"Unlocking Tensor Core Performance with Floating Point Emulation in cuBLAS", Oct 24, 2025 developer.nvidia.com/blog/unlocki...
NVIDIA CUDA Toolkit 13.0 Update 2
FP64 emulation with Blackwell INT8 tensor cores
Sep 16 bsky.app/profile/ogaw...
Jul 1 bsky.app/profile/ogaw...
K. Ozaki, Jul 2
"Unlocking Tensor Core Performance with Floating Point Emulation in cuBLAS", Oct 24, 2025 developer.nvidia.com/blog/unlocki...
NVIDIA CUDA Toolkit 13.0 Update 2
FP64 emulation with Blackwell INT8 tensor cores
Sep 16 bsky.app/profile/ogaw...
Jul 1 bsky.app/profile/ogaw...
K. Ozaki, Jul 2
October 26, 2025 at 2:40 PM
=>
"Unlocking Tensor Core Performance with Floating Point Emulation in cuBLAS", Oct 24, 2025 developer.nvidia.com/blog/unlocki...
NVIDIA CUDA Toolkit 13.0 Update 2
FP64 emulation with Blackwell INT8 tensor cores
Sep 16 bsky.app/profile/ogaw...
Jul 1 bsky.app/profile/ogaw...
K. Ozaki, Jul 2
"Unlocking Tensor Core Performance with Floating Point Emulation in cuBLAS", Oct 24, 2025 developer.nvidia.com/blog/unlocki...
NVIDIA CUDA Toolkit 13.0 Update 2
FP64 emulation with Blackwell INT8 tensor cores
Sep 16 bsky.app/profile/ogaw...
Jul 1 bsky.app/profile/ogaw...
K. Ozaki, Jul 2
Silent Hill f - FSR 4 INT8
https://www.youtube.com/watch?v=h2Lg7FiKcpE
#gaming #silenhillf #rdna3 #fsr4
https://www.youtube.com/watch?v=h2Lg7FiKcpE
#gaming #silenhillf #rdna3 #fsr4
October 24, 2025 at 1:33 PM
Silent Hill f - FSR 4 INT8
https://www.youtube.com/watch?v=h2Lg7FiKcpE
#gaming #silenhillf #rdna3 #fsr4
https://www.youtube.com/watch?v=h2Lg7FiKcpE
#gaming #silenhillf #rdna3 #fsr4
Most Efficient, Least Precise, and Most Performant to Least Efficient, Most Precise, and Least Performant:
INT4
FP4
INT8
FP8
INT16
FP16
INT4
FP4
INT8
FP8
INT16
FP16
October 22, 2025 at 5:40 PM
Most Efficient, Least Precise, and Most Performant to Least Efficient, Most Precise, and Least Performant:
INT4
FP4
INT8
FP8
INT16
FP16
INT4
FP4
INT8
FP8
INT16
FP16
October 22, 2025 at 5:17 PM
Painkiller (2025) - FSR4 INT8 - Epic - RX 7800 XT - 1440p
https://www.youtube.com/watch?v=m_m7ZdiLgYM
https://www.youtube.com/watch?v=m_m7ZdiLgYM
October 22, 2025 at 11:10 AM
Painkiller (2025) - FSR4 INT8 - Epic - RX 7800 XT - 1440p
https://www.youtube.com/watch?v=m_m7ZdiLgYM
https://www.youtube.com/watch?v=m_m7ZdiLgYM
It’s kind of sketchy, unfortunately. I’ve been playing it on my SD OLED on Very Low settings with FSR set to Performance mode, but there are still frequent drops under 30fps.
I want to try the FSR 4 INT8 plugin since that helped image quality significantly on Clair Obscur, another heavy UE5 tile.
I want to try the FSR 4 INT8 plugin since that helped image quality significantly on Clair Obscur, another heavy UE5 tile.
October 19, 2025 at 3:48 PM
It’s kind of sketchy, unfortunately. I’ve been playing it on my SD OLED on Very Low settings with FSR set to Performance mode, but there are still frequent drops under 30fps.
I want to try the FSR 4 INT8 plugin since that helped image quality significantly on Clair Obscur, another heavy UE5 tile.
I want to try the FSR 4 INT8 plugin since that helped image quality significantly on Clair Obscur, another heavy UE5 tile.
