Add GTT memory monitoring for AMD GPUs by jibanez-staticduo · Pull Request #515 · corecoding/Vitals

jibanez-staticduo · 2026-02-06T11:01:24Z

Summary

Adds GTT (Graphics Translation Table) memory monitoring for AMD GPUs detected via DRM sysfs
Reads mem_info_gtt_used and mem_info_gtt_total from /sys/class/drm/cardN/device/ and exposes them as two new GPU sensor fields: GTT Used and GTT Total
Errors are silently caught so discrete AMD GPUs or older drivers that don't expose these files are unaffected

Motivation

GTT memory represents system RAM accessible by the GPU. On AMD APUs — such as Strix Halo (Ryzen AI MAX), Rembrandt, and Phoenix — the dedicated VRAM is typically very small (e.g. 512 MiB), while GTT can be tens to hundreds of GiBs of system memory allocated to the GPU.

For workloads like LLM inference (via ROCm/llama.cpp), the GPU primarily uses GTT memory rather than VRAM. Tools like amdgpu_top already report GTT usage, but Vitals currently only shows VRAM — which gives an incomplete picture on these APUs.

Example on an AMD Ryzen AI MAX+ 395 (Strix Halo):

Memory Type	Total	Typical Usage
VRAM (dedicated)	512 MiB	~455 MiB
GTT (system RAM)	124 GiB	~63 GiB (running LLMs)

Without GTT monitoring, Vitals only shows the 512 MiB VRAM, missing the 63 GiB actually being used by the GPU.

Changes

sensors.js: Added two new FileModule.File reads inside the existing AMD DRM block in _readGpuDrm(), reading mem_info_gtt_used and mem_info_gtt_total. Both are wrapped in .catch() blocks to gracefully handle systems where these files don't exist.

Testing

Tested on Fedora 43 with AMD Ryzen AI MAX+ 395 (Strix Halo) / Radeon 8060S
GTT values correctly displayed alongside existing VRAM values
Verified that the extension loads without errors on systems where GTT sysfs files are present
No regressions observed for existing GPU, temperature, or other sensor readings

Made with Cursor

AMD GPUs expose GTT (Graphics Translation Table) memory usage through sysfs at /sys/class/drm/cardN/device/mem_info_gtt_{used,total}. GTT represents system RAM that is accessible by the GPU, which is particularly important on AMD APUs (e.g. Strix Halo, Rembrandt, Phoenix) where GTT memory can be significantly larger than dedicated VRAM and is the primary memory pool for GPU workloads such as LLM inference or graphics rendering. This patch adds two new GPU sensor fields for AMD cards when detected via DRM: - GTT Used: current GTT memory consumption - GTT Total: total GTT memory available The reads are wrapped in catch blocks so that AMD discrete GPUs or older drivers that do not expose these sysfs files will silently ignore the missing data without affecting existing functionality. Co-authored-by: Cursor <cursoragent@cursor.com>

jibanez-staticduo · 2026-02-06T11:01:24Z

Summary

Adds GTT (Graphics Translation Table) memory monitoring for AMD GPUs detected via DRM sysfs
Reads mem_info_gtt_used and mem_info_gtt_total from /sys/class/drm/cardN/device/ and exposes them as two new GPU sensor fields: GTT Used and GTT Total
Errors are silently caught so discrete AMD GPUs or older drivers that don't expose these files are unaffected

Motivation

GTT memory represents system RAM accessible by the GPU. On AMD APUs — such as Strix Halo (Ryzen AI MAX), Rembrandt, and Phoenix — the dedicated VRAM is typically very small (e.g. 512 MiB), while GTT can be tens to hundreds of GiBs of system memory allocated to the GPU.

For workloads like LLM inference (via ROCm/llama.cpp), the GPU primarily uses GTT memory rather than VRAM. Tools like amdgpu_top already report GTT usage, but Vitals currently only shows VRAM — which gives an incomplete picture on these APUs.

