T-Minus (3) Days Until VEGA!!

[ilikewhey]

Have you ever thought that Vega GPU in MBP will have 35W Power Limit?

The Vega GPU you are talking about, has very little to do with Vega which will land in MBP. It is the same architecture. But apart from that - it is different chip.

i know, thats why i said revised,

but lets wait and see, hopefully its not another thermal disaster.

if they already have this problem with polaris, and vega being more power hungry chip.

 

[CodeJoy]

Is there a proper Vega 64 vs RTX 2070 shootout (not just games, and including thermals) ?

I am thinking about Red Devil, so we are talking about $600.

Yes, there's still 10-bit and FreeSync.

They're a bit hard to find, most reviews seem to be about games. As a rough order of magnitude of performance potential, you could expect the 2070 to be 10-20% faster for games, but the Vega 64 to be ~60-70% faster for compute. How much of that translates into real world performance is hard to say, and it's not even super clear how a reviewer would make an exact apples to apples comparison with currently available software. For anything more exact I think you'll have to say what you expect to use it for, because the relative performance will differ depending on use case.

For thermals, I have no first-hand experience with the Vega 64, but I have a Vega 56 in my box, and when tuned it's actually quite good. I can't give exact numbers, but it runs cool and reasonably quiet under heavy load. Certainly cooler than my MBP 2018 🙂

 

[koyoot]

i know, thats why i said revised,

but lets wait and see, hopefully its not another thermal disaster.

if they already have this problem with polaris, and vega being more power hungry chip.

It is NOT revised chip.

It is completely brand new design.

False. All Coffee Lake CPUs run at or above advertised spec on the 2018 MBP. Which means able to operate pretty much indefinitely on multi-core workflows while drawing 45 watts (the TDP) and running at or above their minimal (base clock). My i9 has no problem maintaining 3.2Ghz (0.3 Ghz over its advertises spec) in this scenario. In single-threaded workouts it reaches 4.5Ghz (which is very close to its real single-core turbo boost of 4.6 Ghz). We had a thread here comparing results — there are many people who's CPUs run slightly better than mine, but this is down to silicone lottery.

Are you sure it runs on 45W TDP Power Limit with 3.2 GHz?

Im asking this seriously, because it is interesting from the perspective of desktop chips. I would actually love to test this on proper motherboard, and try running it passively. What workloads you have that you can test this?

 

[CodeJoy]

upon release the buyers had to underclock their vega out of the box just to get a decent temperature.

in conclusion, apple basically put a power hungry vega chipset(revised i'm sure, but the principle core is still the same) into a alrdy thermal stressed chassis of the mbp.

I think you mean undervolt, not underclock. The whole point of that is to lower power consumption, while NOT lowering performance. And actually, you can get higher performance. I have first-hand experience with this, and it's also confirmed by many reviewers.

And I don't think you have your facts straight about Vega being a power hungry chipset. Vega is significantly more power efficient than Polaris, which was already very power efficient. What caused the thermal outcry at the Vega launch was that the cards were effectively overvolted and slightly overclocked. And as I'm sure you know, it's an exponential curve. Get a little bit out of the sweet spot and you'll have a massive impact on power draw. That's just how the physics work, it's not a Vega specific thing.

Of course, different chipsets can have different exponential curves. Vega is manufactured on the GF 14nm node, which seems to be quite good for lower clocks, but doesn't scale all that well to higher clocks. That's a big reason behind Vega being what it is. You can contrast this with Nvidia's Pascal chips, which seem to be very overclock friendly. In contrast, the Vega chips can overclock their memory quite favorably, and this has a pretty significant performance impact.

So it really depends on how Apple are configuring the Vega chips. They can get something really power efficient, or they can get something that is a thermal disaster. We won't know until reviewers have done their thing.

 

[ilikewhey]

I think you mean undervolt, not underclock. The whole point of that is to lower power consumption, while NOT lowering performance. And actually, you can get higher performance. I have first-hand experience with this, and it's also confirmed by many reviewers.

And I don't think you have your facts straight about Vega being a power hungry chipset. Vega is significantly more power efficient than Polaris, which was already very power efficient. What caused the thermal outcry at the Vega launch was that the cards were effectively overvolted and slightly overclocked. And as I'm sure you know, it's an exponential curve. Get a little bit out of the sweet spot and you'll have a massive impact on power draw. That's just how the physics work, it's not a Vega specific thing.

