ISee
Member
There is an entry in the TimeSpy Database for a new GPU. The unknown AMD graphics card is supposed to be Vega because it is running at a low memory clock of 700 MHz, which would be more typical for HBM memory instead of GDDR5. But 700 MHz are also a bit too low for HBM2, which is expected to run between 800-1000 MHz. Still, this might be an early Vega 10 prototype.
The database also mentions a core clock of 1,200 MHz and a memory configuration of 8 GiByte.
The unknown card reaches 5721 points in the TimeSpy benchmark and is faster than the R9 Fury X (~5200 points, currently the fastest AMD GPU). For comparison: A Geforce GTX 1070 reaches around 5,600 points and a GTX 1080 around 6,800 points (no overclocks).
Please keep in mind: It's normal for prototype like hardware to run at lower speeds. The current drivers are also probably in an early state and final performance will be better.
3D Mark TimeSpy Score
update:
There seems to be a new VEGA patch for linux, which gives us a bit more insight into VEGA specs.
[PATCH 048/100] drm/amdgpu: implement GFX 9.0 support
When comparing raw numbers Vega seems to be at the same level as Fiji, but it is produced at 14nm instead of 28nm and the memory interface is faster. (Still 4096 shader units, 2048-bit wide memory interface).
Overall the differences aren't mind-blowing, on paper. But Vega should at least reach much higher clock speeds then Fiji and there are more architectural benefits, for sure.
update 2:
A new entry in the CompuBench database was spotted. The entry bears the name gfx 9000 (polaris is gfx 8) and has 64 Compute Units. This is also the first AMD GPU reaching 1600 MHz. We don't know if this is a consumer or a professional Vega variant.
The card still has 64 Compute Units. If AMD kept the same principles for GFX9 architecture (64 cores per cluster), then we should expect 4096 shaders. Making this a 13.1 tflop card. Lately, AMD needed significantly more raw power to get to Nvidia performance levels.
The database also mentions a core clock of 1,200 MHz and a memory configuration of 8 GiByte.
The unknown card reaches 5721 points in the TimeSpy benchmark and is faster than the R9 Fury X (~5200 points, currently the fastest AMD GPU). For comparison: A Geforce GTX 1070 reaches around 5,600 points and a GTX 1080 around 6,800 points (no overclocks).
Please keep in mind: It's normal for prototype like hardware to run at lower speeds. The current drivers are also probably in an early state and final performance will be better.
3D Mark TimeSpy Score
update:
There seems to be a new VEGA patch for linux, which gives us a bit more insight into VEGA specs.
[PATCH 048/100] drm/amdgpu: implement GFX 9.0 support
When comparing raw numbers Vega seems to be at the same level as Fiji, but it is produced at 14nm instead of 28nm and the memory interface is faster. (Still 4096 shader units, 2048-bit wide memory interface).
Overall the differences aren't mind-blowing, on paper. But Vega should at least reach much higher clock speeds then Fiji and there are more architectural benefits, for sure.
Code:
switch (adev->asic_type) {
+ case CHIP_VEGA10:
+ adev->gfx.config.max_shader_engines = 4;
+ adev->gfx.config.max_tile_pipes = 8; //??
+ adev->gfx.config.max_cu_per_sh = 16;
+ adev->gfx.config.max_sh_per_se = 1;
+ adev->gfx.config.max_backends_per_se = 4;
+ adev->gfx.config.max_texture_channel_caches = 16;
+ adev->gfx.config.max_gprs = 256;
+ adev->gfx.config.max_gs_threads = 32;
+ adev->gfx.config.max_hw_contexts = 8;
Old Fiji for context:
Code:case CHIP_FIJI: adev->gfx.config.max_shader_engines = 4; adev->gfx.config.max_tile_pipes = 16; adev->gfx.config.max_cu_per_sh = 16; adev->gfx.config.max_sh_per_se = 1; adev->gfx.config.max_backends_per_se = 4; adev->gfx.config.max_texture_channel_caches = 16; adev->gfx.config.max_gprs = 256; adev->gfx.config.max_gs_threads = 32; adev->gfx.config.max_hw_contexts = 8;
Also Videocardz's response regarding the myriad of Vega 'leaks':
https://videocardz.com/69040/amd-ra...p-edition-packaging-leaked#comment-3282921662
update 2:
A new entry in the CompuBench database was spotted. The entry bears the name gfx 9000 (polaris is gfx 8) and has 64 Compute Units. This is also the first AMD GPU reaching 1600 MHz. We don't know if this is a consumer or a professional Vega variant.
The card still has 64 Compute Units. If AMD kept the same principles for GFX9 architecture (64 cores per cluster), then we should expect 4096 shaders. Making this a 13.1 tflop card. Lately, AMD needed significantly more raw power to get to Nvidia performance levels.
Vega sighted:
https://compubench.com/device.jsp?benchmark=compu20d&os=Windows&api=cl&D=AMD+6864:00&testgroup=info
16GB and 1.600 Mhz boost clock
https://www.computerbase.de/2017-05/amd-vega-10-16-gb/