[Rumor, Tweaktown] AMD to launch next-gen Navi graphics cards at E3

[Auer]

Intel doesn't have good brand value for GPUs and their onboard graphics don't have a great reputation. Even before Zen, the AMD APUs were often favored in gaming benchmarks just because Intel's performance was so lackluster. I suppose it's possible that they can achieve parity with AMD/NVidia right out of the gate, but I just don't find it likely. Realistically though they just need a solid mid-range product at a reasonable price and they'll get plenty of people willing to try it out.

If Intel has high prices is will be because they're also wafer constrained. The transition to 10nm has been rocky to say the least so they've been stuck making products on some variation of their 14nm node long after that product should have transitioned to the new node.

My money is on Intel not coming out modestly in any way, price or performance.

Pretty sure they want to make an impact and modest pricing / performance ain't going to achieve that at this stage.

I wouldn't compare anything Intel has done in the past with GPU's with what they are aiming for now.

 

[Stuka87]

My PNY RTX2070 OC at stock from a Superposition bench run:

Thats been proven to not be accurate for total power consumption of the whole graphics card. The only way to get that is to measure all power being consumed by the card, similar to how TPU does it. That way power consumption for the memory, and any efficiency losses in the power distribution and such are measured in as well.

But HWMonitor is fine for giving a rough estimation of what the chip is using.

 

[Mopetar]

My money is on Intel not coming out modestly in any way, price or performance.

Pretty sure they want to make an impact and modest pricing / performance ain't going to achieve that at this stage.

I wouldn't compare anything Intel has done in the past with GPU's with what they are aiming for now.

How do you make an impact that actually matters in any other way than ensuring that people buy your product and start developing brand loyalty? They can bank on some of the loyalty they have from the CPU market and that might just be good enough to get some customers to buy an Intel GPU at equal price/performance, but you have to remember that the people they're trying to convert will also be loyal NVidia customers who are already inclined to buy the next GeForce card all else the same. If you want to lure away loyal customers from someone else, you need to give them a reason to switch.

NVidia has even reported lower sales than expected so the market is not responding well to the prices that currently exist. I have a feeling that if AMD weren't supply constrained, they'd quickly find a lot of surplus inventory of Navi cards building up at their prices. Intel is not going to have a good time designing the massive kind of die they'd need to compete with NVidia's Titan cards and trying to charge a premium for a product that isn't clearly superior.

Instead make something aimed at the mainstream which will yield well and allow Intel to take advantage of their ability to do volume production effectively. Get a solid revenue stream that can support continued development. Look at all of the other categories Intel has tried to get into over the years where they completely failed to gain any appreciable part of the market and eventually axed the product as a result. I think that the management is slowly realizing that outside of where they're already strong and have a history of being on top, no one really gives a hoot about the Intel brand or is willing to pay for it.

 

[guachi]

TBH I expect Intel to be priced high. Possibly even higher than AMD.

On the one hand, Intel has shown they are willing to lose money to gain market share. On the other hand, they may have decided it was a stupid idea not to be repeated.

I could see the place Intel attacks being the laptop market as AMD is weak there.

 

[Glo.]

Last time I checked, it's the Navi architecture having to boost to 1.9ghz out of a stock card just to match/beat TU106 in performance. Nvidia didn't jump on the 7nm train right away because fighting with AMD, Qualcomm, Huawei and Apple for hotly contested, likely expensive early wafer volume is expensive and unnecessary when they still have a significant lead in efficiency due to their architecture and can afford to wait for the process to be cheaper and mature.

RTX 2070 boosts in games to 1.9-2 GHz range. Those types of clocks will not be as easily achievable on 7 nm process, as on 16/12 nm were.

 

[Shivansps]

It's only illegal if there is an agreement in place between the two. This is most likely just a natural result of duopoly. Sucks for the consumer, but at least game devs don't push the envelope as hard as they used to so older cards tend to last longer.