Example on an AMD Ryzen AI MAX+ 395 (Strix Halo):

Memory Type	Total	Typical Usage
VRAM (dedicated)	512 MiB	~455 MiB
GTT (system RAM)	124 GiB	~63 GiB (running LLMs)

Without GTT monitoring, Vitals only shows the 512 MiB VRAM, missing the 63 GiB actually being used by the GPU.

Changes

sensors.js: Added two new FileModule.File reads inside the existing AMD DRM block in _readGpuDrm(), reading mem_info_gtt_used and mem_info_gtt_total. Both are wrapped in .catch() blocks to gracefully handle systems where these files don't exist.

Testing

Tested on Fedora 43 with AMD Ryzen AI MAX+ 395 (Strix Halo) / Radeon 8060S
GTT values correctly displayed alongside existing VRAM values
Verified that the extension loads without errors on systems where GTT sysfs files are present
No regressions observed for existing GPU, temperature, or other sensor readings

Made with Cursor

AMD GPUs expose GTT (Graphics Translation Table) memory usage through sysfs at /sys/class/drm/cardN/device/mem_info_gtt_{used,total}. GTT represents system RAM that is accessible by the GPU, which is particularly important on AMD APUs (e.g. Strix Halo, Rembrandt, Phoenix) where GTT memory can be significantly larger than dedicated VRAM and is the primary memory pool for GPU workloads such as LLM inference or graphics rendering. This patch adds two new GPU sensor fields for AMD cards when detected via DRM: - GTT Used: current GTT memory consumption - GTT Total: total GTT memory available The reads are wrapped in catch blocks so that AMD discrete GPUs or older drivers that do not expose these sysfs files will silently ignore the missing data without affecting existing functionality. Co-authored-by: Cursor <cursoragent@cursor.com>

jibanez-staticduo · 2026-02-06T11:01:24Z

Summary

Adds GTT (Graphics Translation Table) memory monitoring for AMD GPUs detected via DRM sysfs
Reads mem_info_gtt_used and mem_info_gtt_total from /sys/class/drm/cardN/device/ and exposes them as two new GPU sensor fields: GTT Used and GTT Total
Errors are silently caught so discrete AMD GPUs or older drivers that don't expose these files are unaffected

Motivation

GTT memory represents system RAM accessible by the GPU. On AMD APUs — such as Strix Halo (Ryzen AI MAX), Rembrandt, and Phoenix — the dedicated VRAM is typically very small (e.g. 512 MiB), while GTT can be tens to hundreds of GiBs of system memory allocated to the GPU.

For workloads like LLM inference (via ROCm/llama.cpp), the GPU primarily uses GTT memory rather than VRAM. Tools like amdgpu_top already report GTT usage, but Vitals currently only shows VRAM — which gives an incomplete picture on these APUs.

Example on an AMD Ryzen AI MAX+ 395 (Strix Halo):

Memory Type	Total	Typical Usage
VRAM (dedicated)	512 MiB	~455 MiB
GTT (system RAM)	124 GiB	~63 GiB (running LLMs)

Without GTT monitoring, Vitals only shows the 512 MiB VRAM, missing the 63 GiB actually being used by the GPU.

Changes

sensors.js: Added two new FileModule.File reads inside the existing AMD DRM block in _readGpuDrm(), reading mem_info_gtt_used and mem_info_gtt_total. Both are wrapped in .catch() blocks to gracefully handle systems where these files don't exist.

