AMD R9 470X exposure: TDP is only 60W

[Burpo]

"AMD is currently a low-end graphics card Radeon R9 470X exposure, alleged that the card uses the new Polaris 11 architecture, TDP of only 60W, that is to say no external auxiliary power supply for normal operation.
It is reported that, R9 470X will use 1280 stream processors, 1GHz core frequency, 80 texture units, 40 raster units, the use of 4GB GDDR5 memory, memory frequency 7Ghz, memory interface is 128Bits, 112GB / S memory bandwidth."

Retailing at $169 seems like good card for the money. My next AMD purchase? Hmmm, I could use a few of these.
http://www.ithome.com/html/digi/225244.htm

 

[MrTeal]

Just about the perfect replacement for my aging 6950 in the HTPC.

 

[ClockHound]

Intriguing. Price seems a tad high tho. A card without a 6 or 8 pin power header can't possibly worth more than $100-120. Unless it has the Green Marketing Machine powering it.

 

[Lepton87]

This is the same MSRP as Radeon R270 so it seems way too high. R270 has a 256bit bus with a 212mm2 GPU so in terms of size and memory-bus it is more comparable to Polaris. Of course the new R9 470X should be faster than R270 but still it should cost 130$ tops.

 

[VirtualLarry]

So this is basically a 270X, with 4GB rather than 2GB, that doesn't need a 6-pin PCI-E? Sounds like a winner, sort of. Better than the 750Ti, maybe similar to the GTX950(A) boards that lack the additional power connector.

Still, yeah, price is probably a tiny bit too high.

I could see $150. Then again, it's got 4GB, not 2GB, so that's worth some money too.

 

[mohit9206]

Unimpressive.128 bit bus is unimpressive.$169 price point is unimpressive.Aftermarket cards costing $179-189 is even more unimpressive.You can already buy R9 380 for $180 so the only thing this card brings is much lower power consumption.That's lame.And unimpressive.

 

[Madpacket]

Looks like the 750/Ti is finally being dethroned as the most power efficient card. I just hope low profile version exist shortly after launch. These will be the new ideal HTPC / gaming cards.

 

[Madpacket]

Unimpressive.128 bit bus is unimpressive.$169 price point is unimpressive.Aftermarket cards costing $179-189 is even more unimpressive.You can already buy R9 380 for $180 so the only thing this card brings is much lower power consumption.That's lame.And unimpressive.

Who cares about bus bandwidth if it performs on par or better than 256 bit cards?

 

[IEC]

Who cares about bus bandwidth if it performs on par or better than 256 bit cards?

^this.

I could see myself picking up one of these if the price is right. Solid HTPC/light gaming contender if the range of expected performance estimates is in the ballpark.

 

[mohit9206]

So instead of pushing the performance forwards at a given price point, you are happy with AMD just reducing the power consumption instead?
Who cares about power consumption. Instead of 60W 470x performing like a R9 380 for $150,I'd rather have a 100W 470X performing like somewhere between 380X and 390. Its all about price /performance not performance /watt. That's what PC gamers want. More performance at lower price. Not same performance at lower wattage.

 

[Layer8]

Who cares about power consumption.

The OEMs care. They can go with a less powerful power supply and they have one 6-pin ATX cable less.
In this case, a 60W card is way cheaper for them.

More performance at lower price.

Shouldn't that read "more performance for the same price"?

Not same performance at lower wattage.

AMD designed it that way for the above reasons. (And maybe some more.)

You can try to overclock, but you might be better off just buying a similarly priced outdated higher end model. But you won't get that with 100W TDP.

 

[crisium]

If true, I could see this matching 380 performance thanks to much higher clockrates after an overclock. Though at stock it will still essentially be 7870/270/270X performance. As long as AMD can properly compress that low memory bandwidth. That's the same amount that the 960 has and it bites at the heels of a 380 already, so it's not out of the question.

I don't know why they would only do 1000Mhz though. Moving to 14FF should allow for far, far higher frequencies. We already had a 1280SP 1000MHz card nearly 4 and a half years ago, called the Radeon 7870. So why this low frequency? The only reason I can think of is the power consumption of AMD cards must still skyrocket with higher clockrates. AMD must want to make sure every card runs without a power connector and this also let's them advertise good performance-per-watt. Naturally from the 14FF alone I'd expect the old 1280SP 1000Mhz to be the new 1280SP 1400MHz, at the very least. But perhaps we will have this with a 1x6pin variant.

