this is (effectively) an int8 model with int8 activations, packed 2 weights/2 activations into the same word, and the measured quantization error relative to bf16 is only 2.3x.
August 17, 2026 at 5:30 PM
this is (effectively) an int8 model with int8 activations, packed 2 weights/2 activations into the same word, and the measured quantization error relative to bf16 is only 2.3x.
DF Direct Weekly #232 is now on early access to backers of our Supporter Program, talking Dying Light: The Beast, the big Intel/Nvidia deal, FSR4 INT8 running on RDNA 3, the continuing Borderlands 4 controversy and much more: patreon.com/digitalfoundry
September 20, 2025 at 9:13 PM
DF Direct Weekly #232 is now on early access to backers of our Supporter Program, talking Dying Light: The Beast, the big Intel/Nvidia deal, FSR4 INT8 running on RDNA 3, the continuing Borderlands 4 controversy and much more: patreon.com/digitalfoundry
it says f8 so i assumed fp8 but it's int8 quant bruh
March 3, 2026 at 7:59 AM
it says f8 so i assumed fp8 but it's int8 quant bruh
AMD's leaked FSR4 INT8 in depth: image quality vs FP4, RDNA 3, RDNA 2, Steam Deck and ROG Ally tested, ballpark Xbox Series X simulations and... FSR4 on Nvidia - it works but how fast is it vs DLSS? youtu.be/fqDWJCW06Ks
October 14, 2025 at 2:31 PM
AMD's leaked FSR4 INT8 in depth: image quality vs FP4, RDNA 3, RDNA 2, Steam Deck and ROG Ally tested, ballpark Xbox Series X simulations and... FSR4 on Nvidia - it works but how fast is it vs DLSS? youtu.be/fqDWJCW06Ks
July 9, 2026 at 12:17 PM
Bon alors, c'est aujourd'hui qu'on a enfin confirmation que la Switch 2 a un chipset Cortex A78C + Ampere à 1536 cœurs CUDA en Samsung 5nm, augmenté de 48 tensor cores capables de 50 TOPS en INT8 ????????????
April 2, 2025 at 6:51 AM
Bon alors, c'est aujourd'hui qu'on a enfin confirmation que la Switch 2 a un chipset Cortex A78C + Ampere à 1536 cœurs CUDA en Samsung 5nm, augmenté de 48 tensor cores capables de 50 TOPS en INT8 ????????????
We've tested the in-progress INT8 version of FSR 4, designed to run on RDNA3, up against full FSR4 on RDNA4, FSR3 and more youtu.be/yB0qmTCzrmI
AMD Tried To Hide This From You - FSR 4 on RDNA 3 & 2 Tested
YouTube video by Hardware Unboxed
youtu.be
October 2, 2025 at 11:51 AM
We've tested the in-progress INT8 version of FSR 4, designed to run on RDNA3, up against full FSR4 on RDNA4, FSR3 and more youtu.be/yB0qmTCzrmI
Introducing 🥚EGGROLL 🥚(Evolution Guided General Optimization via Low-rank Learning)! 🚀 Scaling backprop-free Evolution Strategies (ES) for billion-parameter models at large population sizes
⚡100x Training Throughput
🎯Fast Convergence
🔢Pure Int8 Pretraining of RNN LLMs
⚡100x Training Throughput
🎯Fast Convergence
🔢Pure Int8 Pretraining of RNN LLMs
November 21, 2025 at 5:56 PM
Introducing 🥚EGGROLL 🥚(Evolution Guided General Optimization via Low-rank Learning)! 🚀 Scaling backprop-free Evolution Strategies (ES) for billion-parameter models at large population sizes
⚡100x Training Throughput
🎯Fast Convergence
🔢Pure Int8 Pretraining of RNN LLMs
⚡100x Training Throughput
🎯Fast Convergence
🔢Pure Int8 Pretraining of RNN LLMs
INT8 LLMs are the future
August 22, 2026 at 10:25 PM
INT8 LLMs are the future
目隠れデザめちゃ好きなのに探索者で作ったのはじめてや
今日KPCでいくINT8の男
今日KPCでいくINT8の男
May 30, 2024 at 4:48 AM
目隠れデザめちゃ好きなのに探索者で作ったのはじめてや
今日KPCでいくINT8の男
今日KPCでいくINT8の男
Leaked FSR4 INT8 version runs on RDNA2 and 3 with 9–13% lower performance, image quality below FSR4 FP8 but still above FSR 3.1
Leaked FSR4 INT8 version runs on RDNA2 and 3 with 9–13% lower performance, image quality below FSR4 FP8 but still above FSR 3.1 - VideoCardz.com
ComputerBase tests leaked FSR 4 INT8 version: works on RDNA 2 & 3, image quality falls short of FP8 but it is still better than FSR 3.1 ComputerBase conducted a deeper analysis of the leaked FSR 4 INT8...
