AMD Announces a Red October: Zen 3 on October 8, RDNA2 on October 28

[Nkd]

I doubt it. Intel's IPC is very far ahead, and memory-latency is always going to be hampered given AMD's architecture. With Intel's Icelake improving IPC (and finally moving to 10nm), I think Intel is moving faster than AMD can catch up.

AMD has the advantage of a large L3 cache. I think as long as AMD is "close enough" to Intel, the extra cores make more sense for the general purpose customer (since those extra cores mean big wins in Server, 3d modeling, video editing, and other harder compute problems).

I don't think the AMD chip needs to be the best at everything, it should be the best at what it is best at, and just "keep pace" with the Intel chips with regards to single-threaded performance.



The XBox Series X / PS5 marketing slides have been pretty nice for RDNA2 so far. The question is how much of that translates to the desktop graphics cards. Raytracing + VRS seem to be big features worth hyping (even if NVidia had them on Turing already).

Are you speaking of this on another planet? how far is that? Intel is moving far ahead of AMD? look I respect intel but I seriously have no idea what you are talking about. Yes intel is faster in gaming so far but Zen 3 is suppose to address that and improve latency even further with each core having full access to cache. let me know when they have 10nm actually going. AMD will be on 5nm by then I believe. Intel talks more about AMD then their own products. When you can have your competition advertise for you, that says alot about your product.

 

[MAXLD]

RDNA 2 is expected to be a significant improvement if I take AMD to their words, but announcing 1.5 months after Ampere means they are going to miss the upgrade train and therefore, lost sales to Nvidia. Moreover announcement is 1 part, the second part is when it will be available for sale. By end Oct, they are already in the start of the holiday season. Assuming they take another 2 to 4 weeks for the product to be widely available, they are deep in the holiday season which would cost them more sales.

With the Ampere limited supply at launch and prices skyrocketing, that "upgrade train" will derail rather quickly. As soon as the first waves are depleted, prices will be pretty pathetic for a while. So who's really that eager to "upgrade" to Ampere immediately after launch?
- nVidia fans that wouldn't buy AMD anyway
- overhyped people that aren't even waiting for Ampere reviews (much less RDNA2)
- (early) buyers that always get the most expensive alternative anyway without second thought (usually still top end nVidia buyers)
- the ones that already sold their massively devalued 2080Ti 's, and are now desperate for a new card (tough luck...)
- people that urgently will need a new GPU by those weeks, because theirs just broke (which just happens during the entire year, so cases are limited)


All the rest, the minimally patient buyer, is waiting for one or more of the following:
- Ampere reviews (both vanilla and several custom partner models) (regardless of being neutral or exclusively nvidia buyer) (and some might not even be convinced by stuff like TDP, etc...)
- Ampere prices and stocks to stabilize (and in the meanwhile check reviews and decide what are the best models to aim for)
- RDNA2 reviews (both vanilla and several custom partner models) (to see if it's a better choice than Ampere)
- RDNA2 prices and stocks to stabilize (and in the meanwhile check reviews and decide what are the best models to aim for)

In conclusion: most people will want to wait and see how all the GPUs perform, check all the cards, make a better level-headed choice of specific models, and also get a "non price inflated" custom partner card, regardless of being Ampere or RDNA2. That means: they won't even bother with the holiday season, they're buying well inside Q1 2021.... so AMD announcing/launching RDNA2 several weeks after nVidia in October2020 makes zero difference. It would make a LOT of difference it AMD had nothing to release for another 6 months. But in this case in particular? Not really. They only need to have a good line up... if it performs well, it will sell well through 2021... because that's what most people do: get stuff that is competitive and a good deal for what they need. If RDNA2 ends ups sucking in reviews, then at least they'll be even more comfortable in buying an Ampere at a fair price in 2021 and laugh at all the jump starters.

 

[watzupken]

...and wondering if they are still tweaking clocks up a bit....

This is what I think they are doing. We can look at it in 2 ways, (1) to make it even better than Nvidia's Ampere, or (2) trying to match Nvidia's Ampere. I feel it will be the latter rather than the former purely based on the fact that they have mentioned nothing about the performance other than a 1.5 months later announcement.

