3 ms·

Looks pretty good against Intel's new Xe Max: https://gfxbench.com/compare.jsp?benchmark=gfx50&did1=90754264&os1=OS+X&api1=metal&hwtype1=GPU&hwname1=Apple+M1&D

by omikun 6y ago

Looks pretty good against Intel's new Xe Max: https://gfxbench.com/compare.jsp?benchmark=gfx50&did1=90754264&os1=OS+X&api1=metal&hwtype1=GPU&hwname1=Apple+M1&D2=Intel%28R%29+Iris%28R%29+Xe+MAX+Graphics https://gfxbench.com/compare.jsp?benchmark=gfx50&did1=907542... It's not the 4-6x in raw graphics improvement advertised but at 10W vs Xe Max's 25W for just the GPU, the M1 getting 50% more fps, that's still 3.75x in perf/watt.

JAlexoid 6y ago
I wouldn't brag, when comparing Metal to OpenGL. Let alone - let's really not forget that this is an HBM type architecture. As a complete package it seems awesome, but we can argue for ages about the performance of GPU cores with no end result.
- fulafel 6y ago
  > this is an HBM type architecture You'd think so, but seems most people[1] think it's just LPDDR wired to the SoC using the standard protocol, inside the same package. (Though it might use > 1 channel I guess?) [1] eg in the spec table here https://www.anandtech.com/show/16235/apple-intros-first-three-apple-silicon-macs-late-2020-macbook-air-13inch-macbook-pro-mac-mini https://www.anandtech.com/show/16235/apple-intros-first-thre... - an interesing thinb in the mini spec table is also the 10x downgrade of the ethernet speed.
  - floatboth 6y ago
    LPDDR channels are weird. https://www.anandtech.com/show/16226/apple-silicon-m1-a14-deep-dive https://www.anandtech.com/show/16226/apple-silicon-m1-a14-de... shows "8x 16b LPDDR4X channels" Which would be the same width as dual channel DDR4 – 8x16 == 2x64 :) Also it's completely ridiculous that some people think it might be HBM. The presentation slide (shown in both anandtech articles) very obviously shows DDR style chips, with their own plastic packaging. That is not how HBM looks :) Also HBM would noticeably raise the price.

3 ms·

Looks pretty good against Intel's new Xe Max: https://gfxbench.com/compare.jsp?benchmark=gfx50&did1=90754264&os1=OS+X&api1=metal&hwtype1=GPU&hwname1=Apple+M1&D

by omikun 6y ago

JAlexoid 6y ago
I wouldn't brag, when comparing Metal to OpenGL. Let alone - let's really not forget that this is an HBM type architecture. As a complete package it seems awesome, but we can argue for ages about the performance of GPU cores with no end result.
- fulafel 6y ago
  > this is an HBM type architecture You'd think so, but seems most people[1] think it's just LPDDR wired to the SoC using the standard protocol, inside the same package. (Though it might use > 1 channel I guess?) [1] eg in the spec table here https://www.anandtech.com/show/16235/apple-intros-first-three-apple-silicon-macs-late-2020-macbook-air-13inch-macbook-pro-mac-mini https://www.anandtech.com/show/16235/apple-intros-first-thre... - an interesing thinb in the mini spec table is also the 10x downgrade of the ethernet speed.
  - floatboth 6y ago
    LPDDR channels are weird. https://www.anandtech.com/show/16226/apple-silicon-m1-a14-deep-dive https://www.anandtech.com/show/16226/apple-silicon-m1-a14-de... shows "8x 16b LPDDR4X channels" Which would be the same width as dual channel DDR4 – 8x16 == 2x64 :) Also it's completely ridiculous that some people think it might be HBM. The presentation slide (shown in both anandtech articles) very obviously shows DDR style chips, with their own plastic packaging. That is not how HBM looks :) Also HBM would noticeably raise the price.