Overclock.net - An Overclocking Community - View Single Post - OC'ing / Pushing my vega56 to the Max

View Single Post
post #4 of (permalink) Old 02-27-2019, 11:12 AM
⤷ αC
AlphaC's Avatar
Join Date: Sep 2012
Posts: 11,177
Rep: 904 (Unique: 590)
Yes it is. Basically Gigabyte decided to dump some money into the GPU cooling (they put a heatpipe into the backplate as a result of the power design) and then cheapen out a bit on the power delivery with two AON6360 as a replacement of sorts of the IRF6894. It isn't as big an issue for gaming or whatnot, but being the only card setup with that sort of power delivery means you inherently use more power. The thermals don't reflect the losses as much since they used 2x low side fets but if you look at Sapphire Nitro results they're at least 10°C cooler even on Vega64.

Six IR3598 interleaved MOSFET drivers, three on the front and three on the back, are responsible for this doubling. The actual voltage conversion for each of the 12 regulator circuits is handled by one Alpha & Omega Semiconductor AON6594 on the high side and two AON6360s in parallel on the low side. This is an inexpensive but acceptable choice of components, especially since the parallel arrangement also means that thermal hot-spots are distributed more evenly.

A look at the board's back side reveals that half of the most heavily-used low-side MOSFETs are relocated there. So, it's safe to assume that ~30% of all voltage converter losses (and associated heat) occur on the back of the card. Naturally, then, it isn't enough to only dissipate heat from the board's front. Cooling is now required on both sides.

3mΩ @ 10V , 5mΩ@4.5V V_GS

7mΩ @ 10V, 11mΩ@4.5V V_GS


Reference uses IRF6894 and IRF6811. The IRF6894 has a 0.9 mΩ RDSON @ 10V , 1.4m mΩ @ 4.5V and the IRF6811 has 2.8 and 4.1 mΩ resistance respectively. Q_G (total gate charge) for IRF6811 high side fet is extremely low at 11nC , which means switching is less lossy.

https://www.infineon.com/dgdl/irf6894mpbf.pdf?fileId=5546d462533600a4015355f0cc0 41abc

If you look at other nonreference designs such as the ASUS STRIX version , it's using 60A IR powerstages (https://www.photopoint.ee/en/compute...vega-56-oc-8gb , https://www.hardocp.com/article/2018...aming_review/3, details gained from VEGA64 version https://www.hardwareluxx.de/index.ph...t.html?start=2).

Powercolor's Red devil is essentially the reference design: https://www.guru3d.com/articles-page...-review,4.html , https://www.hardwareluxx.de/index.ph...t.html?start=2

XFX's Double edition is basically reference Nano with cut down the phase count to 8 (instead of 12): https://www.overclock.net/forum/26648417-post5895.html , https://hardforum.com/threads/xfx-ve...vesty.1953602/

Sapphire's RX VEGA NITRO series is 14 phases up from the 12 on reference.

Sapphire also has a RX VEGA Nano based "Pulse" that has a conservative power limit (minor bump from reference) https://www.computerbase.de/2018-03/...as_winzige_pcb

One issue with the reference Vega 56 in general is that the only aftermarket cooler options are pretty much the Raijintek Morpheus II and waterblocks of some sort (whether they are Asetek AIOs , Alphacool Eiswolf AIOs, or full custom). This is due to HBM memory and the VRM layout. Arctic Accelero Xtreme air cooler doesn't seem to work apparently (that is an option instead of a waterblock for a Nvidia card).

I feel as though Sapphire was the only vendor that did VEGA56 and VEGA64 justice. They have one conservative card that's one centimeter longer than reference and an overkill overclocking powerhouse with 3 PCIE power connectors and a massive triple slot heatsink that only needs ~1200RPM fan speed.

► Recommended GPU Projects: [email protected] , [email protected] (FP64) (AMD moreso) ► Other notable GPU projects: [email protected] (Nvidia), GPUGrid (Nvidia) ► Project list

AlphaC is offline