Testing

Tested on Fedora 43 with AMD Ryzen AI MAX+ 395 (Strix Halo) / Radeon 8060S
GTT values correctly displayed alongside existing VRAM values
Verified that the extension loads without errors on systems where GTT sysfs files are present
No regressions observed for existing GPU, temperature, or other sensor readings

Made with Cursor

AMD GPUs expose GTT (Graphics Translation Table) memory usage through sysfs at /sys/class/drm/cardN/device/mem_info_gtt_{used,total}. GTT represents system RAM that is accessible by the GPU, which is particularly important on AMD APUs (e.g. Strix Halo, Rembrandt, Phoenix) where GTT memory can be significantly larger than dedicated VRAM and is the primary memory pool for GPU workloads such as LLM inference or graphics rendering. This patch adds two new GPU sensor fields for AMD cards when detected via DRM: - GTT Used: current GTT memory consumption - GTT Total: total GTT memory available The reads are wrapped in catch blocks so that AMD discrete GPUs or older drivers that do not expose these sysfs files will silently ignore the missing data without affecting existing functionality. Co-authored-by: Cursor <cursoragent@cursor.com>

jibanez-staticduo · 2026-02-06T11:01:24Z

Summary

Adds GTT (Graphics Translation Table) memory monitoring for AMD GPUs detected via DRM sysfs
Reads mem_info_gtt_used and mem_info_gtt_total from /sys/class/drm/cardN/device/ and exposes them as two new GPU sensor fields: GTT Used and GTT Total
Errors are silently caught so discrete AMD GPUs or older drivers that don't expose these files are unaffected

Motivation

GTT memory represents system RAM accessible by the GPU. On AMD APUs — such as Strix Halo (Ryzen AI MAX), Rembrandt, and Phoenix — the dedicated VRAM is typically very small (e.g. 512 MiB), while GTT can be tens to hundreds of GiBs of system memory allocated to the GPU.

For workloads like LLM inference (via ROCm/llama.cpp), the GPU primarily uses GTT memory rather than VRAM. Tools like amdgpu_top already report GTT usage, but Vitals currently only shows VRAM — which gives an incomplete picture on these APUs.

Example on an AMD Ryzen AI MAX+ 395 (Strix Halo):

Memory Type	Total	Typical Usage
VRAM (dedicated)	512 MiB	~455 MiB
GTT (system RAM)	124 GiB	~63 GiB (running LLMs)

Without GTT monitoring, Vitals only shows the 512 MiB VRAM, missing the 63 GiB actually being used by the GPU.

Changes

sensors.js: Added two new FileModule.File reads inside the existing AMD DRM block in _readGpuDrm(), reading mem_info_gtt_used and mem_info_gtt_total. Both are wrapped in .catch() blocks to gracefully handle systems where these files don't exist.

Testing

Tested on Fedora 43 with AMD Ryzen AI MAX+ 395 (Strix Halo) / Radeon 8060S
GTT values correctly displayed alongside existing VRAM values
Verified that the extension loads without errors on systems where GTT sysfs files are present
No regressions observed for existing GPU, temperature, or other sensor readings

Made with Cursor

AMD GPUs expose GTT (Graphics Translation Table) memory usage through sysfs at /sys/class/drm/cardN/device/mem_info_gtt_{used,total}. GTT represents system RAM that is accessible by the GPU, which is particularly important on AMD APUs (e.g. Strix Halo, Rembrandt, Phoenix) where GTT memory can be significantly larger than dedicated VRAM and is the primary memory pool for GPU workloads such as LLM inference or graphics rendering. This patch adds two new GPU sensor fields for AMD cards when detected via DRM: - GTT Used: current GTT memory consumption - GTT Total: total GTT memory available The reads are wrapped in catch blocks so that AMD discrete GPUs or older drivers that do not expose these sysfs files will silently ignore the missing data without affecting existing functionality. Co-authored-by: Cursor <cursoragent@cursor.com>

jibanez-staticduo · 2026-02-06T11:01:24Z

Summary

Adds GTT (Graphics Translation Table) memory monitoring for AMD GPUs detected via DRM sysfs
Reads mem_info_gtt_used and mem_info_gtt_total from /sys/class/drm/cardN/device/ and exposes them as two new GPU sensor fields: GTT Used and GTT Total
Errors are silently caught so discrete AMD GPUs or older drivers that don't expose these files are unaffected

Motivation

GTT memory represents system RAM accessible by the GPU. On AMD APUs — such as Strix Halo (Ryzen AI MAX), Rembrandt, and Phoenix — the dedicated VRAM is typically very small (e.g. 512 MiB), while GTT can be tens to hundreds of GiBs of system memory allocated to the GPU.