40 raster units? Does that mean it will have 40 ROPs? That would be unusual, as I don't think AMD has ever done anything other than 8, 16, 32, or 64 on a card before.

 

[Qwertilot]

The other potential reason for hypothetical low top clocks would presumably be GF's processes. Seriously hope not of course.

We'll know soon enough

Would not, I think, be a huge surprise if AMD had optimised for low clocks/low power operation. Logically with these two chips they'd have consoles as first priority then notebooks, then desktops.

 

[laamanaator]

If true, I could see this matching 380 performance thanks to much higher clockrates after an overclock. Though at stock it will still essentially be 7870/270/270X performance. As long as AMD can properly compress that low memory bandwidth. That's the same amount that the 960 has and it bites at the heels of a 380 already, so it's not out of the question.

I don't know why they would only do 1000Mhz though. Moving to 14FF should allow for far, far higher frequencies. We already had a 1280SP 1000MHz card nearly 4 and a half years ago, called the Radeon 7870. So why this low frequency? The only reason I can think of is the power consumption of AMD cards must still skyrocket with higher clockrates. AMD must want to make sure every card runs without a power connector and this also let's them advertise good performance-per-watt. Naturally from the 14FF alone I'd expect the old 1280SP 1000Mhz to be the new 1280SP 1400MHz, at the very least. But perhaps we will have this with a 1x6pin variant.

40 raster units? Does that mean it will have 40 ROPs? That would be unusual, as I don't think AMD has ever done anything other than 8, 16, 32, or 64 on a card before.

Well, GCN 4 (Polaris) is basically an almost complete overhaul of previous GCN architectures, so we don't know for example about any possible IPC per SP improvements, that could've been done. So A GCN 4 1280SP GPU could be twice or trice as fast as GCN 1-3 1280SP equivelant at a given clock. :sneaky: But realistically speaking an IPC improvement of around 30 to 60% is IMO much more realistic and plausible than 2-3X. The reason why I think AMD might have achieved that kind of improvement in SP IPC is that they didn't implement the improvements that were designed to be used in 20nm products, which were never released because of the crappy TSMC 20nm process. Three years (after Hawaii launch) is a long time to do some gpu magic.

 

[krumme]

Mobile is going to take all capacity. Therefore there will be no cheap 14nm stuff for desktop the first year.
Samsung process is probably dense and cheap in gf variant but lower freq than tsmc. When gf production ramp up hopefully we can get rid of all the old 28 nm stuff that have been here for far to long.

 

[drwire]

Well, GCN 4 (Polaris) is basically an almost complete overhaul of previous GCN architectures, so we don't know for example about any possible IPC per SP improvements, that could've been done. So A GCN 4 1280SP GPU could be twice or trice as fast as GCN 1-3 1280SP equivelant at a given clock. :sneaky: But realistically speaking an IPC improvement of around 30 to 60% is IMO much more realistic and plausible than 2-3X. The reason why I think AMD might have achieved that kind of improvement in SP IPC is that they didn't implement the improvements that were designed to be used in 20nm products, which were never released because of the crappy TSMC 20nm process. Three years (after Hawaii launch) is a long time to do some gpu magic.

Why not 2.5 "brighter" than previous generation?.

 

[DeathReborn]

I am thinking about 40-50% improvement in performance per core is what GCN 4 brings over previous iterations.

896 GCN4 providing almost identical GFLOPS to the 1280 GCN1 270X.

 

[dark zero]

Intriguing. Price seems a tad high tho. A card without a 6 or 8 pin power header can't possibly worth more than $100-120. Unless it has the Green Marketing Machine powering it.

Or unless is passive too. Passive cards cost a little more than fanned ones.

 

[RussianSensation]

I am thinking about 40-50% improvement in performance per core is what GCN 4 brings over previous iterations.

