I'm looking for a new graphics card and scan seem to be offering this one, which has a faster core clock than the ebuyer one.
Top comments
mrcyco
30 Jun 1610#1
unless you NEED Nvidia (Hairworks, Nvidia controller... moonlight), you should really consider the RX480. on par with the 970 but more ram, new chip set (so more room to grow) and the 8gb wouldn't cost to much more than this.
ReadySetGoGo
30 Jun 168#2
That is a good price for the 970 and means Nvidia have recognised how good AMD's new 480 is. I still went with the 480 though as it's newer technology and likely to improve performance over time compared to the 970.
Elevation
1 Jul 167#15
BetaRomeo
1 Jul 164#31
Everything you've said is correct, but it's stuck in the middle and failed to reach the conclusion.
Maxwell and Pascal can do graphics and compute tasks simultaneously, albeit fewer than AMD's line-up (as you can see when doing too many forces a context shift on Nvidia hardware). So by improving their pre-emption and context switching, Nvidia have allowed their cards to still run at pretty much maximum capacity, even without async compute. Their DX11 instruction scheduling is allowing them to max out their cards.
AMD, on the other hand, have had bigger chips with higher theoretical performance for years (just compare the TFLOPs), and still been beaten by their Nvidia equivalents, as despite their GCN architecture, when it comes to processing their workflow they are at a distinct disadvantage in larger applications such as games (but you're right, mining is handled fine). Their architecture allows instructions to flow without stacking, which results in a reduced flow of calls when compared with scheduling - which might sound like an odd choice, but does produce a significant reduction in latency.
Async compute helps alleviate this scheduling bottleneck that we find in DX11 games on AMD GPUs - again, in larger applications - and allows AMD's chips to reach what we'd expect from them (in addition to DX12 relieving some CPU pressure). This is pretty much spelled out by the hardware specifications of the Nvidia and AMD cards, and their comparative performances.
Throwing async compute onto Maxwell would hardly benefit Nvidia at all (hmmm... although it might improve mining performance?), because the hardware is already approaching its limits while using a scheduler-based architecture - clever pre-emption is the rough equivalent they use instead. I'm still working my way through the Pascal white papers, but it seems to be the same situation there (unless you've found something different..?).
TL;DR: Nvidia's architecture for DX11 was entirely API-appropriate, and maxed out their hardware, while we're only now seeing the true benefits of the untapped potential of AMD's GCN in DX12.
All comments (75)
mrcyco
30 Jun 1610#1
unless you NEED Nvidia (Hairworks, Nvidia controller... moonlight), you should really consider the RX480. on par with the 970 but more ram, new chip set (so more room to grow) and the 8gb wouldn't cost to much more than this.
mikem1989 to mrcyco
1 Jul 16#20
+1
RX480 is much more future proof.
ReadySetGoGo
30 Jun 168#2
That is a good price for the 970 and means Nvidia have recognised how good AMD's new 480 is. I still went with the 480 though as it's newer technology and likely to improve performance over time compared to the 970.
Rhythmeister
30 Jun 16#3
Faster than the GTX 970 in the real world, just ask Techpowerup :wink: Probably better to wait for the GTX 1060 actually, it's getting announced on the 7th of July. Personally I'd rather support the underdog BUT my new monitor happens to support G-Sync and can do 144Hz so I may have to let the rest of you support them for me :man:
LazybeatX
30 Jun 16#4
Decent price and still tempting over the 480 until the aftermarket cards arrive at least.
csf
30 Jun 162#5
Not to hijack the thread but doesn't the rx 480 draw to much juice from the pci slot, more than spec allows? thus causing eventual issues?
Opening post
Top comments
Maxwell and Pascal can do graphics and compute tasks simultaneously, albeit fewer than AMD's line-up (as you can see when doing too many forces a context shift on Nvidia hardware). So by improving their pre-emption and context switching, Nvidia have allowed their cards to still run at pretty much maximum capacity, even without async compute. Their DX11 instruction scheduling is allowing them to max out their cards.
AMD, on the other hand, have had bigger chips with higher theoretical performance for years (just compare the TFLOPs), and still been beaten by their Nvidia equivalents, as despite their GCN architecture, when it comes to processing their workflow they are at a distinct disadvantage in larger applications such as games (but you're right, mining is handled fine). Their architecture allows instructions to flow without stacking, which results in a reduced flow of calls when compared with scheduling - which might sound like an odd choice, but does produce a significant reduction in latency.
Async compute helps alleviate this scheduling bottleneck that we find in DX11 games on AMD GPUs - again, in larger applications - and allows AMD's chips to reach what we'd expect from them (in addition to DX12 relieving some CPU pressure). This is pretty much spelled out by the hardware specifications of the Nvidia and AMD cards, and their comparative performances.
Throwing async compute onto Maxwell would hardly benefit Nvidia at all (hmmm... although it might improve mining performance?), because the hardware is already approaching its limits while using a scheduler-based architecture - clever pre-emption is the rough equivalent they use instead. I'm still working my way through the Pascal white papers, but it seems to be the same situation there (unless you've found something different..?).
TL;DR: Nvidia's architecture for DX11 was entirely API-appropriate, and maxed out their hardware, while we're only now seeing the true benefits of the untapped potential of AMD's GCN in DX12.
All comments (75)
RX480 is much more future proof.
https://www.scan.co.uk/products/4gb-evga-geforce-gtx-970-sc-gaming-acx-20-pcie-30-7010mhz-gddr5-gpu-1165mhz-boost-1317mhz-cores-1664