There is no evidence to support either theory, it could be an arregement or it could not, but there is a reason of why AMD decided to go against what they did with the RX400-500 series.

This I completely disagree. At this moment in time we have zero info on the ability of AMD to produce Navi10 in significant numbers to meet serious demand, hence calling them out for price fixing is premature at best.

@Mopetar already did a great job explaining why supply is likely to be extremely limited for the time being.

There is one small issue with that, AMD cannot supply enoght Navis for the PC market for a rasonable price, but they can supply enoght to make PS5 and xbox scarlet happen that includes what it looks lke a more powerfull version of the Navis 10 we are getting AND 8 Zen 2 cores, making the whole console cost around what just the RX5700XT costs. Im sorry but i dont belive that.

I dont trust AMD, Intel, Nvidia, Microsoft and even less Sony. Nvidia was abusing his position and AMD has some explaining to do here. PC components prices has been increasing over the past couple of years with no aparent stop, there is a clear trend here.

BTW, this is not the first time it happened, remember how the HD5870 was the high end and the HD5770 the midrange? guess what happened after that.... HD6970 was the high end, and the HD6870 the mid range with an increase in prices. Nvidia did the same, this is happening again.

 

[Thala]

Again, when you comparing the Chip you have to compare it in iso perf OR iso power.

10% higher performance dosnt scale linearly with power consumption. You cannot say NAVI 10 is not as efficient as Turing 106 because RX 5700XT will have to be 18% faster than RTX 2070. Currently, We do not know the perf/watt of NAVI 10 when compared at the same performance as the Turing 106 on the RTX 2070 Graphics card.

Do you really know today without having the RX 5700 XT in your bench table how much more efficient the NAVI 10 will be at the same performance as RTX 2070 ??

What nonsense is this? When comparing architectures efficiency, neither need to be at iso power or iso performance - this would be incredibly stupid anyway. If you do want to equalize something it would be voltage.

 

[Stuka87]

What nonsense is this? When comparing architectures efficiency, neither need to be at iso power or iso performance - this would be incredibly stupid anyway. If you do want to equalize something it would be voltage.

Yes, they do. Voltage is irrelevant by itself, as you cannot get wattage without also having the current. And Perf:Watt is never scales linearly. Power Consumption will start to bell curve up with a linear increase in performance.

If comparing GPU A (lets say 14nm) and GPA B (lets say 7nm), you either measure performance while both are at the same power consumption, or you measure power consumption with both at the same performance. Only then can you get a Perf:Watt comparison between the two.

 

[Thala]

If comparing GPU A (lets say 14nm) and GPA B (lets say 7nm), you either measure performance while both are at the same power consumption, or you measure power consumption with both at the same performance. Only then can you get a Perf:Watt comparison between the two.

You can and should do precisely what you describe above, when you want to compare process X at 14nm with process Y at 7nm keeping the implementation invariant. This is precisely the reason you get these two reference points iso-performance and iso-power when a foundry advertises Process X vs Process Y. However when comparing 2 architectures or even 2 implementations of the same architecture requiring iso-power or iso-performance makes no sense at all.

This should get clear to you, when you consider that A might be designed for a much higher performance and power point compared to implementation B. Trying to equalize power or performance would totally skrew your results. What you should do instead is taking voltage and frequency out of the equation. This is typically done by the notion of Cdyn - the dynamic capacitance - with electrical unit F (Farad).

In short, before making such claims be clear of the invariants and which variables you want to reason about.

 

[Ajay]

First: Nvidia GPUs still clock higher than AMD's. Secondly, if TSMC's process woiuld be good enough we would already see Nvidia use it.

That is conjecture, not evidence. We do not know, as of yet, why Nvidia is waiting. I don’t think we even know if NV's next GPU will be on TSMC or SS 7nm EUVL. I suspect, but do not know, that the opportunity cost was too high for NV, as @insertcarehere suggested.

