HW News - RAM & SSD Prices Falling, RTX 3090 Alleged Photos, TSMC Makes One Billion 7nm Chips
Posted on August 26, 2020
This week’s news was highlighted by Hot Chips 32 (2020), which brought plenty of information that we’ll disaggregate below. Another point of interest is AMD launching its budget A520 chipset, but we already have a video and article dedicated to that topic, so we'd point you to that for more information (it’s linked below). Also of importance is Nvidia’s second quarter earnings, DRAM and NAND prices seemingly on the decline, and Internet Explorer being put out to pasture.
01:04 | Medium Modmat Restock & Reforestation Campaign
We just concluded our newest charity campaign for Eden Reforestation Projects and have a donation receipt to share with you.
Last week on store.gamersnexus.net, we restocked the Large Modmats, which we’ve struggled to keep in stock ever since we launched them 3 years ago, the Mouse Mats, and the medium modmats. The large modmats sold completely through in under a day, which is absolutely insane, but we already have another run getting manufactured. In the meantime, we just got a big shipment in of GN Medium Modmats, which use the same high-quality anti-static, rubberized material and are made in the same clean room manufacturing facility we use for the large modmats. The medium mats have a video card PCB layout with a naming reference table, a grid for screw tracking and component disassembly, some to-scale hole spacing diagrams, and is highly portable at its 32” x 16” size, or 811mm by 406mm. The wireframe mousemat restock was our highest quantity order ever, and it’s already down to under 25% inventory remaining after one week. If you buy either soon and it’s marked as “add to cart,” then it’ll ship immediately. You can grab any of these items on store.gamersnexus.net.
After LinusTechTips’ troubleshooting challenge a few weeks ago, we had our own livestream where we announced we’d be donating again to Eden Reforestation Projects between August 8th and August 20th. We offered a portion of our store’s revenue for the period, plus up to $1500 matched to viewer donations for anyone who wanted to bypass us and donate straight to them, plus some additional funding that our distributor committed. Between our distributor’s generous contribution, the GN store, and the $1500 match, we’re happy to present this receipt for $7080 to Eden Reforestation Projects. If you’re not familiar with them, GN worked with the group last year and I’ve personally been supporting them for around 3 years now. The group employs impoverished people to reforest deforested land and does so in a scientific fashion, researching the correct mix of biology and investing in biodiversity in regions, as well as the people who live there. It creates lasting jobs as well, because the planters have to take care of and protect the reforested land. We like the work we do and are happy to make this contribution, and a huge thanks to our viewers, too!
Next week, we’ll list some stuff to benefit Jay’s chosen charity. Keep an eye on HW News next week for that!
04:20 | DRAM & NAND Prices Ready to Fall
Both DRAM and NAND prices look set to fall, according to reports from both Digitimes and DRAMeXchange. It seems memory prices are set to fall throughout the last half of 2020, and extend into at least the first quarter of 2021. As ever, the memory market ebbs and flows, so we’ll have to keep an eye on it.
As is usually the case, the price decline is based on a growing oversupply in the market. As it relates to NAND flash, the oversupply stems from buyers stockpiling memory amidst concerns of the economic effects of the pandemic. That excess inventory has been carried over month over month, as it seems demand remains relatively flat, both in the PC and server space.
“TrendForce believes that, despite the traditional peak season for electronics sales and the release of Apple’s new iPhones in 3Q20, the quarterly decline in NAND Flash ASP will likely reach 10%, due to the client end’s excess inventory under the impact of the pandemic,” according to DRAMeXchange.
The story is much the same with DRAM, as customers stockpiled memory for fear of the supply chain falling out over the course of the pandemic. While DRAM sales were up in general for 2Q 2021, it seems OEMs are still sitting on an excess of inventory. “TrendForce indicates that server OEMs are now carrying a rather high level of DRAM inventory after aggressively stocking up for two consecutive quarters. At the same time, customers of enterprise servers are holding back on procurement because the economic outlook is getting bleaker and more uncertain,” says DRAMeXchange.