Of course, different chipsets can have different exponential curves. Vega is manufactured on the GF 14nm node, which seems to be quite good for lower clocks, but doesn't scale all that well to higher clocks. That's a big reason behind Vega being what it is. You can contrast this with Nvidia's Pascal chips, which seem to be very overclock friendly. In contrast, the Vega chips can overclock their memory quite favorably, and this has a pretty significant performance impact.

So it really depends on how Apple are configuring the Vega chips. They can get something really power efficient, or they can get something that is a thermal disaster. We won't know until reviewers have done their thing.

my apology, yes i meant undervolt, i haven't mess with a vega card since launch, but yes it needs to be undervolt to gain optimal effiency and performance, coming from buying nvidia and used to nvidia's philosphy this doesn't sit right with me.

but yes all of this right now is pure speculation, lets see how the review goes, it won't have an effect on me since i'm going with the mac mini + egpu route anyhow. but for these who have a genuine interest in this i wish you luck
[doublepost=1542136561][/doublepost]

It is NOT revised chip.

It is completely brand new design.

given amd's history of rehashing 290 - 390 i wouldn't hold my breath in "completely brand new design. unless your talking about NAVI thn yes i would definitely agree
[doublepost=1542136947][/doublepost]

False. All Coffee Lake CPUs run at or above advertised spec on the 2018 MBP. Which means able to operate pretty much indefinitely on multi-core workflows while drawing 45 watts (the TDP) and running at or above their minimal (base clock). My i9 has no problem maintaining 3.2Ghz (0.3 Ghz over its advertises spec) in this scenario. In single-threaded workouts it reaches 4.5Ghz (which is very close to its real single-core turbo boost of 4.6 Ghz). We had a thread here comparing results — there are many people who's CPUs run slightly better than mine, but this is down to silicone lottery.

Yes, larger gaming laptops with more cooling power perform better — but this doesn't mean that the MBP performs poorly.



Yes, Vega 10 had some issues, primarily since AMD was forced to compete with Nvidia and its better high-end performance (so they overclocked the cards beyond reasonable limits). Vega 12 uses different internal configuration, a significantly narrower memory bus (1024 bit vs 2048 bit), its die-thinned and probably also uses a different interposer. We know nothing of its thermal properties. I would abstain from drawing conclusions from its larger brethren. Not to mention there is a precedent: AMD and Apple already managed to have a Polaris chip that consumes half the power of the desktop version while running more CUs and offering competitive performance.

please share these links of the test comparison, i would like to see them. from what i read from my own is that i7 and i9 are not much difference in performance underload since i9 gets throttled after a certain period and performance same as the i7.

 

[leman]

Are you sure it runs on 45W TDP Power Limit with 3.2 GHz?

Im asking this seriously, because it is interesting from the perspective of desktop chips. I would actually love to test this on proper motherboard, and try running it passively. What workloads you have that you can test this?

Thats what Power Gadget says at least. I tried different things, from simple memory subsystem overload (piping yes or zero to null) to running multi-core Cinebench or statistical simulations. The results are always the same. It plateaus at around 45Watts, give or take, while reaching approx 3.2-3.3 Ghz and making out at 99C.

When the GPU is doing work at the same time, the CPU clocks obviously drop. In some cases, they can drop below the nominal (I've seen it run at 2.7 Ghz instead of nominal 2.9) — but maxing out both the CPU and GPU at the same time is a rather unrealistic usage scenario and certainly not something this laptop is designed to do.

 

[cube]

They're a bit hard to find, most reviews seem to be about games. As a rough order of magnitude of performance potential, you could expect the 2070 to be 10-20% faster for games, but the Vega 64 to be ~60-70% faster for compute. How much of that translates into real world performance is hard to say, and it's not even super clear how a reviewer would make an exact apples to apples comparison with currently available software. For anything more exact I think you'll have to say what you expect to use it for, because the relative performance will differ depending on use case.

For thermals, I have no first-hand experience with the Vega 64, but I have a Vega 56 in my box, and when tuned it's actually quite good. I can't give exact numbers, but it runs cool and reasonably quiet under heavy load. Certainly cooler than my MBP 2018 🙂

I saw one game benchmarked, and the Vega 64, 1080, and 2070 performance was about the same. I have no idea abot the card models and whether Vega was running too loud or hot.

 

[koyoot]

given amd's history of rehashing 290 - 390 i wouldn't hold my breath in "completely brand new design. unless your talking about NAVI thn yes i would definitely agree

Are you actually saying this because you know, or because you have an opinion?

Have you heard anything about: Uneducated opinion?

How can 4096 GCN Core GPU, with 2048 bit memory bus, 8 GB of HBM2 memory, be the same chip as 1280 GCN core chip, with 1024 bit memory bus, and 4 GB of HBM2?