videocardz.com
October 13, 2025 at 10:02 AM
Leaked FSR4 INT8 version runs on RDNA2 and 3 with 9–13% lower performance, image quality below FSR4 FP8 but still above FSR 3.1
AMD "Zen 6" ISA to Bring AVX512 FP16, VNNI INT8, and More to Desktop PCs
www.techpowerup.com
November 10, 2025 at 9:46 AM
シルバーウイークにMinimax H3高速化にチャレンジ🕺
Sol-AttnやらINT8 ConvRot VAEやら3StepLoRAやら入れて
960 × 544p
1秒1分から1秒20秒に((( ；ﾟДﾟ)))
３べぇ界王拳だ❕
これならガチャが許容範囲かな?🤔
きゃぴ💕
Sol-AttnやらINT8 ConvRot VAEやら3StepLoRAやら入れて
960 × 544p
1秒1分から1秒20秒に((( ；ﾟДﾟ)))
３べぇ界王拳だ❕
これならガチャが許容範囲かな?🤔
きゃぴ💕
September 25, 2026 at 6:46 AM
シルバーウイークにMinimax H3高速化にチャレンジ🕺
Sol-AttnやらINT8 ConvRot VAEやら3StepLoRAやら入れて
960 × 544p
1秒1分から1秒20秒に((( ；ﾟДﾟ)))
３べぇ界王拳だ❕
これならガチャが許容範囲かな?🤔
きゃぴ💕
Sol-AttnやらINT8 ConvRot VAEやら3StepLoRAやら入れて
960 × 544p
1秒1分から1秒20秒に((( ；ﾟДﾟ)))
３べぇ界王拳だ❕
これならガチャが許容範囲かな?🤔
きゃぴ💕
FlashRL, the first open–source & working RL recipe that applies INT8/FP8 for rollout without losing performance compared to BF16!
Blog: fengyao.notion.site/flash-rl
Repo: github.com/yaof20/Flash...
Blog: fengyao.notion.site/flash-rl
Repo: github.com/yaof20/Flash...
August 12, 2025 at 1:44 PM
FlashRL, the first open–source & working RL recipe that applies INT8/FP8 for rollout without losing performance compared to BF16!
Blog: fengyao.notion.site/flash-rl
Repo: github.com/yaof20/Flash...
Blog: fengyao.notion.site/flash-rl
Repo: github.com/yaof20/Flash...
Sony’s new PSSR shares FSR Upscaling roots, but uses an INT8 implementation videocardz.com/newz/sonys-n...
Sony’s new PSSR shares FSR Upscaling roots, but uses an INT8 implementation - VideoCardz.com
Mark Cerny says PS5 Pro’s new PSSR uses INT8 (like the leaked FSR4 for RDNA3/2) Mark Cerny has now given the clearest description yet of how Sony’s
videocardz.com
March 24, 2026 at 9:33 PM
Sony’s new PSSR shares FSR Upscaling roots, but uses an INT8 implementation videocardz.com/newz/sonys-n...
OptiScaler update brings FSR 4 INT8 support to RDNA 2 on newer Radeon drivers
OptiScaler update brings FSR 4 INT8 support to RDNA 2 on newer Radeon drivers - VideoCardz.com
Community FSR 4 INT8 build for RDNA 2 claims fix for ghosting on latest drivers AMD still limits official FSR 4 support to its newer Radeon RX 9000
videocardz.com
March 19, 2026 at 12:33 PM
OptiScaler update brings FSR 4 INT8 support to RDNA 2 on newer Radeon drivers
🏎️ You can perform 200ms search over 40 million texts using just a CPU server, 8GB of RAM, and 45GB of disk space.
The trick: Binary search with int8 rescoring.
I'll show you a demo & how it works in the 🧵:
The trick: Binary search with int8 rescoring.
I'll show you a demo & how it works in the 🧵:
January 6, 2026 at 7:56 PM
🏎️ You can perform 200ms search over 40 million texts using just a CPU server, 8GB of RAM, and 45GB of disk space.