N7 can hit what 4.7 GHz on N7 with Ryzen, seems like the limitation isn’t process, but other factors, as, is likely, power consumption when it comes to large GPUs (and perhaps yields). Anyway, even what I’ve noted is only conjecture - not fact. We don’t know what we don’t know. Simply basing opinions on past experience, without concrete vendor/fab data, leads us down the primrose path to irrelevancy.

 

[prtskg]

In last few years AMD hasn't competed with Nvidia properly. Hence the prices of fast gpus are high now. I have a feeling that by next year prices will be much better.

According to the prices from the above link you provided,


Going from 28nm to 14nm we had a 25% increase in Wafer price

Going from 14nm to 7m we have a 69% increase in Wafer price.


Now, assuming both 14nm and 7nm have the same yields at the same time frame and because Polaris 10 and Navi 10 have almost the same die size,

MSRP for RX480 8GB (POLARIS 10) was $239
MSRP for RX 5700XT (NAVI 10) is $449

~88% increase of price

Since NAVI 10 card MSRP price has increased almost 88% over the POLARIS 10 card and Wafer price only got an increase of 69%, then Margins are higher with NAVI vs Poalris.

But this also assumes that all other BOM cost of the two cards between RX480 and RX 5700XT remained the same.

Looking at these calculations, I think as competition increases AMD can decrese price by about 10-15%. I expect these cards to be more competitve to their counterparts than Polaris was and hence more pricey, at least until Nvidia brings price lower.

 

[Head1985]

This I completely disagree. At this moment in time we have zero info on the ability of AMD to produce Navi10 in significant numbers to meet serious demand, hence calling them out for price fixing is premature at best.

They making 331mm2 vega20 die for how long?Probably close to 1year already.Navi10 is tiny 251mm2 and 7nm is not new at this point.Only reason navi cost what it cost is that AMD milking consumers just like Nvidia.
But 2070 launched 9month ago when Nvidia dint have any competition at ALL so they ovepriced it like crazy.And now 9months later AMD price matching it.If this is not price fixing then i dont know what evidence you need.Tell me when AMD last time undercut NV at launch?
FURY X? nope
RX480? nope
RX580? nope
Vega64? nope
RX590? nope
Radeon7? nope
Navi? yeah 50less vs probaly one of the worst priced card in history...Whuhuu lets celebrate and lets be honest here AIB cards will cost same so they didnt undercut it.

 

[Auer]

Thats been proven to not be accurate for total power consumption of the whole graphics card. The only way to get that is to measure all power being consumed by the card, similar to how TPU does it. That way power consumption for the memory, and any efficiency losses in the power distribution and such are measured in as well.

But HWMonitor is fine for giving a rough estimation of what the chip is using.

I get the same result with HWinfo, and PNY specs the card at 185 TDP.

I'm ok with those numbers.

 

[Glo.]

That is conjecture, not evidence. We do not know, as of yet, why Nvidia is waiting. I don’t think we even know if NV's next GPU will be on TSMC or SS 7nm EUVL. I suspect, but do not know, that the opportunity cost was too high for NV, as @insertcarehere suggested.

N7 can hit what 4.7 GHz on N7 with Ryzen, seems like the limitation isn’t process, but other factors, as, is likely, power consumption when it comes to large GPUs (and perhaps yields). Anyway, even what I’ve noted is only conjecture - not fact. We don’t know what we don’t know. Simply basing opinions on past experience, without concrete vendor/fab data, leads us down the primrose path to irrelevancy.

N7 can hit 4.7 GHz on Matisse, because it is small die. Anything big will not get as good performance and power out of it.

And for this node anything big is bigger than 200 mm2.

 

[DisEnchantment]

And it has arrived, AMDGPU patches for Navi 10.
https://lists.freedesktop.org/archives/amd-gfx/2019-June/035170.html

This is still in the mail archives, it will go to staging soon, after which we can go throught the code line by line
The patch is way bigger than the Vega enablement patches.