The result, it seems, will be a decline in both DRAM shipments and prices, beginning in 3Q2020. “As such, TrendForce forecasts at best a flattening of product shipments and decrease in DRAM prices in 3Q20, with DRAM suppliers suffering a decline in profitability,” DRAMeXchange continues.
It’s tough to balance not wanting to report on rumors and people really wanting to hear it. This one is technically in rumor status. What we know for fact is that RTX 3090 is going to be a product, something we confirmed last week when GamersNexus spoke with board partners off-record and learned of the 3090’s existence. We also heard that the 3090 should launch after the 3080, but we’re not certain if that’s how it’ll ultimately play out. Further, the photos of the shroud design for the 3080 and 3090 appear to be accurate as well, and we also confirmed privately that the 12-pin power connector will be on reference cards, but not most AIB cards. Before getting into the new rumor, we’d also like to highlight that one 12-pin isn’t as simple as adding 6 and 6 together: The wire gauge may change, as may the pinout, and that matters more than the actual count of pins.
For the new rumors, twitter user GarnetSunset, who is competing with momomo_us for Top Leaker, tweeted that the 3090 will allegedly be a three-slot design. The PCB is also plus-sized, running taller than the PCIe slot. Another user in the thread, kman18981, took the time to scale the PCIe slots to match to get a better sense of scale. If this is, in fact, a real 3090, then it’s going to be a large card. There may be room yet for AIBs to compete in the two-slot department, if thermally possible.
10:37 | Hot Chips 32: Xbox Series X SoC Details, Cost is a Factor
At Hot Chips 32, Microsoft offered some more details on the AMD SoC powering its new console. The SoC for the Xbox Series X will be built on a 12-layer stack for the substrate, will consist of 15.3B transistors, and will have a die area of 360.4mm^2. The new SoC is also being built on TSMC’s nebulous N7 Enhanced node -- nebulous because we still don’t know what “enhanced” precisely means, relative to TSMC’s other N7 nodes. We know it’s not N7, but it’s possible it’s some variation of N7P or N7+. So far, everyone has kept surprisingly quiet about it.
Moving on, one of the more important things about this generation is the increased SoC cost. Looking back at the Xbox One X, that SoC was built on 16nm and packed around 6.6B transistors, meaning the transistor density has more than doubled for this generation. Furthermore, Microsoft is noting that N7E added complex steps and structures to an already large and complex heterogeneous SoC. That said, the cost per wafer has greatly increased, and the yield count per wafer has decreased, which means a significant increase in cost per die.
Microsoft seems to be trying to offset these costs by implementing in-hardware engines to help ease power and die cost. The HW Engines, as Microsoft calls them, seem to focus on audio, security, decoding, and decompression.
Additionally, the SoC will feature eight cores spread over two quad-core clusters, with the cores clocked at 3.8 GHz. On the GPU side, there will be 52 CUs in total (26 dual CUs), with four SIMDs and four ALUs. The GPU will come with support for Variable Rate Shading, Sampler Feedback Streaming, and DirectX Ray Tracing.
IBM took the wraps off its upcoming Power10 processor and architecture, which is the successor to the Power9 platform. If you recall, IBM’s Power9 chips are what power the Summit and Sierra supercomputers, which until the arrival of Fugaku, were the first and second most powerful supercomputers in the Top500 list.
Power10 will be based on the Power ISA (specifically Power ISA v3.1), but the overall P10 architecture is being overhauled for better memory latency and bandwidth, improved I/O, and a focus on performance per watt. The Power10 chips will be built on Samsung’s 7nm process, said to make use of an 18 layer metal stack and 18B transistors. The Power10 processors will come as either single-chip modules (SCM) or dual-chip modules (DCM).
Each Power10 chip is built with 16 cores, but will come with 15 enabled in an effort to stabilize yields. Each P10 core is either eight-way (SMT8) or four-way multi-threaded (SMT4), meaning that a single P10 chip is capable of 120 threads per socket (SMT8); a dual-module chip doubles that to 240 threads per socket (SMT8).