 

[Mussels]

If AMD can beat the 2080ti (best of last gen) and compete on modern prices, they're still doing really well. Very few people buy the fastest card of each gen, look how popular the 5700XT is just for its value for money... if it wasnt for the bugs people had with them, they'd have been everywhere.

 

[watzupken]

With the Ampere limited supply at launch and prices skyrocketing, that "upgrade train" will derail rather quickly. As soon as the first waves are depleted, prices will be pretty pathetic for a while. So who's really that eager to "upgrade" to Ampere immediately after launch?
- nVidia fans that wouldn't buy AMD anyway
- overhyped people that aren't even waiting for Ampere reviews (much less RDNA2)
- (early) buyers that always get the most expensive alternative anyway without second thought (usually still top end nVidia buyers)
- the ones that already sold their massively devalued 2080Ti 's, and are now desperate for a new card (tough luck...)
- people that urgently will need a new GPU by those weeks, because theirs just broke (which just happens during the entire year, so cases are limited)


All the rest, the minimally patient buyer, is waiting for one or more of the following:
- Ampere reviews (both vanilla and several custom partner models) (regardless of being neutral or exclusively nvidia buyer) (and some might not even be convinced by stuff like TDP, etc...)
- Ampere prices and stocks to stabilize (and in the meanwhile check reviews and decide what are the best models to aim for)
- RDNA2 reviews (both vanilla and several custom partner models) (to see if it's a better choice than Ampere)
- RDNA2 prices and stocks to stabilize (and in the meanwhile check reviews and decide what are the best models to aim for)

In conclusion: most people will want to wait and see how all the GPUs perform, check all the cards, make a better level-headed choice of specific models, and also get a "non price inflated" custom partner card, regardless of being Ampere or RDNA2. That means: they won't even bother with the holiday season, they're buying well inside Q1 2021.... so AMD announcing/launching RDNA2 several weeks after nVidia in October2020 makes zero difference. It would make a LOT of difference it AMD had nothing to release for another 6 months. But in this case in particular? Not really. They only need to have a good line up... if it performs well, it will sell well through 2021... because that's what most people do: get stuff that is competitive and a good deal for what they need. If RDNA2 ends ups sucking in reviews, then at least they'll be even more comfortable in buying an Ampere at a fair price in 2021 and laugh at all the jump starters.

I think this is nothing different from what we see in the past for any highly anticipated products. Demand will be very strong at the start, with supply trailing. Prices are less of a concern for those that want to get their hands on cutting edge technologies.

As for Nvidia fans, I don't know the population, but I reckon most people are brand agnostic. The only reason why they are flocking to Nvidia is because there are not strong product lineup at the high end. In the mid end, the RX 5700 and 5600 offers very good value, but marred by unstable drivers for quite a long time which is also driving people towards Nvidia since there is no other alternatives.

I agree that most people will wait out since they don't really upgrade as often until they see a good deal. But what AMD will be losing out is the initial few batches of people upgrading not just from a monetary perspective, but also from a market share perspective. Missing half the holiday seasons will also mean losing a sizeable chunk of sales. That is the reason most companies release new products by Sep.

 

[R0H1T]

As for Nvidia fans, I don't know the population, but I reckon most people are brand agnostic.

Nah I'd argue a lot less people are brand agnostic than they were even in the recent past, perhaps it's just me but people tend to be more tribal/extreme (in their views) these days even for things such as PC hardware. Yes that also includes me.

 

[MAXLD]

I agree that most people will wait out since they don't really upgrade as often until they see a good deal. But what AMD will be losing out is the initial few batches of people upgrading not just from a monetary perspective, but also from a market share perspective. Missing half the holiday seasons will also mean losing a sizeable chunk of sales. That is the reason most companies release new products by Sep.

Sure, but nVidia has 70~80% of the discrete GPU market. AMD is not reversing that in a month or so in 2020 right after a launch period of both Ampere and RDNA2, they'll have to do it progressively for a year and beyond, just like they did with ZEN2. And they will have to do it with more mid-range"-ish" cards around the $200~400 bracket. Top GPUs have more profit per card, but there's not a lot of ppl wasting more than a month's salary on a single card. AMD doesn't necessarily need $700~1000 cards to get market share, but they need them for rebuilding brand image over time and get some better profit margins.