896 GCN4 providing almost identical GFLOPS to the 1280 GCN1 270X.

In an interview, AMD themselves said that 30% of the improvement comes from new architecture and 70% from the new 14nm FF node. Therefore, it cannot be a 40-50% increase in IPC when the absolute maximum AMD has said is 30%. The only way to raise IPC to 40-50% per shader is going to be via much higher clocks than existing GCN parts.

The up to 2.5X perf/watt over older GCN GDDR5 parts already includes the 30% improvement from the architecture and 70% improvement from the 14nm node. Take that as you will.

 

[raghu78]

In an interview, AMD themselves said that 30% of the improvement comes from new architecture and 70% from the new 14nm FF node. Therefore, it cannot be a 40-50% increase in IPC when the absolute maximum AMD has said is 30%. The only way to raise IPC to 40-50% per shader is going to be via much higher clocks than existing GCN parts.

The up to 2.5X perf/watt over older GCN GDDR5 parts already includes the 30% improvement from the architecture and 70% improvement from the 14nm node. Take that as you will.

RS 2.5X perf/watt comes from Process node + architecture. The gain is 1.5x or 150% which when split 70/30 would mean 45% from architecture and 105% from process node. So if 45% perf/watt gain purely from architecture means the perf/throughput has increased 45% at a given wattage purely from architecture. Make of that what you will.

btw can you explain why you think AMD refers to older GCN GPUs when the slide refers to 28nm GPUs in the late 2014/2015 timeframe.

 

[Arachnotronic]

RS 2.5X perf/watt comes from Process node + architecture. The gain is 1.5x or 150% which when split 70/30 would mean 45% from architecture and 105% from process node. So if 45% perf/watt gain purely from architecture means the perf/throughput has increased 45% at a given wattage purely from architecture. Make of that what you will.

btw can you explain why you think AMD refers to older GCN GPUs when the slide refers to 28nm GPUs in the late 2014/2015 timeframe.

You are reading way too into that roadmap. Obviously AMD is going to choose the 28nm part that allows it to show the biggest increase in perf/watt, come on now. Marketing 101.

 

[Arachnotronic]

Just about the perfect replacement for my aging 6950 in the HTPC.

The 69xx series cards were quite nice. Let me know if you are going to sell yours, I'd be interested in buying as a souvenir

 

[renderstate]

IPC, by definition, doesn't go up with cores frequency. In an ideal word it doesn't scale at all but in practice it will always go down.

 

[Magic Carpet]

Instead of 60W 470x performing like a R9 380 for $150,I'd rather have a 100W 470X performing like somewhere between 380X and 390. Its all about price /performance not performance /watt

Looks like 470x isn't a true successor to now a four year old 78x0/270.

Should of been named 460 or 450, really. That's what it is, entry level GPU. Name it appropriately for Christs sake! AMD PR :thumbsdown:

If any of that is true.

 

[William Gaatjes]

In an interview, AMD themselves said that 30% of the improvement comes from new architecture and 70% from the new 14nm FF node. Therefore, it cannot be a 40-50% increase in IPC when the absolute maximum AMD has said is 30%. The only way to raise IPC to 40-50% per shader is going to be via much higher clocks than existing GCN parts.

The up to 2.5X perf/watt over older GCN GDDR5 parts already includes the 30% improvement from the architecture and 70% improvement from the 14nm node. Take that as you will.

This is all guessing from my part :

I am wondering about how much clock cycles a given gcn instruction needs to complete when compared between the different gcn generations.
I am still trying to find some information about it, but it makes sense that AMD already optimized GCN as much as possible.

So, they cannot shave of much (if at all) clock cycles per instruction to increase IPC that way. Because usually for higher clocks, a longer pipeline is needed. I assume here that gcn is pipelined because that makes sense for a gpu which does the same iterative work over and over again. But then again, a new smaller process may allow for less stages for some instructions, so that these instructions complete faster with less clock cycles. This may limit of course the maximum clock speed. But if that is not necessary because the reachable clock speed plus IPC improvements are already enough.

Do you know if there is somewhere a guide about gcn instructionset ?

We could all compare the difference between 1.0,1.1 and 1.2 and make an educated guess how much gcn has improved and how 1.3 may have improved.

edit:

While looking, i found something right here from anandtech.

http://www.anandtech.com/show/4455/amds-graphics-core-next-preview-amd-architects-for-compute/2