The reworked memory system for P10 supports both the Open Memory Interface (OMI) and PowerAxon interface, with both interfaces capable of 1TB/sec. Power10 will also support DDR5 and PCIe 5.0, with support for 64 PCIe Gen5 lanes. Clock speeds look to be in the range of 3.5 GHz to over 4.0 GHz.
Systems based on Power10 are expected sometime towards the end of 2021.
Last week, Taiwan Semiconductor announced that it had fabricated its one-billionth defect-free 7nm die. The company began volume production in April of 2018, and includes Apple, AMD, Huawei, car companies, and now Intel as customers of its foundry. For fun, TSMC calculated that its one billion 7nm chips would be equal to 13 Manhattan city blocks, although TSMC didn’t say if it smells the same as 13 Manhattan city blocks.
TSMC highlighted that its yields have improved with practice, but says that it’s still not easy to make chips on 7nm process. As a reminder, Intel’s version of 7nm and TSMC’s version aren’t equivalent, and a lower number isn’t inherently better alone; that said, TSMC’s 7nm does exist, and Intel’s 10 barely does, so that point doesn’t really matter anymore. TSMC is in the lead, period.
TSMC had this to say of its accomplishment:
“We have extended our 7nm technology in to a new family member, our N6 process. N6 is in volume production today, using EUV to replace conventional immersion layers. TSMC’s N6 offers a new standard cell with nearly 20% logic density improvement. Its design rules are completely compatible with its N7 predecessor, and delivers an excellent cost-effective option for our customers’ next wave of 7nm generation designs.”
Nvidia reported its Q2 2021 earnings this past week, and it seems the company has notched a record quarter. Most notably, Nvidia’s datacenter segment earned more than its gaming segment for the first time in history. We should note that this is 2 years after RTX, so that could shift come September.
In total, Nvidia is reporting a record revenue of $3.87B, which is a 50% increase YoY, and roughly 26% QoQ. A couple of key things have happened in Q2 that have driven Nvidia’s revenue: The closing of the Mellanox acquisition, and ramping A100 accelerator sales. Nvidia closed the Mellanox deal back in April, and that added no less than 14% to Nvidia’s revenue, all of which is accounted for in the datacenter. Additionally, Nvidia’s A100 accelerators had the full run of Q2 to drive sales, unlike the previous quarter.
By segment, Nvidia’s datacenter segment saw $1.75B in revenue, up 54% QoQ and 167% YoY. Again, that’s thanks to Mellanox and Nvidia’s continued sales of A100 systems. For gaming, Nvidia recorded $1.65B in revenue, up 24% QoQ and 26% YoY. Nvidia’s gaming segment revenue was fueled by a spike in demand for gaming, GPUs, and laptops -- primarily thanks to the pandemic that has forced everyone to work, learn, and play from home the last several months.
Shoring up Nvidia’s revenue was also its professional visualization and automotive markets, which brought in $203M and $111M, respectively. Looking ahead to Q3 2021, Nvidia is forecasting a $4.4B quarter. Finally, Q3 should also see the arrival of new Ampere GeForce cards, but that likely won’t be fully reflected in Nvidia’s earnings until Q4.
19:32 | Microsoft Begins Phasing Out Internet Explorer
Microsoft's Internet Explorer and legacy Edge browser (non-Chromium) are getting ever closer to making that trip to the great browser window in the sky. While IE11 isn’t going away entirely, Microsoft is taking big steps in what appears to be the beginning of the death knell for the browser. And with Microsoft’s Chromium-based version of Edge, there’s no need for legacy Edge at this point.
Microsoft announced that after August 17, 2021, IE11 will no longer be supported by any Microsoft Office 365 app, which is significant. Microsoft Teams support ends even sooner, with November 30, 2020, marking the end of support for IE11. As Microsoft notes, some enterprises have made specific application investments in IE11 that require that browser. To that end, Microsoft states that IE11 isn’t going away entirely, and that Edge’s Internet Explore Mode will offer legacy support as well.