The trick: Binary search with int8 rescoring.
I'll show you a demo & how it works in the 🧵:
The trick: Binary search with int8 rescoring.
I'll show you a demo & how it works in the 🧵:
In Software:
int8 + int8 = ?
int8 * int8 = ?
int8 + int8 = ?
int8 * int8 = ?
August 12, 2023 at 6:20 PM
In Software:
int8 + int8 = ?
int8 * int8 = ?
int8 + int8 = ?
int8 * int8 = ?
🤗 Perplexity has released 4 open-weights state-of-the-art multilingual embedding models designed for retrieval tasks!
pplx-embed-v1 and pplx-embed-context-v1
Specifically trained for int8 and binary embeddings, they'll be viable for massive search problems.
Details in 🧵
pplx-embed-v1 and pplx-embed-context-v1
Specifically trained for int8 and binary embeddings, they'll be viable for massive search problems.
Details in 🧵
February 27, 2026 at 2:36 PM
🤗 Perplexity has released 4 open-weights state-of-the-art multilingual embedding models designed for retrieval tasks!
pplx-embed-v1 and pplx-embed-context-v1
Specifically trained for int8 and binary embeddings, they'll be viable for massive search problems.
Details in 🧵
pplx-embed-v1 and pplx-embed-context-v1
Specifically trained for int8 and binary embeddings, they'll be viable for massive search problems.
Details in 🧵
How Do INT8 Edge Inference Systems Work for Real-Time AI in 2026? — https://skymineral.com/knowledge/how_do_int8_edge_inference_systems_work_for_real-time_ai_in_2026.php
September 25, 2026 at 5:46 AM
How Do INT8 Edge Inference Systems Work for Real-Time AI in 2026? — https://skymineral.com/knowledge/how_do_int8_edge_inference_systems_work_for_real-time_ai_in_2026.php
Wow super interesting. Quantizing might not be as big of a cost saver as I thought.
"... reducing from float32 (32 bits) to float16 (16 bits) or int8 (8 bits) should yield ideal 2× or 4× gains, respectively. However, such improvements are not observed in practice."
"... reducing from float32 (32 bits) to float16 (16 bits) or int8 (8 bits) should yield ideal 2× or 4× gains, respectively. However, such improvements are not observed in practice."
"numerical precision reduction yields the most benefit in the prefill phase of large models, where compute dominates. In contrast, the decode phase remains memory-limited, and aggressive quantization (e.g., int8 or int4) may incur overheads that outweigh theoretical savings"
Understanding Efficiency: Quantization, Batching, and Serving Strategies in LLM Energy Use
Large Language Models (LLMs) are increasingly deployed in production, contributing towards shifting the burden in terms of computational resources and energy demands from training to inference. While ...
arxiv.org
February 2, 2026 at 6:23 PM
Wow super interesting. Quantizing might not be as big of a cost saver as I thought.
"... reducing from float32 (32 bits) to float16 (16 bits) or int8 (8 bits) should yield ideal 2× or 4× gains, respectively. However, such improvements are not observed in practice."
"... reducing from float32 (32 bits) to float16 (16 bits) or int8 (8 bits) should yield ideal 2× or 4× gains, respectively. However, such improvements are not observed in practice."
you're probably running something on the scale of INT8-quantized DeepSeek, which has a power draw similar to the game itself (i.e., you've got your whole graphics card pegged at 100% running it.)
July 17, 2025 at 3:28 AM
you're probably running something on the scale of INT8-quantized DeepSeek, which has a power draw similar to the game itself (i.e., you've got your whole graphics card pegged at 100% running it.)
Mistral AI Introduces Codestral Embed, the new state-of-the-art embedding model for code.
Codestral Embed can output embeddings with different dimensions and precisions. Codestral Embed with dimension 256 and int8 precision still performs better than any model from our competitors.
Codestral Embed can output embeddings with different dimensions and precisions. Codestral Embed with dimension 256 and int8 precision still performs better than any model from our competitors.
May 29, 2025 at 11:52 PM
Mistral AI Introduces Codestral Embed, the new state-of-the-art embedding model for code.
Codestral Embed can output embeddings with different dimensions and precisions. Codestral Embed with dimension 256 and int8 precision still performs better than any model from our competitors.
Codestral Embed can output embeddings with different dimensions and precisions. Codestral Embed with dimension 256 and int8 precision still performs better than any model from our competitors.