For workloads like LLM inference (via ROCm/llama.cpp), the GPU primarily uses GTT memory rather than VRAM. Tools like amdgpu_top already report GTT usage, but Vitals currently only shows VRAM — which gives an incomplete picture on these APUs.

Example on an AMD Ryzen AI MAX+ 395 (Strix Halo):

Memory Type	Total	Typical Usage
VRAM (dedicated)	512 MiB	~455 MiB
GTT (system RAM)	124 GiB	~63 GiB (running LLMs)

Without GTT monitoring, Vitals only shows the 512 MiB VRAM, missing the 63 GiB actually being used by the GPU.

Changes

sensors.js: Added two new FileModule.File reads inside the existing AMD DRM block in _readGpuDrm(), reading mem_info_gtt_used and mem_info_gtt_total. Both are wrapped in .catch() blocks to gracefully handle systems where these files don't exist.

Testing

Tested on Fedora 43 with AMD Ryzen AI MAX+ 395 (Strix Halo) / Radeon 8060S
GTT values correctly displayed alongside existing VRAM values
Verified that the extension loads without errors on systems where GTT sysfs files are present
No regressions observed for existing GPU, temperature, or other sensor readings

Made with Cursor

AMD GPUs expose GTT (Graphics Translation Table) memory usage through sysfs at /sys/class/drm/cardN/device/mem_info_gtt_{used,total}. GTT represents system RAM that is accessible by the GPU, which is particularly important on AMD APUs (e.g. Strix Halo, Rembrandt, Phoenix) where GTT memory can be significantly larger than dedicated VRAM and is the primary memory pool for GPU workloads such as LLM inference or graphics rendering. This patch adds two new GPU sensor fields for AMD cards when detected via DRM: - GTT Used: current GTT memory consumption - GTT Total: total GTT memory available The reads are wrapped in catch blocks so that AMD discrete GPUs or older drivers that do not expose these sysfs files will silently ignore the missing data without affecting existing functionality. Co-authored-by: Cursor <cursoragent@cursor.com>

jibanez-staticduo · 2026-02-06T11:01:24Z

Summary

Adds GTT (Graphics Translation Table) memory monitoring for AMD GPUs detected via DRM sysfs
Reads mem_info_gtt_used and mem_info_gtt_total from /sys/class/drm/cardN/device/ and exposes them as two new GPU sensor fields: GTT Used and GTT Total
Errors are silently caught so discrete AMD GPUs or older drivers that don't expose these files are unaffected

Motivation

GTT memory represents system RAM accessible by the GPU. On AMD APUs — such as Strix Halo (Ryzen AI MAX), Rembrandt, and Phoenix — the dedicated VRAM is typically very small (e.g. 512 MiB), while GTT can be tens to hundreds of GiBs of system memory allocated to the GPU.

For workloads like LLM inference (via ROCm/llama.cpp), the GPU primarily uses GTT memory rather than VRAM. Tools like amdgpu_top already report GTT usage, but Vitals currently only shows VRAM — which gives an incomplete picture on these APUs.

Example on an AMD Ryzen AI MAX+ 395 (Strix Halo):

Memory Type	Total	Typical Usage
VRAM (dedicated)	512 MiB	~455 MiB
GTT (system RAM)	124 GiB	~63 GiB (running LLMs)

Without GTT monitoring, Vitals only shows the 512 MiB VRAM, missing the 63 GiB actually being used by the GPU.

Changes

sensors.js: Added two new FileModule.File reads inside the existing AMD DRM block in _readGpuDrm(), reading mem_info_gtt_used and mem_info_gtt_total. Both are wrapped in .catch() blocks to gracefully handle systems where these files don't exist.