 

[Deleted member 185088]

I doubt it. Intel's IPC is very far ahead, and memory-latency is always going to be hampered given AMD's architecture. With Intel's Icelake improving IPC (and finally moving to 10nm), I think Intel is moving faster than AMD can catch up.

AMD has the advantage of a large L3 cache. I think as long as AMD is "close enough" to Intel, the extra cores make more sense for the general purpose customer (since those extra cores mean big wins in Server, 3d modeling, video editing, and other harder compute problems).

I don't think the AMD chip needs to be the best at everything, it should be the best at what it is best at, and just "keep pace" with the Intel chips with regards to single-threaded performance.

Not really AMD is holding IPC crown and SMT is superior to HT for that matter, only in games like you stated AMD is still a bit behind, the latency of Zen2 is worse than Zen+ but still it's way faster, I think Intel run their CPUs at much higher frequencies AMD doesn't, and games are more optimised for Intel than AMD, would be interesting to see clock for clock who's faster.

 

[AsRock]

AMD late to the GPU party once again...

yeah and it's going be another month so Novemeber, not to sure if it's worth waiting tbh.

They need to get some numbers "COUGH" honest ones too.

well now that the date is set. AIBs will likely be getting more info. They have been super tight this time around. I am sure we will start getting leaks later this month or early October. Or even by AMD themselves secretly lol.

I want facts not just the typical BS they been throwing around.

You make some good points. I think you might be onto something about them waiting to see what real world performance looks like and adjusting price accordingly. AMD would be silly to make business decisions solely based on claims made by their competitor. Or, conversely, it could be the fact that they don't have anything that competitive. If they revealed non-competitive products reactively, there's nothing to stop people from buying Nvidia right now. However, waiting to reveal details regarding performance keeps people wondering what RDNA2 will offer and could therefore forego people from buying Nvidia without waiting for RDNA2. Although then you could argue that all they're doing is delaying the inevitable...

No not in all cases, if AMD stay silent fuck it i will most likely get a nVidia card. Although still thinking if i should bother still at all but leaning more of just picking up a 3070 and have done with it.

 

[BoboOOZ]

It's as if they are ashamed of RDNA2...

Well, it's not their most successful product. Less priority than zen and console APU.

 

[Trovaricon]

Intel absolutely has faster PDEP, PEXT, DIV, and IDIV instructions. I know, because I've measured them personally. In the case of PDEP / PEXT, 1000% faster or more (single-clock PDEP / PEXT on Intel, but over 15 clocks on AMD).

AMD and Intel both have single-clock add / subtract. AMD had the first dual-AES pipeline (supporting two concurrent AES instructions), so AMD's AES IPC was superior to Intel's AES IPC on Skylake.

What instructions, what conditions, what memory? Its actually very complicated. If we're sitting around adding/subtracting numbers inside of registers all day, both processors tie. MAC instructions, I forget from the top of my head... but you get the picture. Its pretty complicated, because all these CPUs have different speeds for every instruction.

DIV/IDIV
Context!
32bit operands? Sure, rock-solid throughput of 6 cycles for consecutive instructions > 1/6 DIV per cycle (xxx Lake uarch).
What about 64bit operands? Well, sides reverse don't they? (I assume) Micro-coded calculation, variable latency & variable throughput, twice slower than Zen2. What is even more intriguing, slower than Goldmond Plus uarch which does not spot fancy fixed 6cycle reciprocal throughput for operands shorter than 64bits!

PDEP/PEXT
GPR scatter-gather niche? Sure Intel created those instructions (BMI2), if they would be "micro-coded" on their flagship uarch, then I don't see a reason for even making them.

