Explanation of INT8 ConvRot (FP8 is no longer needed)

3 hours ago

INT8 ConvRot is a new quantization method natively supported in ComfyUI v0.27.0 (released July 1, 2026).
It is particularly beneficial for GeForce RTX 20/30 series but also outperforms FP8 and FP8 Scaled on RTX 40/50 series.
INT8 ConvRot uses INT8 encoding with a ConvRot (Convolution + Rotation) technique to distribute outliers, solving INT8's accuracy issues.
It is faster than FP8 Scaled even on RTX 5090 (35% reduction in warm start time) and is expected to become the standard 8-bit format.
ComfyUI natively supports it; Forge Neo added support in commit 6d0bc6e (2026/7/7).
Hardware support for INT8 exists on all RTX 20/30/40/50 series, unlike FP8 which is only supported on RTX 40/50.
Performance reports indicate INT8 ConvRot > NVFP4 in speed, with accuracy approaching GGUF Q8.
Conversion tools and custom nodes (ComfyUI-INT8-Fast) exist, but native nodes are recommended after v0.27.0.
The article includes a detailed classification table of model formats (encoding, scaling, quantization methods).
Author measured Krea2 generation speed: INT8 ConvRot 2.56s warm start vs FP8 Scaled 3.91s on RTX 5090.

Go 1.27 introduces type parameters on methods, allowing generic operations on types to be defined as methods instead of package-level functions.
Struct literal keys can now use promoted fields from embedded structs directly, simplifying initialization syntax.
Function type inference is generalized to work in conversions, composite literals, and other contexts, reducing the need for explicit type arguments.

NVIDIA introduces Nemotron 3 Ultra, their most capable model with 550 billion total parameters and 55 billion active parameters.
Employs Mixture-of-Experts Hybrid Mamba-Attention architecture, LatentMoE, and MTP layers for speculative decoding and faster inference.
Supports inference time reasoning budget control and is pretrained in NVFP4 format.

BTL-3 is a 27B open-weight agent model designed for agentic coding, structural tool use, repository work, failure recovery, and long multi-turn execution.
It achieves high scores on benchmarks: 95.1% HumanEval pass@1, 88.5% BFCL v4 AST, and 88.1% LiveCodeBench v6.
A Compact edition (8.39 GB) packages the full model into a single file with 92.2% behavior retention, enabling native local inference.

The Nebula driver for Nvidia GPUs currently only supports Turing and Ampere architectures, with no blobs for Lovelace or newer cards like the RTX 5090.
A user reports that the GTX 2060 runs well at 2560x1440 resolution with refresh rates of 60/120/144 Hz, and that DisplayPort support is ready but not yet released.
DisplayPort offers advantages such as higher refresh rates and easier multi-monitor setups, while HDMI can be inconvenient for dual-booting and VR headsets.

G1 becomes the default garbage collector in all environments, having been improved to match or exceed the performance of Serial GC.
Post-quantum hybrid key exchange for TLS 1.3 is introduced to enhance security against future quantum computing threats.
Compact object headers are enabled by default, reducing memory usage and improving performance.

Go 1.27 introduces type parameters on methods, allowing generic operations on types to be defined as methods instead of package-level functions.
Struct literal keys can now use promoted fields from embedded structs directly, simplifying initialization syntax.
Function type inference is generalized to work in conversions, composite literals, and other contexts, reducing the need for explicit type arguments.

NVIDIA introduces Nemotron 3 Ultra, their most capable model with 550 billion total parameters and 55 billion active parameters.
Employs Mixture-of-Experts Hybrid Mamba-Attention architecture, LatentMoE, and MTP layers for speculative decoding and faster inference.
Supports inference time reasoning budget control and is pretrained in NVFP4 format.

BTL-3 is a 27B open-weight agent model designed for agentic coding, structural tool use, repository work, failure recovery, and long multi-turn execution.
It achieves high scores on benchmarks: 95.1% HumanEval pass@1, 88.5% BFCL v4 AST, and 88.1% LiveCodeBench v6.
A Compact edition (8.39 GB) packages the full model into a single file with 92.2% behavior retention, enabling native local inference.

The Nebula driver for Nvidia GPUs currently only supports Turing and Ampere architectures, with no blobs for Lovelace or newer cards like the RTX 5090.
A user reports that the GTX 2060 runs well at 2560x1440 resolution with refresh rates of 60/120/144 Hz, and that DisplayPort support is ready but not yet released.
DisplayPort offers advantages such as higher refresh rates and easier multi-monitor setups, while HDMI can be inconvenient for dual-booting and VR headsets.

G1 becomes the default garbage collector in all environments, having been improved to match or exceed the performance of Serial GC.
Post-quantum hybrid key exchange for TLS 1.3 is introduced to enhance security against future quantum computing threats.
Compact object headers are enabled by default, reducing memory usage and improving performance.

Hasty Briefsbeta