Testing

Tested on Fedora 43 with AMD Ryzen AI MAX+ 395 (Strix Halo) / Radeon 8060S
GTT values correctly displayed alongside existing VRAM values
Verified that the extension loads without errors on systems where GTT sysfs files are present
No regressions observed for existing GPU, temperature, or other sensor readings

Made with Cursor

AMD GPUs expose GTT (Graphics Translation Table) memory usage through sysfs at /sys/class/drm/cardN/device/mem_info_gtt_{used,total}. GTT represents system RAM that is accessible by the GPU, which is particularly important on AMD APUs (e.g. Strix Halo, Rembrandt, Phoenix) where GTT memory can be significantly larger than dedicated VRAM and is the primary memory pool for GPU workloads such as LLM inference or graphics rendering. This patch adds two new GPU sensor fields for AMD cards when detected via DRM: - GTT Used: current GTT memory consumption - GTT Total: total GTT memory available The reads are wrapped in catch blocks so that AMD discrete GPUs or older drivers that do not expose these sysfs files will silently ignore the missing data without affecting existing functionality. Co-authored-by: Cursor <cursoragent@cursor.com>

jibanez-staticduo · 2026-02-06T11:01:24Z

Summary

Adds GTT (Graphics Translation Table) memory monitoring for AMD GPUs detected via DRM sysfs
Reads mem_info_gtt_used and mem_info_gtt_total from /sys/class/drm/cardN/device/ and exposes them as two new GPU sensor fields: GTT Used and GTT Total
Errors are silently caught so discrete AMD GPUs or older drivers that don't expose these files are unaffected

Motivation

GTT memory represents system RAM accessible by the GPU. On AMD APUs — such as Strix Halo (Ryzen AI MAX), Rembrandt, and Phoenix — the dedicated VRAM is typically very small (e.g. 512 MiB), while GTT can be tens to hundreds of GiBs of system memory allocated to the GPU.

For workloads like LLM inference (via ROCm/llama.cpp), the GPU primarily uses GTT memory rather than VRAM. Tools like amdgpu_top already report GTT usage, but Vitals currently only shows VRAM — which gives an incomplete picture on these APUs.

Example on an AMD Ryzen AI MAX+ 395 (Strix Halo):

Memory Type	Total	Typical Usage
VRAM (dedicated)	512 MiB	~455 MiB
GTT (system RAM)	124 GiB	~63 GiB (running LLMs)

Without GTT monitoring, Vitals only shows the 512 MiB VRAM, missing the 63 GiB actually being used by the GPU.

Changes

sensors.js: Added two new FileModule.File reads inside the existing AMD DRM block in _readGpuDrm(), reading mem_info_gtt_used and mem_info_gtt_total. Both are wrapped in .catch() blocks to gracefully handle systems where these files don't exist.

Testing

Tested on Fedora 43 with AMD Ryzen AI MAX+ 395 (Strix Halo) / Radeon 8060S
GTT values correctly displayed alongside existing VRAM values
Verified that the extension loads without errors on systems where GTT sysfs files are present
No regressions observed for existing GPU, temperature, or other sensor readings

Made with Cursor

AMD GPUs expose GTT (Graphics Translation Table) memory usage through sysfs at /sys/class/drm/cardN/device/mem_info_gtt_{used,total}. GTT represents system RAM that is accessible by the GPU, which is particularly important on AMD APUs (e.g. Strix Halo, Rembrandt, Phoenix) where GTT memory can be significantly larger than dedicated VRAM and is the primary memory pool for GPU workloads such as LLM inference or graphics rendering. This patch adds two new GPU sensor fields for AMD cards when detected via DRM: - GTT Used: current GTT memory consumption - GTT Total: total GTT memory available The reads are wrapped in catch blocks so that AMD discrete GPUs or older drivers that do not expose these sysfs files will silently ignore the missing data without affecting existing functionality. Co-authored-by: Cursor <cursoragent@cursor.com>