It's good that Microsoft is doing this now. They can see at scale how different applications perform on different platforms with real-world usage. This is also similar to how genAI required a 30+ TOPS (Int8) NPU on a smartphone to run genAI applications, but Google did it with ~4 TOPS.
October 17, 2025 at 10:17 PM
It's good that Microsoft is doing this now. They can see at scale how different applications perform on different platforms with real-world usage. This is also similar to how genAI required a 30+ TOPS (Int8) NPU on a smartphone to run genAI applications, but Google did it with ~4 TOPS.
from the datacenter GPU PoV closest equivalent to the DGX Spark's GPU is the NVIDIA L4
~ same memory bandwidth, FP32 tflops and FP16/BF16/FP8/INT8 tensor core throughput (but DGX Spark has mxfp4 at twice that rate)
But you get 128GB of memory capacity, and a *very strong* CPU
~ same memory bandwidth, FP32 tflops and FP16/BF16/FP8/INT8 tensor core throughput (but DGX Spark has mxfp4 at twice that rate)
But you get 128GB of memory capacity, and a *very strong* CPU
October 16, 2025 at 10:19 PM
from the datacenter GPU PoV closest equivalent to the DGX Spark's GPU is the NVIDIA L4
~ same memory bandwidth, FP32 tflops and FP16/BF16/FP8/INT8 tensor core throughput (but DGX Spark has mxfp4 at twice that rate)
But you get 128GB of memory capacity, and a *very strong* CPU
~ same memory bandwidth, FP32 tflops and FP16/BF16/FP8/INT8 tensor core throughput (but DGX Spark has mxfp4 at twice that rate)
But you get 128GB of memory capacity, and a *very strong* CPU
CB-Funk-Podcast #140: Hallo FSR 4 mit INT8 (?), M5 und BF6, Tschüss Xbox und Win 10
CB-Funk-Podcast #140: Hallo FSR 4 mit INT8 (?), M5 und BF6, Tschüss Xbox und Win 10
Im Podcast geht es diese Woche um Benchmarks zu Battlefield 6, Zahlenformat-Wirrwar bei FSR 4, die neueste Xbox und das Aus für Windows 10.
www.computerbase.de
October 16, 2025 at 5:00 PM
CB-Funk-Podcast #140: Hallo FSR 4 mit INT8 (?), M5 und BF6, Tschüss Xbox und Win 10
[Wccftech]AMD FSR 4 INT8 Xbox Series X Simulation Shows Possible Advantage in Image Quality and Frame Rate - Wccftech Digital Foundry ran a simulation of Xbox Series X-like hardware running the leaked AMD FSR 4 INT8 version, and the result is intriguing.
October 16, 2025 at 11:30 AM
[Wccftech]AMD FSR 4 INT8 Xbox Series X Simulation Shows Possible Advantage in Image Quality and Frame Rate - Wccftech Digital Foundry ran a simulation of Xbox Series X-like hardware running the leaked AMD FSR 4 INT8 version, and the result is intriguing.
Okay so on the vector math thing I am working on a sheet with Copilot to calculate the memory size needed... It produced this banger:
SWITCH(B6,"Float16",2,"Int8",1,"Binary (1-bit)",0)
Sure, technically a binary is 0 bytes but we need to determine how much memory is consumed with a gazillion bits
SWITCH(B6,"Float16",2,"Int8",1,"Binary (1-bit)",0)
Sure, technically a binary is 0 bytes but we need to determine how much memory is consumed with a gazillion bits
October 15, 2025 at 7:40 PM
Okay so on the vector math thing I am working on a sheet with Copilot to calculate the memory size needed... It produced this banger:
SWITCH(B6,"Float16",2,"Int8",1,"Binary (1-bit)",0)
Sure, technically a binary is 0 bytes but we need to determine how much memory is consumed with a gazillion bits
SWITCH(B6,"Float16",2,"Int8",1,"Binary (1-bit)",0)
Sure, technically a binary is 0 bytes but we need to determine how much memory is consumed with a gazillion bits
Digital Foundry ran a simulation of Xbox Series X-like hardware running the leaked AMD FSR 4 INT8 version, and the result is intriguing.
AMD FSR 4 INT8 Xbox Series X Simulation Shows Possible Advantage in Image Quality and Frame Rate
wccftech.com
October 15, 2025 at 12:00 PM
Digital Foundry ran a simulation of Xbox Series X-like hardware running the leaked AMD FSR 4 INT8 version, and the result is intriguing.