And that is DESIGNED chip. It is different design, different chip. Completely.

 

[leman]

please share these links of the test comparison, i would like to see them. from what i read from my own is that i7 and i9 are not much difference in performance underload since i9 gets throttled after a certain period and performance same as the i7.

It's not that the i9 gets throttled. Since they are essentially the same chip, both the i9 and i7 tend to reach the same plateau. Intel prevented it in previous CPUs by limiting their clock wen running heavy multi-threaded workouts, but it seems that the limit in mobile Coffee Lake is set up very high (if it even exists). Since the i9 is higher binned, it tends to perform slightly better on average, but this all again depends on your luck in the silicon lottery. The point is that they perform better than what has been guaranteed by the manufacturer, so in any case you are getting what you have been promised. The difference is that with the base model, you are probably getting substantially more of that 🙂

I guess that all this confusion is because people who bough the i9 expect it to be much faster overall and get upset that their more expensive investment doesn't give them much of an edge in some situations. The entire topic is much more about marketing and customer expectation than about technology. And since people have difficulty understanding that all these clocks and TDP and whatever are all marketing concepts, we have this rather unhealthy discussion as the result.

My point is that with the i9, you are paying for a higher chance to get a slightly faster CPU overall (but you can also get lucky with the i7), more cache, as well as higher single-threaded boost clocks.
[doublepost=1542138242][/doublepost]

How can 4096 GCN Core GPU, with 2048 bit memory bus, 8 GB of HBM2 memory, be the same chip as 1280 GCN core chip, with 1024 bit memory bus, and 4 GB of HBM2?

Well, wait, I understand what you are saying, but its also not like this chip has been completely designed new from ground up. It uses the building blocks of the Vega design, just arranged a bit differently. And some components, like the interposer are likely to be entirely new design, sure, but the processors, memory controller, and the entire bus design is pretty much the same, just scaled down.

 

[CodeJoy]

I saw one game benchmarked, and the Vega 64, 1080, and 2070 performance was about the same. I have no idea abot the card models and whether Vega was running too loud or hot.

Yeah. You'll find some benchmarks where the Vega 64 is ahead, some where it's behind, and some where it's the same. Generally, my own rule of thumb is that if performance is within 15% then you probably can't perceive the difference anyway, so I tend to consider them "the same" for practical purposes.

What's a bit interesting is that I think Vega 56/64 have been doing better in some modern titles, relative to 1080.

 

[koyoot]

Well, wait, I understand what you are saying, but its also not like this chip has been completely designed new from ground up. It uses the building blocks of the Vega design, just arranged a bit differently. And some components, like the interposer are likely to be entirely new design, sure, but the processors, memory controller, and the entire bus design is pretty much the same, just scaled down.

No. GCN is ISA. Vega, Polaris, etc are just implementations of the ISA in the GPU. THEN you have the design. We have talked about this, in the Vega thread.

I will put it in another way. Is Polaris 10 and 11 the same design?

No.

Vega 10(4096 cores, 2048 Bit memory bus, 8 GB HBM2) is not the same design as Vega 12(1280 GCN cores, 1024 bit memory bus, 4 GB HBM2).

 

[leman]

What's a bit interesting is that I think Vega 56/64 have been doing better in some modern titles, relative to 1080.

I assume that's titles that either use DX12 or Vulcan. The secret is game-specific optimisations in the driver (and the overall driver quality). DX12/Vulcan/Metal provide computational models that are much closer to how hardware operates compared to earlier APIs. IHVs have been "cheating" their way around the inefficiencies of the older APIs by optimising their drivers for specific cases. And its something Nvidia has always been very good at. With modern APIs, this is less of an issue, and the gap between Nvidia and AMD has shrunk.

 

[cube]

Yeah. You'll find some benchmarks where the Vega 64 is ahead, some where it's behind, and some where it's the same. Generally, my own rule of thumb is that if performance is within 15% then you probably can't perceive the difference anyway, so I tend to consider them "the same" for practical purposes.

What's a bit interesting is that I think Vega 56/64 have been doing better in some modern titles, relative to 1080.

I would consider 5% on average to be negligible.

But it has to be at the same price and without thermal or acoustic excess.

In any case, the big draws of AMD's consumer cards are compute performance and FreeSync.

 

[leman]

Vega 10(4096 cores, 2048 Bit memory bus, 8 GB HBM2) is not the same design as Vega 12(1280 GCN cores, 1024 bit memory bus, 4 GB HBM2).