First glance:
So many new blocks
VCN2
DCN2
new SMU
XGMI ?? Navi 10 has IF2

And it is here
https://cgit.freedesktop.org/~agd5f/linux/log/?h=amd-staging-drm-next-navi10

{0x1002, 0x7310, PCI_ANY_ID, PCI_ANY_ID, 0, 0, CHIP_NAVI10},
{0x1002, 0x7312, PCI_ANY_ID, PCI_ANY_ID, 0, 0, CHIP_NAVI10},
{0x1002, 0x7318, PCI_ANY_ID, PCI_ANY_ID, 0, 0, CHIP_NAVI10},
{0x1002, 0x7319, PCI_ANY_ID, PCI_ANY_ID, 0, 0, CHIP_NAVI10},
{0x1002, 0x731A, PCI_ANY_ID, PCI_ANY_ID, 0, 0, CHIP_NAVI10},
{0x1002, 0x731B, PCI_ANY_ID, PCI_ANY_ID, 0, 0, CHIP_NAVI10},
{0x1002, 0x731F, PCI_ANY_ID, PCI_ANY_ID, 0, 0, CHIP_NAVI10},

 

[ElFenix]

but there is a reason of why AMD decided to go against what they did with the RX400-500 series.

RX480 8GB came out at "$239", which was barely under the 1060 6GB's "$249."* if anything, AMD's SRP is more price competitive now, assuming its claims about in game (nonRT) performance are anywhere near correct



*non FirstadoptersEdition pricing

 

[tviceman]

The problem here is that it seams you dont understand the concept of perf/watt.

You can compare Graphics cards and come up with a perf/watt but that is compering the graphics card as a product and NOT the Chip.

In that example you may compare RX 5700 XT vs RTX2070 and say that RX 5700XT Grapics Card has lower efficiency in perf/watt than RTX 2070 Graphics Card.

IF you want to compare the NAVI 10 Chip in perf/watt then you have to do it in iso perf OR at iso power.

In that example you take both cards (RX 5700XT and RTX2070) but this time you put them at the same performance (fps) OR at the same power (watt) and you come up with the Chip perf/watt efficiency.

So when you are comparing NAVI 10 to TU106 you must state a context, you either compare the Graphics Cards (RX 5700XT vs RTX2070) or you compare the chips at iso perf OR at iso power.

What you did above was comparing only the Graphics Card perf/watt since you compared the perf/TDP of the Graphics Cards and not the perf/watt at iso Perf or at iso Power of the Chip as described above.

Absolutely nobody, and I mean nobody except AMD, uses the power consumption only of the chip to derive the perf/w metric of the product when giving reviews or performance figures. The consumer product is not just the GPU - it's entire board. And the entire board is what EVERYONE derives their power consumption figures from.

This obscure metric you're trying to presumptively use to put Navi in a better light is not a metric any reputable site will use. Techpowerup, Anandtech, Techreport, Hardwarecanucks, guru3d.... every single site uses the power consumption of the board to test perf/w because that is the power being consumed by the product.

 

[tviceman]

Most 3rd party RTX 2070's consume between 215-230W when gaming. But until we have actual reviews, we simply do not know.

So to reiterate, no, we DO NOT KNOW if that isn't the case. It may turn out to be, it may turn out to not be. But until we have actual facts, don't claim conjecture as facts.

Most 3rd party graphics cards from all vendors, regardless of AMD or Nvidia, use more wattage than reference cards because they come with more fans, stronger power phases, clocked higher, etc. Reference to reference has been and continues to be the only acceptable comparison when finding performance figures, power consumption, general overclocking capabilities, etc.

If aftermarket cards are used in comparisons then it's most fair if every card is after market and of the same caliber.

 

[Stuka87]

I get the same result with HWinfo, and PNY specs the card at 185 TDP.

I'm ok with those numbers.