Differences between capabilities of ALUs of Intel and AMD architectures (not counting front-end, cache & memory) are much bigger than many "IPC" discussions admit.
What about POPCNT/LZCNT/TZCNT? (number of bits set to 1, number of leading zeroes, number of trailing zeroes - simple explanation for those, who are "not in the know")
Latency in Lake uarch is 3, AMD 1 (or 2 for TZ). That is not all. On Intel only a single port is capable of handling these instructions (throughput is 1 instruction per cycle). On AMD? All four ALUs are able to execute POP and LZ instructions (4 "IPC") and 2 TZ instructions & 2 "IPC".
There is one more implication based on the differences of Intel "Super-ALUs" vs. AMD clustered ALUs and separate GPR & SIMD "engines". By executing these three instructions you block port1 which is really an abstraction of a "super ALU" (take a look at https://en.wikichip.org/wiki/intel/microarchitectures/coffee_lake what else it is supposed to handle).
AMD design philosophy of a separate "vector engine" means, that they do not block each other (see https://en.wikichip.org/wiki/amd/microarchitectures/zen_2 ).
Which I suppose is the reason why AMD sees higher gains when SMT is enabled as there is no way to utilize such wide engine in a single thread / one non-artificial instruction stream.

[/END of OT]

I am looking forward to Zen3 mostly because of unified L3 cache. At least in Azure, Zen2 CCX are exposed as NUMA nodes (starting with 16 "virtual core"/thread VMs) to let those not intimate with the details know, that spreading threads across all cores will randomly end up with cache misses even if you expect, from the code perspective, that it "should" be present in the cache.
Cache-sensitive workloads see actually higher perf degradation (vs theoretical *2) on Zen2 than on Cascade Lake when you double the size of VMs (from 4c/8t to 8c/16t).

 

[sutyi]

AMD late to the GPU party once again...

nVIDIA launching two highend SKUs with probably very limited availability next week, making it almost a paper launch in case not 90% of the stock is bought up by miners.
AMD RDNA2 launch event in 7 weeks or a month and half, oh the horror.

 

[olymind1]

AMD late to the GPU party once again...

Maybe, maybe not, i for one don't buy hardware at day one. Few years ago i bought my old RX480 3-4 months after release and probably that was my fastest hw purchase, i got my Ryzen 2600 1 year after its release.

Still i'm wating for Zen3, but i'll most likely wait for it till Zen4 releases. Also with the increased mid range prices over the years, i'll wait a lot more with RDNA2 or with nv's Ampere.

 

[Lionheart]

I doubt it. Intel's IPC is very far ahead, and memory-latency is always going to be hampered given AMD's architecture. With Intel's Icelake improving IPC (and finally moving to 10nm), I think Intel is moving faster than AMD can catch up.

AMD has the advantage of a large L3 cache. I think as long as AMD is "close enough" to Intel, the extra cores make more sense for the general purpose customer (since those extra cores mean big wins in Server, 3d modeling, video editing, and other harder compute problems).

I don't think the AMD chip needs to be the best at everything, it should be the best at what it is best at, and just "keep pace" with the Intel chips with regards to single-threaded performance.



The XBox Series X / PS5 marketing slides have been pretty nice for RDNA2 so far. The question is how much of that translates to the desktop graphics cards. Raytracing + VRS seem to be big features worth hyping (even if NVidia had them on Turing already).

What are you smoking.

 

[ShurikN]

AMD RDNA2 launch event in 7 weeks or a month and half, oh the horror.

From what was said in the teaser, it looks more like an announcement event and not a launch. I'd be surprised if it launches in October.
For AMD every day matters considering people are already heavily hyped for Ampere.

 

[Th3pwn3r]

I love how AMD has been very quiet, it most likely signals something great down the line. I really wanted a juicy tidbit to bite on in the meanwhile though.

I think it's stupid. A lot of people will already have purchased 3070s,80s and 90s. Give people something to wait for and they will. If they release information showing what they have a lot of customers that would buy Nvidia will go with AMD instead. Hell, I might wait either way but who knows. The sooner they give us details the better in my opinion.

 

[sutyi]

From what was said in the teaser, it looks more like an announcement event and not a launch. I'd be surprised if it launches in October.
For AMD every day matters considering people are already heavily hyped for Ampere.

RDNA2 consumer cards were said to be coming before consoles. New XBox variants are launching November 10th.
So they either launch in that two weak time frame (Oct. 28.-Nov. 01.) with at least one SKU being available in retail / etail or there will be a lot of upset people and the proverbial ball had been dropped again.