jibanez-staticduo · 2026-02-06T11:01:24Z

Summary

Adds GTT (Graphics Translation Table) memory monitoring for AMD GPUs detected via DRM sysfs
Reads mem_info_gtt_used and mem_info_gtt_total from /sys/class/drm/cardN/device/ and exposes them as two new GPU sensor fields: GTT Used and GTT Total
Errors are silently caught so discrete AMD GPUs or older drivers that don't expose these files are unaffected

Motivation

GTT memory represents system RAM accessible by the GPU. On AMD APUs — such as Strix Halo (Ryzen AI MAX), Rembrandt, and Phoenix — the dedicated VRAM is typically very small (e.g. 512 MiB), while GTT can be tens to hundreds of GiBs of system memory allocated to the GPU.

For workloads like LLM inference (via ROCm/llama.cpp), the GPU primarily uses GTT memory rather than VRAM. Tools like amdgpu_top already report GTT usage, but Vitals currently only shows VRAM — which gives an incomplete picture on these APUs.

Example on an AMD Ryzen AI MAX+ 395 (Strix Halo):

Memory Type	Total	Typical Usage
VRAM (dedicated)	512 MiB	~455 MiB
GTT (system RAM)	124 GiB	~63 GiB (running LLMs)

Without GTT monitoring, Vitals only shows the 512 MiB VRAM, missing the 63 GiB actually being used by the GPU.

Changes

sensors.js: Added two new FileModule.File reads inside the existing AMD DRM block in _readGpuDrm(), reading mem_info_gtt_used and mem_info_gtt_total. Both are wrapped in .catch() blocks to gracefully handle systems where these files don't exist.

Testing

Tested on Fedora 43 with AMD Ryzen AI MAX+ 395 (Strix Halo) / Radeon 8060S
GTT values correctly displayed alongside existing VRAM values
Verified that the extension loads without errors on systems where GTT sysfs files are present
No regressions observed for existing GPU, temperature, or other sensor readings

Made with Cursor

AMD GPUs expose GTT (Graphics Translation Table) memory usage through sysfs at /sys/class/drm/cardN/device/mem_info_gtt_{used,total}. GTT represents system RAM that is accessible by the GPU, which is particularly important on AMD APUs (e.g. Strix Halo, Rembrandt, Phoenix) where GTT memory can be significantly larger than dedicated VRAM and is the primary memory pool for GPU workloads such as LLM inference or graphics rendering. This patch adds two new GPU sensor fields for AMD cards when detected via DRM: - GTT Used: current GTT memory consumption - GTT Total: total GTT memory available The reads are wrapped in catch blocks so that AMD discrete GPUs or older drivers that do not expose these sysfs files will silently ignore the missing data without affecting existing functionality. Co-authored-by: Cursor <cursoragent@cursor.com>

Uh oh!

Conversation

jibanez-staticduo commented Feb 6, 2026

Summary

Motivation

Example on an AMD Ryzen AI MAX+ 395 (Strix Halo):

Changes

Testing

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

jibanez-staticduo commented Feb 6, 2026

Summary

Motivation

Example on an AMD Ryzen AI MAX+ 395 (Strix Halo):

Changes

Testing

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

jibanez-staticduo commented Feb 6, 2026

Summary

Motivation

Example on an AMD Ryzen AI MAX+ 395 (Strix Halo):

Changes

Testing

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

jibanez-staticduo commented Feb 6, 2026

Summary

Motivation

Example on an AMD Ryzen AI MAX+ 395 (Strix Halo):

Changes

Testing

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

jibanez-staticduo commented Feb 6, 2026

Summary

Motivation

Example on an AMD Ryzen AI MAX+ 395 (Strix Halo):

Changes

Testing

Uh oh!

Reviewers

Assignees

Labels