Yeah, GPUz/HWMonitor both measure the same thing, which is estimated chip power. And the 185 TDP is most likely spot on. But TDP equals Thermal Design Power, which is a measurement for the chip itself, not Typical Board Power, which is power consumption of the entire board. Which gets back to my original point of board power for the 2070 is very similar to what AMD is stating board power of the 5700 XT is. We just have to wait for tests to see how close AMD's spec is to real world.

 

[tviceman]

RX480 8GB came out at "$239", which was barely under the 1060 6GB's "$249."* if anything, AMD's SRP is more price competitive now, assuming its claims about in game (nonRT) performance are anywhere near correct



*non FirstadoptersEdition pricing

It would be nice if AMD's figures hold true, but given their Vega 64 "you can't tell a difference" campaign and their internal benchmarks putting Fury X above the 980 TI, I remain highly, highly skeptical until independent reviews are out.

 

[Stuka87]

Most 3rd party graphics cards from all vendors, regardless of AMD or Nvidia, use more wattage than reference cards because they come with more fans, stronger power phases, clocked higher, etc. Reference to reference has been and continues to be the only acceptable comparison when finding performance figures, power consumption, general overclocking capabilities, etc.

If aftermarket cards are used in comparisons then it's most fair if every card is after market and of the same caliber.

Yes, you are correct, the aftermarket cards are using around 20W more than the reference. Which means the 5700 XT would need to use less than advertised and have a performance advantage over the 2070 to have equal Perf:Watt.

 

[lobz]

In a competitive climate, for everyone. If nvidia make a great product, it sells extremely well, then both the consumer and shareholder get benefit. If nvidia make a bad product, it sells poorly, both the consumer and shareholder are hurt. This is really basic, but the reality falls somewhere on the shades of gray of these scenarios.

In the case you and Hans Gruber are discussing, nvidia only need to lower their prices if their sales fall, because in reality, even if the AMD cards are better, nvidia has the mindshare and can overprice their cards a little more if they wish, and still make a healthy profit. Price them too high, and overall profit decreases due to sales decline. Price too low, the same happens because margin per card is too low.

However, nvidia may choose to lower prices not due to a sales decrease, but to maintain marketshare and mindshare, thus hurting AMD's efforts to gain on them. Because nvidia has a lot of capital and market cap to work with, they can take some "hypercompetitive" routes to continued success against AMD. Such a strategy might not work against Intel, but they (nvidia) may consider it easier to fight a giant (Intel) when you don't have a yippy dog (AMD) nipping at your ankles.

that's exactly what I tried to imply in the current environment, it's actually very unlucky for us, customers

 

[Ajay]

N7 can hit 4.7 GHz on Matisse, because it is small die. Anything big will not get as good performance and power out of it.

And for this node anything big is bigger than 200 mm2.

Whoosh! I think anything larger than 171.283 mm2 is big on N7

 

[Shivansps]

RX480 8GB came out at "$239", which was barely under the 1060 6GB's "$249."* if anything, AMD's SRP is more price competitive now, assuming its claims about in game (nonRT) performance are anywhere near correct



*non FirstadoptersEdition pricing

The RX480 came out first, matching GTX970 perf at GTX960 prices, GTX1060 came out a month later or so. RX480 replacement now matches RTX2070 non RT perf at RTX2070 prices. I dont see how that is more price competitive in any way. What they did with the RX480 was normal, matching old tech perf at a lower price.

As i said if Navi 10 were at RTX2060 and GTX1660TI prices, while not perfect(it should be GTX1660TI and non TI) it would be similar to the RX480 launch, they would still have a very considerable price bump comprared to RX400 series, but it would be another thing.

Imagine Nvidia charging $449 for the GTX1050 2Gb just because it has GTX680 2GB perf +5%, that is exactly what AMD did here, saving the distances.

 

[piesquared]

Sounds like the people claiming to be upset about the 5700XT price better high tale it over to nvidia forums or email them directly with their complaints since they are the ones setting the pricing in the GPU market currently.