It's not the same design but they use same modules. The ALUs, the memory controller, the cache - they will be basically "copy-pasted" parts between different designs. Maybe with some incremental changes and improvements. But they certainly didn't come up with completely new circuitry to implement the ISA. That wouldn't make any sense commercially. The strength of Vega is its modular, scalable design — its like a box fo Lego parts that can be assembled together in different configs with relative ease. Thats why AMD can use them for dGPUs and integrated GPUs at the same time.

 

[CodeJoy]

Well, wait, I understand what you are saying, but its also not like this chip has been completely designed new from ground up. It uses the building blocks of the Vega design, just arranged a bit differently. And some components, like the interposer are likely to be entirely new design, sure, but the processors, memory controller, and the entire bus design is pretty much the same, just scaled down.

Well, arranged differently does mean a different design, no? Same architecture, different design.

 

[leman]

Well, arranged differently does mean a different design, no? Same architecture, different design.

Yeah, I think that's what koyoot's point is. I can't escape the impression that 90% of the time people at these forums are just arguing semantics anyway 😀

 

[koyoot]

It's not the same design but they use same modules. The ALUs, the memory controller, the cache - they will be basically "copy-pasted" parts between different designs. Maybe with some incremental changes and improvements. But they certainly didn't come up with completely new circuitry to implement the ISA. That wouldn't make any sense commercially. The strength of Vega is its modular, scalable design — its like a box fo Lego parts that can be assembled together in different configs with relative ease. Thats why AMD can use them for dGPUs and integrated GPUs at the same time.

When you design new implementation of an ISA, you usually design different circuit design...

Voltage curve can be different, that is the first thing...

 

[CodeJoy]

I would consider 5% on average to be negligible.

But it has to be at the same price and without thermal or acoustic excess.

In any case, the big draws of AMD's consumer cards are compute performance and FreeSync.

Yes, and also simpler for hackintosh builds. When I got mine, it was also cheaper than Nvidia's equivalent cards, and I've found thermal and noise to be on par with Nvidia when tuned (not stock though).

For gaming only, I don't think either Vega or RTX are the sweet spot for value.
[doublepost=1542139742][/doublepost]

Yeah, I think that's what koyoot's point is. I can't escape the impression that 90% of the time people at these forums are just arguing semantics anyway 😀

It's probably the case that in general, most people here do. But in this case I think koyoot has a valid point, if I've understood it correctly. Point is that you can't draw too many conclusions on Vega mobile based on performance of the Vega desktop chips. They can have different power consumption profiles, and will clearly run at different clocks anyway, and with different amounts of memory, and therefore also with a different memory bus and bandwidth. From what I've seen so far, I suspect that the chip design is actually quite good. Then what Apple does with it is a different story, we'll know soon enough.

 

[ilikewhey]

Are you actually saying this because you know, or because you have an opinion?

Have you heard anything about: Uneducated opinion?

How can 4096 GCN Core GPU, with 2048 bit memory bus, 8 GB of HBM2 memory, be the same chip as 1280 GCN core chip, with 1024 bit memory bus, and 4 GB of HBM2?

And that is DESIGNED chip. It is different design, different chip. Completely.

easy there, no need to attack others, since all of this is speculation unless you are in the industry, we all are forming uneducated opinions

difference is your uneducated opinion is optimistic

 

[CodeJoy]

easy there, no need to attack others, since all of this is speculation unless you are in the industry, we all are forming uneducated opinions

Flaming aside, the chip design is necessarily different, and there's no need to be in the industry to know that. It having 4GB HBM2 is official information. Given the way GPUs work, this immediately implies half the memory bus size, and automatically lower memory bandwidth. We also know officially that there are fewer compute units. In theory, it could just have been a regular desktop chip with stuff disabled, but we know that's not the case since we've seen the die, and it's clearly smaller. So it's not the same design, it's different. There is no speculation involved, just logical deduction from official sources.

 

[dan9700]

apple on the phone said its not out tomorrow end of November

 

[koyoot]

easy there, no need to attack others, since all of this is speculation unless you are in the industry, we all are forming uneducated opinions

difference is your uneducated opinion is optimistic

What I have written, about Vega GPU is not an opinion. I am not speculating about this chip. That is the difference.

 

[cube]

No wonder the RTX 4000 FP64 specs are hidden. Still a lot less than Vega 64.

 

[dan9700]

not dropping tomorrow now, end of November

 

[maflynn]

not dropping tomorrow now, end of November

In all fairness to apple, their website does state late November. I don't think November 14th qualifies as late November