Currently they are not a year but only 1-2 months behind big green in product cycle and that is not as bad as you make it out to be. But honestly they had lost tremendous mindshare in the past years with overhyped releases that "failed to deliver" in one way or another and then came the driver debacle with NAVI cards that were amplified by the release of the new 2020 software suite.

Honestly I wouldn't even mind if they would release only in December but have a rock stable driver stack / software suite with all the bells and whistles working as they should for RDNA2. Cause in the end no matter how long those bars in the charts are, when people question their own decisions to buy the product due to software issues.

 

[kapone32]

I'm really hoping the new Radeons will be good. But tbh I dont expect it. AMDs GPUs have not been competitive for along time now and I dont expect them to change now.
They always made a big show and somehow could not delive in the end, first they just got too hot (Vega), then they were just too expansive (VII) or they fucked up driver support (Navi). Its time to bring something that is just good!
Dont get me wrong I am not some kind of fanboy I always buy what I belive gets me the best value for money (I just got a Ryzen 3900x this year) but AMD GPUs did not seem to give me that for a long time..

No love for Polaris?

 

[NeuralNexus]

Intel has higher clocks, but based on what I'm seeing for Icelake, Intel is taking a big step forward with IPC again.

Remember that Skylake only had a 32kB L1 data-cache and 256kB L2 cache. Zen2 clobbers that with a 64kB L1 dCache and 512kB L2 cache.

Icelake comes back with 64kB L1 cache and 512kB L2 cache. But not only that, its improving the ROB (register renaming buffers or something like that), and other bits that make Icelake look much wider and probably much better from an IPC perspective.

I don't think Zen2 was ever "clearly better" than Skylake with regards to IPC. It was close, but not necessarily an IPC king. Icelake has a bunch of core-enhancements that should improve IPC dramatically. The question is what clocks can Intel get from their 10nm node. Intel's 14nm node clearly is cost-effective and has very high clocks, so the 10nm IceLake may lose a bit of clock speed.

Something is serious wrong with you lol

 

[Legacy-ZA]

I think it's stupid. A lot of people will already have purchased 3070s,80s and 90s. Give people something to wait for and they will. If they release information showing what they have a lot of customers that would buy Nvidia will go with AMD instead. Hell, I might wait either way but who knows. The sooner they give us details the better in my opinion.

While I agree, I think AMD is trying to stick it to nVidia, like they did with Intel, make them release their over-inflated prices and hit them hard and fast. Anyways, time will tell, I hope I am right.

 

[EarthDog]

While I agree, I think AMD is trying to stick it to nVidia, like they did with Intel, make them release their over-inflated prices and hit them hard and fast. Anyways, time will tell, I hope I am right.

We all do, but it doesn't seem like it will happen. AMD will be lucky to beat the 3080 on non RT/DLSS loads if things pan out...

 

[dragontamer5788]

PDEP/PEXT
GPR scatter-gather niche? Sure Intel created those instructions (BMI2), if they would be "micro-coded" on their flagship uarch, then I don't see a reason for even making them.

Today it is niche, but in my experience, surprisingly flexible.

Spoiled for offtopic:

Spoiler

My introduction to PDEP / PEXT was chess sliding attacks. For this original problem, consider a chess row: [_ x R _ _ x _ x], where "R" is the location of a Rook, and x are locations of other pieces. Where can the rook move to? The output of sliding piece attack would be: [0 1 1 1 1 1 0 0] (inclusive, because the Rook may "capture" the [0 x . . . x 0 0] positions. A same-color check can be done later at very little cost). The PEXT solution to this problem was so elegant, so fast, that I knew I had to play with the instruction more.

Now instead of doing it row-by-row or column-by-column, do it across all 64 locations of a chessboard (which happens to fit on a 64-bit number, one bit per position).