As for the non-Chromium build of Edge, it’s done. Love it or hate it, Microsoft’s Chromium-based Edge has been a success, and there’s no turning back. Legacy Edge will sunset after March 9, 2021.
Meanwhile, IE6 remains in safe refuge at college campuses and government offices across America.
21:14 | Intel Defends AVX-512, Frequency Improving
AVX-512 seems to keep coming up in the news cycle lately, mostly in a non-flattering way. First, Linux creator Linus Torvalds wished it a painful death and criticized the frequency penalty for AVX-512 workloads, relating it to a power virus. Then, former Intel Chief Engineer Francois Piednoel echoed Torvalds’ sentiment to an extent, calling AVX-512 a mistake.
However, Intel is here to take up for its beloved instruction set, defending it against such critics. The defense comes from Intel’s current Chief Architect, Raja Koduri. “AVX-512 is a great feature. Our HPC community, AI community, love it. Our customers on the data center side really, really, really love it,” Koduri told PCWorld.
There’s been a frequency and performance penalty for executing AVX-512 instructions, as evidenced by the downclocking on Skylake-X and Skylake-SP CPUs when running an AVX-512 workload. The wider vector instruction set has also been criticised for taking up too much transistor and die budget for chips that don’t really need them; i.e., mobile chips and generalist Xeons.
Koduri addressed this by stating that the AVX-512 frequencies are improving generation-over-generation. That’s something that Travis Downs, who runs the Performance Matters blog, was able to substantiate. Downs tested an Ice Lake i5-1035 CPU for AVX-512 downclocking, and he summarized his findings as follows:
“The Ice Lake i5-1035 CPU exhibits only 100 MHz of licence-based downclock with 1 active core when running 512-bit instructions.
There is no downclock in any other scenario.
The all-core 512-bit turbo frequency of 3.3 GHz is 89% of the maximum single-core scalar frequency of 3.7 GHz, so within power and thermal limits this chip has a very “flat” frequency profile.
Unlike SKX, this Ice Lake chip does not distinguish between “light” and “heavy” instructions for frequency scaling purposes: FMA operations behave the same as lighter operations.”
On the topic of certain chips not needing AVX-512 support, Kodrui stated that Intel has to develop for the x86 ecosystem, which spans everything from laptops to servers.
“So when we do a CPU core and add an instruction to it, historically the power of x86 and our instruction set extensions have been that we made them available everywhere. Because of that, when we have an IP like Sunny Cove and it appears both in a server like an Ice Lake server and on a client, like an Ice Lake client, you get the commonality of the instruction set,” Koduri told PCWorld.
So, as we’ve said previously, despite its detractors, AVX-512 isn't going anywhere yet.
In a report from OC3D, it seems there’s a shortage of PWM doublers and that shortage is expected to force motherboard OEMs to redesign some of their boards. In fact, according to OC3D, Gigabyte has already made moves to revise two of its more budget oriented B550 motherboards.
PWM phase doublers are prevalent across all motherboard price segments, but the impact of the doubler shortage is likely to be concentrated in the budget segment. According to an OC3D exclusive, AMD’s B550 chipset-based boards are expected to be hit the hardest. OC3D is exclusively reporting that Gigabyte has designed two new SKUs due to the shortage: The B550 AORUS Elite V2 and the B550 AORUS Elite AX V2. Both SKUs carry the “V2” designation to mark the absence of PWM phase doublers. Other manufacturers are expected to redesign boards around the shortage, but it’s presently unclear how transparent they will be about the revisions.
Coincidentally, this may lead to better motherboards with more efficient VRMs and tighter voltage tolerances. Motherboard makers have long used dodgy marketing around a board’s VRM design in how VRM phases are advertised. This type of advertising often leads consumers to believe a particular motherboard’s VRM is better than it really is.