U64 arrAttacks  [...]; // ~840 KByte all rook and bishop attacks
U64 arrRookMask  [64]; // 10..12 relevant occupancy bits per rook square
U64 arrBishopMask[64]; //  5.. 9 relevant occupancy bits per bishop square
U32 arrRookBase  [64]; // arrAttacks base offset per rook square
U32 arrBishopBase[64]; // arrAttacks base offset per bishop square

U64 rookAttack(U64 occ, enumSquare sq) {
  return arrAttacks[arrRookBase[sq] + _pext_u64(occ, arrRookMask[sq])];
}

U64 bishopAttack(U64 occ, enumSquare sq) {
  return arrAttacks[arrBishopBase[sq] + _pext_u64(occ, arrBishopMask[sq])];
}

"Occupancy" is the 64-bit "other pieces" bitboard (aka: the "x" pieces in my [_ x R _ _ x _ x] 8-bit example). arrRookMask[sq] selects the 14-locations a rook can attack (the vertical and horizontal movements, it doesn't make sense in the 8-bit row example but when you consider the full sized board its useful). The 64-bit number returned is all positions the Rook can attack. In just 3-table lookups and 1 pext.

I've played around with it for a weekend. I found other good uses of PEXT and PDEP: not just in bitfields but also Hashing (PDEP / PEXT can really "mix up the bits"). I really think that PDEP / PEXT are the most "creative" new instructions in recent memory.

Differences between capabilities of ALUs of Intel and AMD architectures (not counting front-end, cache & memory) are much bigger than many "IPC" discussions admit.
What about POPCNT/LZCNT/TZCNT? (number of bits set to 1, number of leading zeroes, number of trailing zeroes - simple explanation for those, who are "not in the know")
Latency in Lake uarch is 3, AMD 1 (or 2 for TZ). That is not all. On Intel only a single port is capable of handling these instructions (throughput is 1 instruction per cycle). On AMD? All four ALUs are able to execute POP and LZ instructions (4 "IPC") and 2 TZ instructions & 2 "IPC".
There is one more implication based on the differences of Intel "Super-ALUs" vs. AMD clustered ALUs and separate GPR & SIMD "engines". By executing these three instructions you block port1 which is really an abstraction of a "super ALU" (take a look at https://en.wikichip.org/wiki/intel/microarchitectures/coffee_lake what else it is supposed to handle).
AMD design philosophy of a separate "vector engine" means, that they do not block each other (see https://en.wikichip.org/wiki/amd/microarchitectures/zen_2 ).
Which I suppose is the reason why AMD sees higher gains when SMT is enabled as there is no way to utilize such wide engine in a single thread / one non-artificial instruction stream.

I'm betting you got it reversed by the way. TZCNT is equivalent to popcnt((x & -x) - 1), which means the TZCNT instruction can recycle the adder and negation circuits on a CPU. LZCNT on the other hand, cannot be emulated quickly by any standard CPU.

POPCNT: That's cool and I wasn't aware of this fact. POPCNT is one of my favorite instructions.

TZCNT As stated earlier: TZCNT is popcnt((x & (-x)) - 1), which is 400% slower than a dedicated instruction (but still very fast at 1x popcnt((x & (-x)) - 1) per clocktick bandwidth on Zen processors: 4-instructions on 4 pipelines, that can almost certainly be out-of-order scheduled pretty finely).

LZCNT: Okay, that's nifty. Can't be replicated quickly on any machine. I fortunately haven't had the need for LZCNT (Given the efficiency of TZCNT, I "favor" the TZCNT direction heavily whenever I'm playing with bitsets).

Fun fact: POPCNT, TZCNT, and LZCNT are efficiently executed on NVidia and AMD GPUs. So learning to use these instructions is great for GPU programming. And GPUs are surprisingly good at 32-bit integer bitset operations, thanks to single-cycle implementations of these instructions.

 

[Crustybeaver]

I love how AMD has been very quiet, it most likely signals something great down the line. I really wanted a juicy tidbit to bite on in the meanwhile though.

Explain the logic behind being quite correlates to greatness???

 

[TheoneandonlyMrK]

We all do, but it doesn't seem like it will happen. AMD will be lucky to beat the 3080 on non RT/DLSS loads if things pan out...

Nah the 3090'S in sight the 3080 will get ti'd into obscurity ,on shelves within six months of its launch. IMHO.
Ps5 comes first for me though

 

[EarthDog]

Nah the 3090'S in sight the 3080 will get ti'd into obscurity ,on shelves within six months of its launch. IMHO.
Ps5 comes first for me though

You keep saying that... and we hope you're right... but the magic 8 ball tells otherwise me thinks.